root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-27 03:44:35 +00:00
Code Issues Actions 4 Packages Projects Releases Wiki Activity
8a86b95e87
llama.cpp/ggml-quants.c

7281 lines
282 KiB
C
Raw Normal View History

ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
#include "ggml-quants.h"
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
#include "ggml.h"
#include <math.h>
#include <string.h>
#include <assert.h>
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
#include <float.h>
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
#ifdef __ARM_NEON
// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
//
// $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
//
#include <arm_neon.h>
k-quants : fix build on armv7 (android only) (#2920) * k-quants : fix build on armv7 * ggml : cleanup unused arm32 specific impl * k-quants : avoid some unused vzero / mzero define * ggml-alloc : use 4g for MEASURE_MAX_SIZE in 32-bit arm
2023-09-02 12:23:45 +00:00
#if !defined(__aarch64__)
inline static int32_t vaddvq_s16(int16x8_t v) {
return
(int32_t)vgetq_lane_s16(v, 0) + (int32_t)vgetq_lane_s16(v, 1) +
(int32_t)vgetq_lane_s16(v, 2) + (int32_t)vgetq_lane_s16(v, 3) +
(int32_t)vgetq_lane_s16(v, 4) + (int32_t)vgetq_lane_s16(v, 5) +
(int32_t)vgetq_lane_s16(v, 6) + (int32_t)vgetq_lane_s16(v, 7);
}
inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
return vcombine_s16(a0, b0);
}
inline static int32_t vaddvq_s32(int32x4_t v) {
return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
}
#endif
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
#else
#ifdef __wasm_simd128__
#include <wasm_simd128.h>
#else
#ifdef __POWER9_VECTOR__
#include <altivec.h>
#undef bool
#define bool _Bool
#else
#if defined(_MSC_VER) || defined(__MINGW32__)
#include <intrin.h>
#else
gguf : support big endian platform (#3552) * check whether platform is 390x if yes->do not import immintrin.h * support s390x big endian * support --bigendian option for s390x 1. verified with baichuan7b-chat with float 16 on s390x 2. verified with baichuan7b-chat 3. verified with chinese-alpaca-2-13b-f16 * update format based on editor-config checker result * Update convert-baichuan-hf-to-gguf.py * 1. check in ggml.c if endianess is not match 2. update GGUF version 3. change get_pack_prefix to property 4. update information log * always use "GGUF" as beginng of GGUF file * Compare "GGUF" with file header char by char 1. Set GGUF_MAGIC to "GGUF" string instead of int value 2. Compare "GGUF" char by char to ensure its byte order 3. Move bytes swap code from convert.py to gguf.py write_tensor_data --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-10-20 11:19:40 +00:00
#if !defined(__riscv) && !defined(__s390__)
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
#include <immintrin.h>
#endif
#endif
#endif
#endif
#endif
ggml : add RISC-V Vector Support for K-Quants and improved the existing intrinsics (#3453) * Added RVV intrinsics support for Q8 quantize row and also improved the existing dot product function for risc-v. The RVV intrinsics is added for the following quantize row functions quantize_row_q8_0 quantize_row_q8_1 The following dot product functions have also been optimized by using LMUL = 1/2 instead of LMUL = 1 ggml_vec_dot_q4_0_q8_0 ggml_vec_dot_q4_1_q8_1 ggml_vec_dot_q5_0_q8_0 ggml_vec_dot_q5_1_q8_1 And vector initialization in Q5 by temporary array is also replaced by the vid intrinsics Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai> * Added RVV intrinsics support for k_quants This adds RISC-V Vector intrinsics support for the following K_quants functions for both QKK = 256 and QKK = 64 ggml_vec_dot_q2_K_q8_K ggml_vec_dot_q3_K_q8_K ggml_vec_dot_q4_K_q8_K ggml_vec_dot_q5_K_q8_K ggml_vec_dot_q6_K_q8_K Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai> --------- Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai>
2023-10-03 18:38:19 +00:00
#ifdef __riscv_v_intrinsic
#include <riscv_vector.h>
#endif
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
#undef MIN
#undef MAX
#define MIN(a, b) ((a) < (b) ? (a) : (b))
#define MAX(a, b) ((a) > (b) ? (a) : (b))
ggml : workaround for missing _mm256_setr_m128i in GCC < 8 in k_quants.c (#2405)
2023-07-28 18:17:45 +00:00
#define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__)
// multiply int8_t, add results pairwise twice
static inline __m128i mul_sum_i8_pairs(const __m128i x, const __m128i y) {
// Get absolute values of x vectors
const __m128i ax = _mm_sign_epi8(x, x);
// Sign the values of the y vectors
const __m128i sy = _mm_sign_epi8(y, x);
// Perform multiplication and create 16-bit values
const __m128i dot = _mm_maddubs_epi16(ax, sy);
const __m128i ones = _mm_set1_epi16(1);
return _mm_madd_epi16(ones, dot);
}
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
#if __AVX__ || __AVX2__ || __AVX512F__
// horizontally add 8 floats
static inline float hsum_float_8(const __m256 x) {
__m128 res = _mm256_extractf128_ps(x, 1);
res = _mm_add_ps(res, _mm256_castps256_ps128(x));
res = _mm_add_ps(res, _mm_movehl_ps(res, res));
res = _mm_add_ss(res, _mm_movehdup_ps(res));
return _mm_cvtss_f32(res);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// horizontally add 8 int32_t
static inline int hsum_i32_8(const __m256i a) {
const __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(a), _mm256_extractf128_si256(a, 1));
const __m128i hi64 = _mm_unpackhi_epi64(sum128, sum128);
const __m128i sum64 = _mm_add_epi32(hi64, sum128);
const __m128i hi32 = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1));
return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32));
}
// horizontally add 4 int32_t
static inline int hsum_i32_4(const __m128i a) {
const __m128i hi64 = _mm_unpackhi_epi64(a, a);
const __m128i sum64 = _mm_add_epi32(hi64, a);
const __m128i hi32 = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1));
return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32));
}
#if defined(__AVX2__) || defined(__AVX512F__)
// spread 32 bits to 32 bytes { 0x00, 0xFF }
static inline __m256i bytes_from_bits_32(const uint8_t * x) {
uint32_t x32;
memcpy(&x32, x, sizeof(uint32_t));
const __m256i shuf_mask = _mm256_set_epi64x(
0x0303030303030303, 0x0202020202020202,
0x0101010101010101, 0x0000000000000000);
__m256i bytes = _mm256_shuffle_epi8(_mm256_set1_epi32(x32), shuf_mask);
const __m256i bit_mask = _mm256_set1_epi64x(0x7fbfdfeff7fbfdfe);
bytes = _mm256_or_si256(bytes, bit_mask);
return _mm256_cmpeq_epi8(bytes, _mm256_set1_epi64x(-1));
}
// Unpack 32 4-bit fields into 32 bytes
// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval
static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi)
{
const __m128i tmp = _mm_loadu_si128((const __m128i *)rsi);
const __m256i bytes = MM256_SET_M128I(_mm_srli_epi16(tmp, 4), tmp);
const __m256i lowMask = _mm256_set1_epi8( 0xF );
return _mm256_and_si256(lowMask, bytes);
}
// add int16_t pairwise and return as float vector
static inline __m256 sum_i16_pairs_float(const __m256i x) {
const __m256i ones = _mm256_set1_epi16(1);
const __m256i summed_pairs = _mm256_madd_epi16(ones, x);
return _mm256_cvtepi32_ps(summed_pairs);
}
static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) {
#if __AVXVNNI__
const __m256i zero = _mm256_setzero_si256();
const __m256i summed_pairs = _mm256_dpbusd_epi32(zero, ax, sy);
return _mm256_cvtepi32_ps(summed_pairs);
#else
// Perform multiplication and create 16-bit values
const __m256i dot = _mm256_maddubs_epi16(ax, sy);
return sum_i16_pairs_float(dot);
#endif
}
// multiply int8_t, add results pairwise twice and return as float vector
static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) {
#if __AVXVNNIINT8__
const __m256i zero = _mm256_setzero_si256();
const __m256i summed_pairs = _mm256_dpbssd_epi32(zero, x, y);
return _mm256_cvtepi32_ps(summed_pairs);
#else
// Get absolute values of x vectors
const __m256i ax = _mm256_sign_epi8(x, x);
// Sign the values of the y vectors
const __m256i sy = _mm256_sign_epi8(y, x);
return mul_sum_us8_pairs_float(ax, sy);
#endif
}
static inline __m128i packNibbles( __m256i bytes )
{
// Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh
#if __AVX512F__
const __m256i bytes_srli_4 = _mm256_srli_epi16(bytes, 4); // 0000_0000_abcd_0000
bytes = _mm256_or_si256(bytes, bytes_srli_4); // 0000_abcd_abcd_efgh
return _mm256_cvtepi16_epi8(bytes); // abcd_efgh
#else
const __m256i lowByte = _mm256_set1_epi16( 0xFF );
__m256i high = _mm256_andnot_si256( lowByte, bytes );
__m256i low = _mm256_and_si256( lowByte, bytes );
high = _mm256_srli_epi16( high, 4 );
bytes = _mm256_or_si256( low, high );
// Compress uint16_t lanes into bytes
__m128i r0 = _mm256_castsi256_si128( bytes );
__m128i r1 = _mm256_extracti128_si256( bytes, 1 );
return _mm_packus_epi16( r0, r1 );
#endif
}
#elif defined(__AVX__)
// spread 32 bits to 32 bytes { 0x00, 0xFF }
static inline __m256i bytes_from_bits_32(const uint8_t * x) {
uint32_t x32;
memcpy(&x32, x, sizeof(uint32_t));
const __m128i shuf_maskl = _mm_set_epi64x(0x0101010101010101, 0x0000000000000000);
const __m128i shuf_maskh = _mm_set_epi64x(0x0303030303030303, 0x0202020202020202);
__m128i bytesl = _mm_shuffle_epi8(_mm_set1_epi32(x32), shuf_maskl);
__m128i bytesh = _mm_shuffle_epi8(_mm_set1_epi32(x32), shuf_maskh);
const __m128i bit_mask = _mm_set1_epi64x(0x7fbfdfeff7fbfdfe);
bytesl = _mm_or_si128(bytesl, bit_mask);
bytesh = _mm_or_si128(bytesh, bit_mask);
bytesl = _mm_cmpeq_epi8(bytesl, _mm_set1_epi64x(-1));
bytesh = _mm_cmpeq_epi8(bytesh, _mm_set1_epi64x(-1));
return MM256_SET_M128I(bytesh, bytesl);
}
// Unpack 32 4-bit fields into 32 bytes
// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval
static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi)
{
// Load 16 bytes from memory
__m128i tmpl = _mm_loadu_si128((const __m128i *)rsi);
__m128i tmph = _mm_srli_epi16(tmpl, 4);
const __m128i lowMask = _mm_set1_epi8(0xF);
tmpl = _mm_and_si128(lowMask, tmpl);
tmph = _mm_and_si128(lowMask, tmph);
return MM256_SET_M128I(tmph, tmpl);
}
// add int16_t pairwise and return as float vector
static inline __m256 sum_i16_pairs_float(const __m128i xh, const __m128i xl) {
const __m128i ones = _mm_set1_epi16(1);
const __m128i summed_pairsl = _mm_madd_epi16(ones, xl);
const __m128i summed_pairsh = _mm_madd_epi16(ones, xh);
const __m256i summed_pairs = MM256_SET_M128I(summed_pairsh, summed_pairsl);
return _mm256_cvtepi32_ps(summed_pairs);
}
static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) {
const __m128i axl = _mm256_castsi256_si128(ax);
const __m128i axh = _mm256_extractf128_si256(ax, 1);
const __m128i syl = _mm256_castsi256_si128(sy);
const __m128i syh = _mm256_extractf128_si256(sy, 1);
// Perform multiplication and create 16-bit values
const __m128i dotl = _mm_maddubs_epi16(axl, syl);
const __m128i doth = _mm_maddubs_epi16(axh, syh);
return sum_i16_pairs_float(doth, dotl);
}
// multiply int8_t, add results pairwise twice and return as float vector
static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) {
const __m128i xl = _mm256_castsi256_si128(x);
const __m128i xh = _mm256_extractf128_si256(x, 1);
const __m128i yl = _mm256_castsi256_si128(y);
const __m128i yh = _mm256_extractf128_si256(y, 1);
// Get absolute values of x vectors
const __m128i axl = _mm_sign_epi8(xl, xl);
const __m128i axh = _mm_sign_epi8(xh, xh);
// Sign the values of the y vectors
const __m128i syl = _mm_sign_epi8(yl, xl);
const __m128i syh = _mm_sign_epi8(yh, xh);
// Perform multiplication and create 16-bit values
const __m128i dotl = _mm_maddubs_epi16(axl, syl);
const __m128i doth = _mm_maddubs_epi16(axh, syh);
return sum_i16_pairs_float(doth, dotl);
}
static inline __m128i packNibbles( __m128i bytes1, __m128i bytes2 )
{
// Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh
const __m128i lowByte = _mm_set1_epi16( 0xFF );
__m128i high = _mm_andnot_si128( lowByte, bytes1 );
__m128i low = _mm_and_si128( lowByte, bytes1 );
high = _mm_srli_epi16( high, 4 );
bytes1 = _mm_or_si128( low, high );
high = _mm_andnot_si128( lowByte, bytes2 );
low = _mm_and_si128( lowByte, bytes2 );
high = _mm_srli_epi16( high, 4 );
bytes2 = _mm_or_si128( low, high );
return _mm_packus_epi16( bytes1, bytes2);
}
#endif
#elif defined(__SSSE3__)
// horizontally add 4x4 floats
static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128 c, const __m128 d) {
__m128 res_0 =_mm_hadd_ps(a, b);
__m128 res_1 =_mm_hadd_ps(c, d);
__m128 res =_mm_hadd_ps(res_0, res_1);
res =_mm_hadd_ps(res, res);
res =_mm_hadd_ps(res, res);
return _mm_cvtss_f32(res);
}
#endif // __AVX__ || __AVX2__ || __AVX512F__
#endif // defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__)
#if defined(__ARM_NEON)
#if !defined(__aarch64__)
inline static int32_t vaddvq_s32(int32x4_t v) {
return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
}
inline static float vaddvq_f32(float32x4_t v) {
return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
}
inline static float vmaxvq_f32(float32x4_t v) {
return
MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
}
inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) {
int32x4_t res;
res[0] = roundf(vgetq_lane_f32(v, 0));
res[1] = roundf(vgetq_lane_f32(v, 1));
res[2] = roundf(vgetq_lane_f32(v, 2));
res[3] = roundf(vgetq_lane_f32(v, 3));
return res;
}
#endif
#endif
#if defined(__ARM_NEON) || defined(__wasm_simd128__)
#define B1(c,s,n) 0x ## n ## c , 0x ## n ## s
#define B2(c,s,n) B1(c,s,n ## c), B1(c,s,n ## s)
#define B3(c,s,n) B2(c,s,n ## c), B2(c,s,n ## s)
#define B4(c,s,n) B3(c,s,n ## c), B3(c,s,n ## s)
#define B5(c,s,n) B4(c,s,n ## c), B4(c,s,n ## s)
#define B6(c,s,n) B5(c,s,n ## c), B5(c,s,n ## s)
#define B7(c,s,n) B6(c,s,n ## c), B6(c,s,n ## s)
#define B8(c,s ) B7(c,s, c), B7(c,s, s)
// precomputed tables for expanding 8bits to 8 bytes:
static const uint64_t table_b2b_0[1 << 8] = { B8(00, 10) }; // ( b) << 4
static const uint64_t table_b2b_1[1 << 8] = { B8(10, 00) }; // (!b) << 4
#endif
// reference implementation for deterministic creation of model files
void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k) {
static const int qk = QK4_0;
assert(k % qk == 0);
const int nb = k / qk;
for (int i = 0; i < nb; i++) {
float amax = 0.0f; // absolute max
float max = 0.0f;
for (int j = 0; j < qk; j++) {
const float v = x[i*qk + j];
if (amax < fabsf(v)) {
amax = fabsf(v);
max = v;
}
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
const float d = max / -8;
const float id = d ? 1.0f/d : 0.0f;
y[i].d = ggml_fp32_to_fp16(d);
for (int j = 0; j < qk/2; ++j) {
const float x0 = x[i*qk + 0 + j]*id;
const float x1 = x[i*qk + qk/2 + j]*id;
const uint8_t xi0 = MIN(15, (int8_t)(x0 + 8.5f));
const uint8_t xi1 = MIN(15, (int8_t)(x1 + 8.5f));
y[i].qs[j] = xi0;
y[i].qs[j] |= xi1 << 4;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
}
void quantize_row_q4_0(const float * restrict x, void * restrict y, int k) {
quantize_row_q4_0_reference(x, y, k);
}
void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict y, int k) {
const int qk = QK4_1;
assert(k % qk == 0);
const int nb = k / qk;
for (int i = 0; i < nb; i++) {
float min = FLT_MAX;
float max = -FLT_MAX;
for (int j = 0; j < qk; j++) {
const float v = x[i*qk + j];
if (v < min) min = v;
if (v > max) max = v;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
const float d = (max - min) / ((1 << 4) - 1);
const float id = d ? 1.0f/d : 0.0f;
y[i].d = ggml_fp32_to_fp16(d);
y[i].m = ggml_fp32_to_fp16(min);
for (int j = 0; j < qk/2; ++j) {
const float x0 = (x[i*qk + 0 + j] - min)*id;
const float x1 = (x[i*qk + qk/2 + j] - min)*id;
const uint8_t xi0 = MIN(15, (int8_t)(x0 + 0.5f));
const uint8_t xi1 = MIN(15, (int8_t)(x1 + 0.5f));
y[i].qs[j] = xi0;
y[i].qs[j] |= xi1 << 4;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
}
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
void quantize_row_q4_1(const float * restrict x, void * restrict y, int k) {
quantize_row_q4_1_reference(x, y, k);
}
void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict y, int k) {
static const int qk = QK5_0;
assert(k % qk == 0);
const int nb = k / qk;
for (int i = 0; i < nb; i++) {
float amax = 0.0f; // absolute max
float max = 0.0f;
for (int j = 0; j < qk; j++) {
const float v = x[i*qk + j];
if (amax < fabsf(v)) {
amax = fabsf(v);
max = v;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
const float d = max / -16;
const float id = d ? 1.0f/d : 0.0f;
y[i].d = ggml_fp32_to_fp16(d);
uint32_t qh = 0;
for (int j = 0; j < qk/2; ++j) {
const float x0 = x[i*qk + 0 + j]*id;
const float x1 = x[i*qk + qk/2 + j]*id;
const uint8_t xi0 = MIN(31, (int8_t)(x0 + 16.5f));
const uint8_t xi1 = MIN(31, (int8_t)(x1 + 16.5f));
y[i].qs[j] = (xi0 & 0x0F) | ((xi1 & 0x0F) << 4);
// get the 5-th bit and store it in qh at the right position
qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
qh |= ((xi1 & 0x10u) >> 4) << (j + qk/2);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
memcpy(&y[i].qh, &qh, sizeof(qh));
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
void quantize_row_q5_0(const float * restrict x, void * restrict y, int k) {
quantize_row_q5_0_reference(x, y, k);
}
void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict y, int k) {
const int qk = QK5_1;
assert(k % qk == 0);
const int nb = k / qk;
for (int i = 0; i < nb; i++) {
float min = FLT_MAX;
float max = -FLT_MAX;
for (int j = 0; j < qk; j++) {
const float v = x[i*qk + j];
if (v < min) min = v;
if (v > max) max = v;
}
const float d = (max - min) / ((1 << 5) - 1);
const float id = d ? 1.0f/d : 0.0f;
y[i].d = ggml_fp32_to_fp16(d);
y[i].m = ggml_fp32_to_fp16(min);
uint32_t qh = 0;
for (int j = 0; j < qk/2; ++j) {
const float x0 = (x[i*qk + 0 + j] - min)*id;
const float x1 = (x[i*qk + qk/2 + j] - min)*id;
const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
y[i].qs[j] = (xi0 & 0x0F) | ((xi1 & 0x0F) << 4);
// get the 5-th bit and store it in qh at the right position
qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
qh |= ((xi1 & 0x10u) >> 4) << (j + qk/2);
}
memcpy(&y[i].qh, &qh, sizeof(y[i].qh));
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
}
void quantize_row_q5_1(const float * restrict x, void * restrict y, int k) {
quantize_row_q5_1_reference(x, y, k);
}
// reference implementation for deterministic creation of model files
void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * restrict y, int k) {
assert(k % QK8_0 == 0);
const int nb = k / QK8_0;
for (int i = 0; i < nb; i++) {
float amax = 0.0f; // absolute max
for (int j = 0; j < QK8_0; j++) {
const float v = x[i*QK8_0 + j];
amax = MAX(amax, fabsf(v));
}
const float d = amax / ((1 << 7) - 1);
const float id = d ? 1.0f/d : 0.0f;
y[i].d = ggml_fp32_to_fp16(d);
for (int j = 0; j < QK8_0; ++j) {
const float x0 = x[i*QK8_0 + j]*id;
y[i].qs[j] = roundf(x0);
}
}
}
void quantize_row_q8_0(const float * restrict x, void * restrict vy, int k) {
assert(QK8_0 == 32);
assert(k % QK8_0 == 0);
const int nb = k / QK8_0;
block_q8_0 * restrict y = vy;
#if defined(__ARM_NEON)
for (int i = 0; i < nb; i++) {
float32x4_t srcv [8];
float32x4_t asrcv[8];
float32x4_t amaxv[8];
for (int j = 0; j < 8; j++) srcv[j] = vld1q_f32(x + i*32 + 4*j);
for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[j]);
for (int j = 0; j < 4; j++) amaxv[2*j] = vmaxq_f32(asrcv[2*j], asrcv[2*j+1]);
for (int j = 0; j < 2; j++) amaxv[4*j] = vmaxq_f32(amaxv[4*j], amaxv[4*j+2]);
for (int j = 0; j < 1; j++) amaxv[8*j] = vmaxq_f32(amaxv[8*j], amaxv[8*j+4]);
const float amax = vmaxvq_f32(amaxv[0]);
const float d = amax / ((1 << 7) - 1);
const float id = d ? 1.0f/d : 0.0f;
y[i].d = ggml_fp32_to_fp16(d);
for (int j = 0; j < 8; j++) {
const float32x4_t v = vmulq_n_f32(srcv[j], id);
const int32x4_t vi = vcvtnq_s32_f32(v);
y[i].qs[4*j + 0] = vgetq_lane_s32(vi, 0);
y[i].qs[4*j + 1] = vgetq_lane_s32(vi, 1);
y[i].qs[4*j + 2] = vgetq_lane_s32(vi, 2);
y[i].qs[4*j + 3] = vgetq_lane_s32(vi, 3);
}
}
#elif defined(__wasm_simd128__)
for (int i = 0; i < nb; i++) {
v128_t srcv [8];
v128_t asrcv[8];
v128_t amaxv[8];
for (int j = 0; j < 8; j++) srcv[j] = wasm_v128_load(x + i*32 + 4*j);
for (int j = 0; j < 8; j++) asrcv[j] = wasm_f32x4_abs(srcv[j]);
for (int j = 0; j < 4; j++) amaxv[2*j] = wasm_f32x4_max(asrcv[2*j], asrcv[2*j+1]);
for (int j = 0; j < 2; j++) amaxv[4*j] = wasm_f32x4_max(amaxv[4*j], amaxv[4*j+2]);
for (int j = 0; j < 1; j++) amaxv[8*j] = wasm_f32x4_max(amaxv[8*j], amaxv[8*j+4]);
const float amax = MAX(MAX(wasm_f32x4_extract_lane(amaxv[0], 0),
wasm_f32x4_extract_lane(amaxv[0], 1)),
MAX(wasm_f32x4_extract_lane(amaxv[0], 2),
wasm_f32x4_extract_lane(amaxv[0], 3)));
const float d = amax / ((1 << 7) - 1);
const float id = d ? 1.0f/d : 0.0f;
y[i].d = ggml_fp32_to_fp16(d);
for (int j = 0; j < 8; j++) {
const v128_t v = wasm_f32x4_mul(srcv[j], wasm_f32x4_splat(id));
const v128_t vi = wasm_i32x4_trunc_sat_f32x4(v);
y[i].qs[4*j + 0] = wasm_i32x4_extract_lane(vi, 0);
y[i].qs[4*j + 1] = wasm_i32x4_extract_lane(vi, 1);
y[i].qs[4*j + 2] = wasm_i32x4_extract_lane(vi, 2);
y[i].qs[4*j + 3] = wasm_i32x4_extract_lane(vi, 3);
}
}
#elif defined(__AVX2__) || defined(__AVX__)
for (int i = 0; i < nb; i++) {
// Load elements into 4 AVX vectors
__m256 v0 = _mm256_loadu_ps( x );
__m256 v1 = _mm256_loadu_ps( x + 8 );
__m256 v2 = _mm256_loadu_ps( x + 16 );
__m256 v3 = _mm256_loadu_ps( x + 24 );
x += 32;
// Compute max(abs(e)) for the block
const __m256 signBit = _mm256_set1_ps( -0.0f );
__m256 maxAbs = _mm256_andnot_ps( signBit, v0 );
maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) );
maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) );
maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) );
__m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) );
max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) );
max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) );
const float maxScalar = _mm_cvtss_f32( max4 );
// Quantize these floats
const float d = maxScalar / 127.f;
y[i].d = ggml_fp32_to_fp16(d);
const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f;
const __m256 mul = _mm256_set1_ps( id );
// Apply the multiplier
v0 = _mm256_mul_ps( v0, mul );
v1 = _mm256_mul_ps( v1, mul );
v2 = _mm256_mul_ps( v2, mul );
v3 = _mm256_mul_ps( v3, mul );
// Round to nearest integer
v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST );
v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST );
v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST );
v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST );
// Convert floats to integers
__m256i i0 = _mm256_cvtps_epi32( v0 );
__m256i i1 = _mm256_cvtps_epi32( v1 );
__m256i i2 = _mm256_cvtps_epi32( v2 );
__m256i i3 = _mm256_cvtps_epi32( v3 );
#if defined(__AVX2__)
// Convert int32 to int16
i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15
i2 = _mm256_packs_epi32( i2, i3 ); // 16, 17, 18, 19, 24, 25, 26, 27, 20, 21, 22, 23, 28, 29, 30, 31
// Convert int16 to int8
i0 = _mm256_packs_epi16( i0, i2 ); // 0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27, 4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31
// We got our precious signed bytes, but the order is now wrong
// These AVX2 pack instructions process 16-byte pieces independently
// The following instruction is fixing the order
const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 );
i0 = _mm256_permutevar8x32_epi32( i0, perm );
_mm256_storeu_si256((__m256i *)y[i].qs, i0);
#else
// Since we don't have in AVX some necessary functions,
// we split the registers in half and call AVX2 analogs from SSE
__m128i ni0 = _mm256_castsi256_si128( i0 );
__m128i ni1 = _mm256_extractf128_si256( i0, 1);
__m128i ni2 = _mm256_castsi256_si128( i1 );
__m128i ni3 = _mm256_extractf128_si256( i1, 1);
__m128i ni4 = _mm256_castsi256_si128( i2 );
__m128i ni5 = _mm256_extractf128_si256( i2, 1);
__m128i ni6 = _mm256_castsi256_si128( i3 );
__m128i ni7 = _mm256_extractf128_si256( i3, 1);
// Convert int32 to int16
ni0 = _mm_packs_epi32( ni0, ni1 );
ni2 = _mm_packs_epi32( ni2, ni3 );
ni4 = _mm_packs_epi32( ni4, ni5 );
ni6 = _mm_packs_epi32( ni6, ni7 );
// Convert int16 to int8
ni0 = _mm_packs_epi16( ni0, ni2 );
ni4 = _mm_packs_epi16( ni4, ni6 );
_mm_storeu_si128((__m128i *)(y[i].qs + 0), ni0);
_mm_storeu_si128((__m128i *)(y[i].qs + 16), ni4);
#endif
}
#elif defined(__riscv_v_intrinsic)
size_t vl = __riscv_vsetvl_e32m4(QK8_0);
for (int i = 0; i < nb; i++) {
// load elements
vfloat32m4_t v_x = __riscv_vle32_v_f32m4(x+i*QK8_0, vl);
vfloat32m4_t vfabs = __riscv_vfabs_v_f32m4(v_x, vl);
vfloat32m1_t tmp = __riscv_vfmv_v_f_f32m1(0.0f, vl);
vfloat32m1_t vmax = __riscv_vfredmax_vs_f32m4_f32m1(vfabs, tmp, vl);
float amax = __riscv_vfmv_f_s_f32m1_f32(vmax);
const float d = amax / ((1 << 7) - 1);
const float id = d ? 1.0f/d : 0.0f;
y[i].d = ggml_fp32_to_fp16(d);
vfloat32m4_t x0 = __riscv_vfmul_vf_f32m4(v_x, id, vl);
// convert to integer
vint16m2_t vi = __riscv_vfncvt_x_f_w_i16m2(x0, vl);
vint8m1_t vs = __riscv_vncvt_x_x_w_i8m1(vi, vl);
// store result
__riscv_vse8_v_i8m1(y[i].qs , vs, vl);
}
#else
// scalar
quantize_row_q8_0_reference(x, y, k);
#endif
}
// reference implementation for deterministic creation of model files
void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict y, int k) {
assert(QK8_1 == 32);
assert(k % QK8_1 == 0);
const int nb = k / QK8_1;
for (int i = 0; i < nb; i++) {
float amax = 0.0f; // absolute max
for (int j = 0; j < QK8_1; j++) {
const float v = x[i*QK8_1 + j];
amax = MAX(amax, fabsf(v));
}
const float d = amax / ((1 << 7) - 1);
const float id = d ? 1.0f/d : 0.0f;
y[i].d = d;
int sum = 0;
for (int j = 0; j < QK8_1/2; ++j) {
const float v0 = x[i*QK8_1 + j]*id;
const float v1 = x[i*QK8_1 + QK8_1/2 + j]*id;
y[i].qs[ j] = roundf(v0);
y[i].qs[QK8_1/2 + j] = roundf(v1);
sum += y[i].qs[ j];
sum += y[i].qs[QK8_1/2 + j];
}
y[i].s = sum*d;
}
}
void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
assert(k % QK8_1 == 0);
const int nb = k / QK8_1;
block_q8_1 * restrict y = vy;
#if defined(__ARM_NEON)
for (int i = 0; i < nb; i++) {
float32x4_t srcv [8];
float32x4_t asrcv[8];
float32x4_t amaxv[8];
for (int j = 0; j < 8; j++) srcv[j] = vld1q_f32(x + i*32 + 4*j);
for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[j]);
for (int j = 0; j < 4; j++) amaxv[2*j] = vmaxq_f32(asrcv[2*j], asrcv[2*j+1]);
for (int j = 0; j < 2; j++) amaxv[4*j] = vmaxq_f32(amaxv[4*j], amaxv[4*j+2]);
for (int j = 0; j < 1; j++) amaxv[8*j] = vmaxq_f32(amaxv[8*j], amaxv[8*j+4]);
const float amax = vmaxvq_f32(amaxv[0]);
const float d = amax / ((1 << 7) - 1);
const float id = d ? 1.0f/d : 0.0f;
y[i].d = d;
int32x4_t accv = vdupq_n_s32(0);
for (int j = 0; j < 8; j++) {
const float32x4_t v = vmulq_n_f32(srcv[j], id);
const int32x4_t vi = vcvtnq_s32_f32(v);
y[i].qs[4*j + 0] = vgetq_lane_s32(vi, 0);
y[i].qs[4*j + 1] = vgetq_lane_s32(vi, 1);
y[i].qs[4*j + 2] = vgetq_lane_s32(vi, 2);
y[i].qs[4*j + 3] = vgetq_lane_s32(vi, 3);
accv = vaddq_s32(accv, vi);
}
y[i].s = d * vaddvq_s32(accv);
}
#elif defined(__wasm_simd128__)
for (int i = 0; i < nb; i++) {
v128_t srcv [8];
v128_t asrcv[8];
v128_t amaxv[8];
for (int j = 0; j < 8; j++) srcv[j] = wasm_v128_load(x + i*32 + 4*j);
for (int j = 0; j < 8; j++) asrcv[j] = wasm_f32x4_abs(srcv[j]);
for (int j = 0; j < 4; j++) amaxv[2*j] = wasm_f32x4_max(asrcv[2*j], asrcv[2*j+1]);
for (int j = 0; j < 2; j++) amaxv[4*j] = wasm_f32x4_max(amaxv[4*j], amaxv[4*j+2]);
for (int j = 0; j < 1; j++) amaxv[8*j] = wasm_f32x4_max(amaxv[8*j], amaxv[8*j+4]);
const float amax = MAX(MAX(wasm_f32x4_extract_lane(amaxv[0], 0),
wasm_f32x4_extract_lane(amaxv[0], 1)),
MAX(wasm_f32x4_extract_lane(amaxv[0], 2),
wasm_f32x4_extract_lane(amaxv[0], 3)));
const float d = amax / ((1 << 7) - 1);
const float id = d ? 1.0f/d : 0.0f;
y[i].d = d;
v128_t accv = wasm_i32x4_splat(0);
for (int j = 0; j < 8; j++) {
const v128_t v = wasm_f32x4_mul(srcv[j], wasm_f32x4_splat(id));
const v128_t vi = wasm_i32x4_trunc_sat_f32x4(v);
y[i].qs[4*j + 0] = wasm_i32x4_extract_lane(vi, 0);
y[i].qs[4*j + 1] = wasm_i32x4_extract_lane(vi, 1);
y[i].qs[4*j + 2] = wasm_i32x4_extract_lane(vi, 2);
y[i].qs[4*j + 3] = wasm_i32x4_extract_lane(vi, 3);
accv = wasm_i32x4_add(accv, vi);
}
y[i].s = d * (wasm_i32x4_extract_lane(accv, 0) +
wasm_i32x4_extract_lane(accv, 1) +
wasm_i32x4_extract_lane(accv, 2) +
wasm_i32x4_extract_lane(accv, 3));
}
#elif defined(__AVX2__) || defined(__AVX__)
for (int i = 0; i < nb; i++) {
// Load elements into 4 AVX vectors
__m256 v0 = _mm256_loadu_ps( x );
__m256 v1 = _mm256_loadu_ps( x + 8 );
__m256 v2 = _mm256_loadu_ps( x + 16 );
__m256 v3 = _mm256_loadu_ps( x + 24 );
x += 32;
// Compute max(abs(e)) for the block
const __m256 signBit = _mm256_set1_ps( -0.0f );
__m256 maxAbs = _mm256_andnot_ps( signBit, v0 );
maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) );
maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) );
maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) );
__m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) );
max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) );
max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) );
const float maxScalar = _mm_cvtss_f32( max4 );
// Quantize these floats
const float d = maxScalar / 127.f;
y[i].d = d;
const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f;
const __m256 mul = _mm256_set1_ps( id );
// Apply the multiplier
v0 = _mm256_mul_ps( v0, mul );
v1 = _mm256_mul_ps( v1, mul );
v2 = _mm256_mul_ps( v2, mul );
v3 = _mm256_mul_ps( v3, mul );
// Round to nearest integer
v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST );
v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST );
v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST );
v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST );
// Convert floats to integers
__m256i i0 = _mm256_cvtps_epi32( v0 );
__m256i i1 = _mm256_cvtps_epi32( v1 );
__m256i i2 = _mm256_cvtps_epi32( v2 );
__m256i i3 = _mm256_cvtps_epi32( v3 );
#if defined(__AVX2__)
// Compute the sum of the quants and set y[i].s
y[i].s = d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3)));
// Convert int32 to int16
i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15
i2 = _mm256_packs_epi32( i2, i3 ); // 16, 17, 18, 19, 24, 25, 26, 27, 20, 21, 22, 23, 28, 29, 30, 31
// Convert int16 to int8
i0 = _mm256_packs_epi16( i0, i2 ); // 0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27, 4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31
// We got our precious signed bytes, but the order is now wrong
// These AVX2 pack instructions process 16-byte pieces independently
// The following instruction is fixing the order
const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 );
i0 = _mm256_permutevar8x32_epi32( i0, perm );
_mm256_storeu_si256((__m256i *)y[i].qs, i0);
#else
// Since we don't have in AVX some necessary functions,
// we split the registers in half and call AVX2 analogs from SSE
__m128i ni0 = _mm256_castsi256_si128( i0 );
__m128i ni1 = _mm256_extractf128_si256( i0, 1);
__m128i ni2 = _mm256_castsi256_si128( i1 );
__m128i ni3 = _mm256_extractf128_si256( i1, 1);
__m128i ni4 = _mm256_castsi256_si128( i2 );
__m128i ni5 = _mm256_extractf128_si256( i2, 1);
__m128i ni6 = _mm256_castsi256_si128( i3 );
__m128i ni7 = _mm256_extractf128_si256( i3, 1);
// Compute the sum of the quants and set y[i].s
const __m128i s0 = _mm_add_epi32(_mm_add_epi32(ni0, ni1), _mm_add_epi32(ni2, ni3));
const __m128i s1 = _mm_add_epi32(_mm_add_epi32(ni4, ni5), _mm_add_epi32(ni6, ni7));
y[i].s = d * hsum_i32_4(_mm_add_epi32(s0, s1));
// Convert int32 to int16
ni0 = _mm_packs_epi32( ni0, ni1 );
ni2 = _mm_packs_epi32( ni2, ni3 );
ni4 = _mm_packs_epi32( ni4, ni5 );
ni6 = _mm_packs_epi32( ni6, ni7 );
// Convert int16 to int8
ni0 = _mm_packs_epi16( ni0, ni2 );
ni4 = _mm_packs_epi16( ni4, ni6 );
_mm_storeu_si128((__m128i *)(y[i].qs + 0), ni0);
_mm_storeu_si128((__m128i *)(y[i].qs + 16), ni4);
#endif
}
#elif defined(__riscv_v_intrinsic)
size_t vl = __riscv_vsetvl_e32m4(QK8_1);
for (int i = 0; i < nb; i++) {
// load elements
vfloat32m4_t v_x = __riscv_vle32_v_f32m4(x+i*QK8_1, vl);
vfloat32m4_t vfabs = __riscv_vfabs_v_f32m4(v_x, vl);
vfloat32m1_t tmp = __riscv_vfmv_v_f_f32m1(0.0, vl);
vfloat32m1_t vmax = __riscv_vfredmax_vs_f32m4_f32m1(vfabs, tmp, vl);
float amax = __riscv_vfmv_f_s_f32m1_f32(vmax);
const float d = amax / ((1 << 7) - 1);
const float id = d ? 1.0f/d : 0.0f;
y[i].d = d;
vfloat32m4_t x0 = __riscv_vfmul_vf_f32m4(v_x, id, vl);
// convert to integer
vint16m2_t vi = __riscv_vfncvt_x_f_w_i16m2(x0, vl);
vint8m1_t vs = __riscv_vncvt_x_x_w_i8m1(vi, vl);
// store result
__riscv_vse8_v_i8m1(y[i].qs , vs, vl);
// compute sum for y[i].s
vint16m1_t tmp2 = __riscv_vmv_v_x_i16m1(0, vl);
vint16m1_t vwrs = __riscv_vwredsum_vs_i8m1_i16m1(vs, tmp2, vl);
// set y[i].s
int sum = __riscv_vmv_x_s_i16m1_i16(vwrs);
y[i].s = sum*d;
}
#else
// scalar
quantize_row_q8_1_reference(x, y, k);
#endif
}
void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int k) {
static const int qk = QK4_0;
assert(k % qk == 0);
const int nb = k / qk;
for (int i = 0; i < nb; i++) {
const float d = ggml_fp16_to_fp32(x[i].d);
for (int j = 0; j < qk/2; ++j) {
const int x0 = (x[i].qs[j] & 0x0F) - 8;
const int x1 = (x[i].qs[j] >> 4) - 8;
y[i*qk + j + 0 ] = x0*d;
y[i*qk + j + qk/2] = x1*d;
}
}
}
void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int k) {
static const int qk = QK4_1;
assert(k % qk == 0);
const int nb = k / qk;
for (int i = 0; i < nb; i++) {
const float d = ggml_fp16_to_fp32(x[i].d);
const float m = ggml_fp16_to_fp32(x[i].m);
for (int j = 0; j < qk/2; ++j) {
const int x0 = (x[i].qs[j] & 0x0F);
const int x1 = (x[i].qs[j] >> 4);
y[i*qk + j + 0 ] = x0*d + m;
y[i*qk + j + qk/2] = x1*d + m;
}
}
}
void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int k) {
static const int qk = QK5_0;
assert(k % qk == 0);
const int nb = k / qk;
for (int i = 0; i < nb; i++) {
const float d = ggml_fp16_to_fp32(x[i].d);
uint32_t qh;
memcpy(&qh, x[i].qh, sizeof(qh));
for (int j = 0; j < qk/2; ++j) {
const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10;
const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10;
const int32_t x0 = ((x[i].qs[j] & 0x0F) | xh_0) - 16;
const int32_t x1 = ((x[i].qs[j] >> 4) | xh_1) - 16;
y[i*qk + j + 0 ] = x0*d;
y[i*qk + j + qk/2] = x1*d;
}
}
}
void dequantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int k) {
static const int qk = QK5_1;
assert(k % qk == 0);
const int nb = k / qk;
for (int i = 0; i < nb; i++) {
const float d = ggml_fp16_to_fp32(x[i].d);
const float m = ggml_fp16_to_fp32(x[i].m);
uint32_t qh;
memcpy(&qh, x[i].qh, sizeof(qh));
for (int j = 0; j < qk/2; ++j) {
const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10;
const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10;
const int x0 = (x[i].qs[j] & 0x0F) | xh_0;
const int x1 = (x[i].qs[j] >> 4) | xh_1;
y[i*qk + j + 0 ] = x0*d + m;
y[i*qk + j + qk/2] = x1*d + m;
}
}
}
void dequantize_row_q8_0(const block_q8_0 * restrict x, float * restrict y, int k) {
static const int qk = QK8_0;
assert(k % qk == 0);
const int nb = k / qk;
for (int i = 0; i < nb; i++) {
const float d = ggml_fp16_to_fp32(x[i].d);
for (int j = 0; j < qk; ++j) {
y[i*qk + j] = x[i].qs[j]*d;
}
}
}
//
// 2-6 bit quantization in super-blocks
//
//
// ===================== Helper functions
//
static inline int nearest_int(float fval) {
assert(fval <= 4194303.f);
float val = fval + 12582912.f;
int i; memcpy(&i, &val, sizeof(int));
return (i & 0x007fffff) - 0x00400000;
}
static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t * restrict L, int rmse_type) {
float max = 0;
float amax = 0;
for (int i = 0; i < n; ++i) {
float ax = fabsf(x[i]);
if (ax > amax) { amax = ax; max = x[i]; }
}
if (amax < 1e-30f) { // all zero
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
for (int i = 0; i < n; ++i) {
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
L[i] = 0;
}
return 0.f;
}
float iscale = -nmax / max;
if (rmse_type == 0) {
for (int i = 0; i < n; ++i) {
int l = nearest_int(iscale * x[i]);
L[i] = nmax + MAX(-nmax, MIN(nmax-1, l));
}
return 1/iscale;
}
bool return_early = false;
if (rmse_type < 0) {
rmse_type = -rmse_type;
return_early = true;
}
int weight_type = rmse_type%2;
float sumlx = 0;
float suml2 = 0;
for (int i = 0; i < n; ++i) {
int l = nearest_int(iscale * x[i]);
l = MAX(-nmax, MIN(nmax-1, l));
L[i] = l + nmax;
float w = weight_type == 1 ? x[i] * x[i] : 1;
sumlx += w*x[i]*l;
suml2 += w*l*l;
}
float scale = sumlx/suml2;
if (return_early) return suml2 > 0 ? 0.5f*(scale + 1/iscale) : 1/iscale;
float best = scale * sumlx;
for (int is = -9; is <= 9; ++is) {
if (is == 0) {
continue;
}
iscale = -(nmax + 0.1f*is) / max;
sumlx = suml2 = 0;
for (int i = 0; i < n; ++i) {
int l = nearest_int(iscale * x[i]);
l = MAX(-nmax, MIN(nmax-1, l));
float w = weight_type == 1 ? x[i] * x[i] : 1;
sumlx += w*x[i]*l;
suml2 += w*l*l;
}
if (suml2 > 0 && sumlx*sumlx > best*suml2) {
for (int i = 0; i < n; ++i) {
int l = nearest_int(iscale * x[i]);
L[i] = nmax + MAX(-nmax, MIN(nmax-1, l));
}
scale = sumlx/suml2; best = scale*sumlx;
}
}
return scale;
}
static float make_q3_quants(int n, int nmax, const float * restrict x, int8_t * restrict L, bool do_rmse) {
float max = 0;
float amax = 0;
for (int i = 0; i < n; ++i) {
float ax = fabsf(x[i]);
if (ax > amax) { amax = ax; max = x[i]; }
}
if (!amax) { // all zero
for (int i = 0; i < n; ++i) { L[i] = 0; }
return 0.f;
}
float iscale = -nmax / max;
if (do_rmse) {
float sumlx = 0;
float suml2 = 0;
for (int i = 0; i < n; ++i) {
int l = nearest_int(iscale * x[i]);
l = MAX(-nmax, MIN(nmax-1, l));
L[i] = l;
float w = x[i]*x[i];
sumlx += w*x[i]*l;
suml2 += w*l*l;
}
for (int itry = 0; itry < 5; ++itry) {
int n_changed = 0;
for (int i = 0; i < n; ++i) {
float w = x[i]*x[i];
float slx = sumlx - w*x[i]*L[i];
if (slx > 0) {
float sl2 = suml2 - w*L[i]*L[i];
int new_l = nearest_int(x[i] * sl2 / slx);
new_l = MAX(-nmax, MIN(nmax-1, new_l));
if (new_l != L[i]) {
slx += w*x[i]*new_l;
sl2 += w*new_l*new_l;
if (sl2 > 0 && slx*slx*suml2 > sumlx*sumlx*sl2) {
L[i] = new_l; sumlx = slx; suml2 = sl2;
++n_changed;
}
}
}
}
if (!n_changed) {
break;
}
}
for (int i = 0; i < n; ++i) {
L[i] += nmax;
}
return sumlx / suml2;
}
for (int i = 0; i < n; ++i) {
int l = nearest_int(iscale * x[i]);
l = MAX(-nmax, MIN(nmax-1, l));
L[i] = l + nmax;
}
return 1/iscale;
}
static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t * restrict L, float * restrict the_min,
int ntry, float alpha) {
float min = x[0];
float max = x[0];
for (int i = 1; i < n; ++i) {
if (x[i] < min) min = x[i];
if (x[i] > max) max = x[i];
}
if (max == min) {
for (int i = 0; i < n; ++i) L[i] = 0;
*the_min = 0;
return 0.f;
}
if (min > 0) min = 0;
float iscale = nmax/(max - min);
float scale = 1/iscale;
for (int itry = 0; itry < ntry; ++itry) {
float sumlx = 0; int suml2 = 0;
bool did_change = false;
for (int i = 0; i < n; ++i) {
int l = nearest_int(iscale*(x[i] - min));
l = MAX(0, MIN(nmax, l));
if (l != L[i]) {
L[i] = l;
did_change = true;
}
sumlx += (x[i] - min)*l;
suml2 += l*l;
}
scale = sumlx/suml2;
float sum = 0;
for (int i = 0; i < n; ++i) {
sum += x[i] - scale*L[i];
}
min = alpha*min + (1 - alpha)*sum/n;
if (min > 0) min = 0;
iscale = 1/scale;
if (!did_change) break;
}
*the_min = -min;
return scale;
}
static float make_qkx2_quants(int n, int nmax, const float * restrict x, const float * restrict weights,
uint8_t * restrict L, float * restrict the_min, uint8_t * restrict Laux,
float rmin, float rdelta, int nstep, bool use_mad) {
float min = x[0];
float max = x[0];
float sum_w = weights[0];
float sum_x = sum_w * x[0];
for (int i = 1; i < n; ++i) {
if (x[i] < min) min = x[i];
if (x[i] > max) max = x[i];
float w = weights[i];
sum_w += w;
sum_x += w * x[i];
}
if (min > 0) min = 0;
if (max == min) {
for (int i = 0; i < n; ++i) L[i] = 0;
*the_min = -min;
return 0.f;
}
float iscale = nmax/(max - min);
float scale = 1/iscale;
float best_mad = 0;
for (int i = 0; i < n; ++i) {
int l = nearest_int(iscale*(x[i] - min));
L[i] = MAX(0, MIN(nmax, l));
float diff = scale * L[i] + min - x[i];
diff = use_mad ? fabsf(diff) : diff * diff;
float w = weights[i];
best_mad += w * diff;
}
if (nstep < 1) {
*the_min = -min;
return scale;
}
for (int is = 0; is <= nstep; ++is) {
iscale = (rmin + rdelta*is + nmax)/(max - min);
float sum_l = 0, sum_l2 = 0, sum_xl = 0;
for (int i = 0; i < n; ++i) {
int l = nearest_int(iscale*(x[i] - min));
l = MAX(0, MIN(nmax, l));
Laux[i] = l;
float w = weights[i];
sum_l += w*l;
sum_l2 += w*l*l;
sum_xl += w*l*x[i];
}
float D = sum_w * sum_l2 - sum_l * sum_l;
if (D > 0) {
float this_scale = (sum_w * sum_xl - sum_x * sum_l)/D;
float this_min = (sum_l2 * sum_x - sum_l * sum_xl)/D;
if (this_min > 0) {
this_min = 0;
this_scale = sum_xl / sum_l2;
}
float mad = 0;
for (int i = 0; i < n; ++i) {
float diff = this_scale * Laux[i] + this_min - x[i];
diff = use_mad ? fabsf(diff) : diff * diff;
float w = weights[i];
mad += w * diff;
}
if (mad < best_mad) {
for (int i = 0; i < n; ++i) {
L[i] = Laux[i];
}
best_mad = mad;
scale = this_scale;
min = this_min;
}
}
}
*the_min = -min;
return scale;
}
#if QK_K == 256
static inline void get_scale_min_k4(int j, const uint8_t * restrict q, uint8_t * restrict d, uint8_t * restrict m) {
if (j < 4) {
*d = q[j] & 63; *m = q[j + 4] & 63;
} else {
*d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4);
*m = (q[j+4] >> 4) | ((q[j-0] >> 6) << 4);
}
}
#endif
//========================- 2-bit (de)-quantization
void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict y, int k) {
assert(k % QK_K == 0);
const int nb = k / QK_K;
uint8_t L[QK_K];
uint8_t Laux[16];
float weights[16];
float mins[QK_K/16];
float scales[QK_K/16];
const float q4scale = 15.f;
for (int i = 0; i < nb; i++) {
float max_scale = 0; // as we are deducting the min, scales are always positive
float max_min = 0;
for (int j = 0; j < QK_K/16; ++j) {
for (int l = 0; l < 16; ++l) weights[l] = fabsf(x[16*j + l]);
scales[j] = make_qkx2_quants(16, 3, x + 16*j, weights, L + 16*j, &mins[j], Laux, -0.5f, 0.1f, 15, true);
float scale = scales[j];
if (scale > max_scale) {
max_scale = scale;
}
float min = mins[j];
if (min > max_min) {
max_min = min;
}
}
if (max_scale > 0) {
float iscale = q4scale/max_scale;
for (int j = 0; j < QK_K/16; ++j) {
int l = nearest_int(iscale*scales[j]);
y[i].scales[j] = l;
}
y[i].d = ggml_fp32_to_fp16(max_scale/q4scale);
} else {
for (int j = 0; j < QK_K/16; ++j) y[i].scales[j] = 0;
y[i].d = ggml_fp32_to_fp16(0.f);
}
if (max_min > 0) {
float iscale = q4scale/max_min;
for (int j = 0; j < QK_K/16; ++j) {
int l = nearest_int(iscale*mins[j]);
y[i].scales[j] |= (l << 4);
}
y[i].dmin = ggml_fp32_to_fp16(max_min/q4scale);
} else {
y[i].dmin = ggml_fp32_to_fp16(0.f);
}
for (int j = 0; j < QK_K/16; ++j) {
const float d = ggml_fp16_to_fp32(y[i].d) * (y[i].scales[j] & 0xF);
if (!d) continue;
const float dm = ggml_fp16_to_fp32(y[i].dmin) * (y[i].scales[j] >> 4);
for (int ii = 0; ii < 16; ++ii) {
int l = nearest_int((x[16*j + ii] + dm)/d);
l = MAX(0, MIN(3, l));
L[16*j + ii] = l;
}
}
#if QK_K == 256
for (int j = 0; j < QK_K; j += 128) {
for (int l = 0; l < 32; ++l) {
y[i].qs[j/4 + l] = L[j + l] | (L[j + l + 32] << 2) | (L[j + l + 64] << 4) | (L[j + l + 96] << 6);
}
}
#else
for (int l = 0; l < 16; ++l) {
y[i].qs[l] = L[l] | (L[l + 16] << 2) | (L[l + 32] << 4) | (L[l + 48] << 6);
}
#endif
x += QK_K;
}
}
void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int k) {
assert(k % QK_K == 0);
const int nb = k / QK_K;
for (int i = 0; i < nb; i++) {
const float d = ggml_fp16_to_fp32(x[i].d);
const float min = ggml_fp16_to_fp32(x[i].dmin);
const uint8_t * q = x[i].qs;
#if QK_K == 256
int is = 0;
float dl, ml;
for (int n = 0; n < QK_K; n += 128) {
int shift = 0;
for (int j = 0; j < 4; ++j) {
uint8_t sc = x[i].scales[is++];
dl = d * (sc & 0xF); ml = min * (sc >> 4);
for (int l = 0; l < 16; ++l) *y++ = dl * ((int8_t)((q[l] >> shift) & 3)) - ml;
sc = x[i].scales[is++];
dl = d * (sc & 0xF); ml = min * (sc >> 4);
for (int l = 0; l < 16; ++l) *y++ = dl * ((int8_t)((q[l+16] >> shift) & 3)) - ml;
shift += 2;
}
q += 32;
}
#else
float dl1 = d * (x[i].scales[0] & 0xF), ml1 = min * (x[i].scales[0] >> 4);
float dl2 = d * (x[i].scales[1] & 0xF), ml2 = min * (x[i].scales[1] >> 4);
float dl3 = d * (x[i].scales[2] & 0xF), ml3 = min * (x[i].scales[2] >> 4);
float dl4 = d * (x[i].scales[3] & 0xF), ml4 = min * (x[i].scales[3] >> 4);
for (int l = 0; l < 16; ++l) {
y[l+ 0] = dl1 * ((int8_t)((q[l] >> 0) & 3)) - ml1;
y[l+16] = dl2 * ((int8_t)((q[l] >> 2) & 3)) - ml2;
y[l+32] = dl3 * ((int8_t)((q[l] >> 4) & 3)) - ml3;
y[l+48] = dl4 * ((int8_t)((q[l] >> 6) & 3)) - ml4;
}
y += QK_K;
#endif
}
}
void quantize_row_q2_K(const float * restrict x, void * restrict vy, int k) {
quantize_row_q2_K_reference(x, vy, k);
}
size_t ggml_quantize_q2_K(const float * restrict src, void * restrict dst, int n, int k, int64_t * restrict hist) {
(void)hist; // TODO: collect histograms
for (int j = 0; j < n; j += k) {
block_q2_K * restrict y = (block_q2_K *)dst + j/QK_K;
quantize_row_q2_K_reference(src + j, y, k);
}
return (n/QK_K*sizeof(block_q2_K));
}
//========================= 3-bit (de)-quantization
void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict y, int k) {
assert(k % QK_K == 0);
const int nb = k / QK_K;
int8_t L[QK_K];
float scales[QK_K / 16];
for (int i = 0; i < nb; i++) {
float max_scale = 0;
float amax = 0;
for (int j = 0; j < QK_K/16; ++j) {
scales[j] = make_q3_quants(16, 4, x + 16*j, L + 16*j, true);
float scale = fabsf(scales[j]);
if (scale > amax) {
amax = scale; max_scale = scales[j];
}
}
#if QK_K == 256
memset(y[i].scales, 0, 12);
if (max_scale) {
float iscale = -32.f/max_scale;
for (int j = 0; j < QK_K/16; ++j) {
int8_t l = nearest_int(iscale*scales[j]);
l = MAX(-32, MIN(31, l)) + 32;
if (j < 8) {
y[i].scales[j] = l & 0xF;
} else {
y[i].scales[j-8] |= ((l & 0xF) << 4);
}
l >>= 4;
y[i].scales[j%4 + 8] |= (l << (2*(j/4)));
}
y[i].d = ggml_fp32_to_fp16(1/iscale);
} else {
y[i].d = ggml_fp32_to_fp16(0.f);
}
int8_t sc;
for (int j = 0; j < QK_K/16; ++j) {
sc = j < 8 ? y[i].scales[j] & 0xF : y[i].scales[j-8] >> 4;
sc = (sc | (((y[i].scales[8 + j%4] >> (2*(j/4))) & 3) << 4)) - 32;
float d = ggml_fp16_to_fp32(y[i].d) * sc;
if (!d) {
continue;
}
for (int ii = 0; ii < 16; ++ii) {
int l = nearest_int(x[16*j + ii]/d);
l = MAX(-4, MIN(3, l));
L[16*j + ii] = l + 4;
}
}
#else
if (max_scale) {
float iscale = -8.f/max_scale;
for (int j = 0; j < QK_K/16; j+=2) {
int l1 = nearest_int(iscale*scales[j]);
l1 = 8 + MAX(-8, MIN(7, l1));
int l2 = nearest_int(iscale*scales[j+1]);
l2 = 8 + MAX(-8, MIN(7, l2));
y[i].scales[j/2] = l1 | (l2 << 4);
}
y[i].d = ggml_fp32_to_fp16(1/iscale);
} else {
for (int j = 0; j < QK_K/16; j+=2) {
y[i].scales[j/2] = 0;
}
y[i].d = ggml_fp32_to_fp16(0.f);
}
for (int j = 0; j < QK_K/16; ++j) {
int s = j%2 == 0 ? y[i].scales[j/2] & 0xF : y[i].scales[j/2] >> 4;
float d = ggml_fp16_to_fp32(y[i].d) * (s - 8);
if (!d) {
continue;
}
for (int ii = 0; ii < 16; ++ii) {
int l = nearest_int(x[16*j + ii]/d);
l = MAX(-4, MIN(3, l));
L[16*j + ii] = l + 4;
}
}
#endif
memset(y[i].hmask, 0, QK_K/8);
// We put the high-bit for the 1st 8 quants into bit 0, the next 8 into bit 1, etc.
int m = 0;
uint8_t hm = 1;
for (int j = 0; j < QK_K; ++j) {
if (L[j] > 3) {
y[i].hmask[m] |= hm;
L[j] -= 4;
}
if (++m == QK_K/8) {
m = 0; hm <<= 1;
}
}
#if QK_K == 256
for (int j = 0; j < QK_K; j += 128) {
for (int l = 0; l < 32; ++l) {
y[i].qs[j/4 + l] = L[j + l] | (L[j + l + 32] << 2) | (L[j + l + 64] << 4) | (L[j + l + 96] << 6);
}
}
#else
for (int l = 0; l < 16; ++l) {
y[i].qs[l] = L[l] | (L[l + 16] << 2) | (L[l + 32] << 4) | (L[l + 48] << 6);
}
#endif
x += QK_K;
}
}
#if QK_K == 256
void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k) {
assert(k % QK_K == 0);
const int nb = k / QK_K;
const uint32_t kmask1 = 0x03030303;
const uint32_t kmask2 = 0x0f0f0f0f;
uint32_t aux[4];
const int8_t * scales = (const int8_t*)aux;
for (int i = 0; i < nb; i++) {
const float d_all = ggml_fp16_to_fp32(x[i].d);
const uint8_t * restrict q = x[i].qs;
const uint8_t * restrict hm = x[i].hmask;
uint8_t m = 1;
memcpy(aux, x[i].scales, 12);
uint32_t tmp = aux[2];
aux[2] = ((aux[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
aux[3] = ((aux[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
aux[0] = (aux[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
aux[1] = (aux[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
int is = 0;
float dl;
for (int n = 0; n < QK_K; n += 128) {
int shift = 0;
for (int j = 0; j < 4; ++j) {
dl = d_all * (scales[is++] - 32);
for (int l = 0; l < 16; ++l) {
*y++ = dl * ((int8_t)((q[l+ 0] >> shift) & 3) - ((hm[l+ 0] & m) ? 0 : 4));
}
dl = d_all * (scales[is++] - 32);
for (int l = 0; l < 16; ++l) {
*y++ = dl * ((int8_t)((q[l+16] >> shift) & 3) - ((hm[l+16] & m) ? 0 : 4));
}
shift += 2;
m <<= 1;
}
q += 32;
}
}
}
#else
void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k) {
assert(k % QK_K == 0);
assert(QK_K == 64);
const int nb = k / QK_K;
for (int i = 0; i < nb; i++) {
const float d_all = ggml_fp16_to_fp32(x[i].d);
const uint8_t * restrict q = x[i].qs;
const uint8_t * restrict hm = x[i].hmask;
const float d1 = d_all * ((x[i].scales[0] & 0xF) - 8);
const float d2 = d_all * ((x[i].scales[0] >> 4) - 8);
const float d3 = d_all * ((x[i].scales[1] & 0xF) - 8);
const float d4 = d_all * ((x[i].scales[1] >> 4) - 8);
for (int l=0; l<8; ++l) {
uint8_t h = hm[l];
y[l+ 0] = d1 * ((int8_t)((q[l+0] >> 0) & 3) - ((h & 0x01) ? 0 : 4));
y[l+ 8] = d1 * ((int8_t)((q[l+8] >> 0) & 3) - ((h & 0x02) ? 0 : 4));
y[l+16] = d2 * ((int8_t)((q[l+0] >> 2) & 3) - ((h & 0x04) ? 0 : 4));
y[l+24] = d2 * ((int8_t)((q[l+8] >> 2) & 3) - ((h & 0x08) ? 0 : 4));
y[l+32] = d3 * ((int8_t)((q[l+0] >> 4) & 3) - ((h & 0x10) ? 0 : 4));
y[l+40] = d3 * ((int8_t)((q[l+8] >> 4) & 3) - ((h & 0x20) ? 0 : 4));
y[l+48] = d4 * ((int8_t)((q[l+0] >> 6) & 3) - ((h & 0x40) ? 0 : 4));
y[l+56] = d4 * ((int8_t)((q[l+8] >> 6) & 3) - ((h & 0x80) ? 0 : 4));
}
y += QK_K;
}
}
#endif
void quantize_row_q3_K(const float * restrict x, void * restrict vy, int k) {
quantize_row_q3_K_reference(x, vy, k);
}
size_t ggml_quantize_q3_K(const float * restrict src, void * restrict dst, int n, int k, int64_t * restrict hist) {
(void)hist; // TODO: collect histograms
for (int j = 0; j < n; j += k) {
block_q3_K * restrict y = (block_q3_K *)dst + j/QK_K;
quantize_row_q3_K_reference(src + j, y, k);
}
return (n/QK_K*sizeof(block_q3_K));
}
// ====================== 4-bit (de)-quantization
void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict y, int k) {
assert(k % QK_K == 0);
const int nb = k / QK_K;
uint8_t L[QK_K];
uint8_t Laux[32];
float weights[32];
float mins[QK_K/32];
float scales[QK_K/32];
for (int i = 0; i < nb; i++) {
float max_scale = 0; // as we are deducting the min, scales are always positive
float max_min = 0;
for (int j = 0; j < QK_K/32; ++j) {
//scales[j] = make_qkx1_quants(32, 15, x + 32*j, L + 32*j, &mins[j], 9, 0.5f);
float sum_x2 = 0;
for (int l = 0; l < 32; ++l) sum_x2 += x[32*j + l] * x[32*j + l];
float av_x = sqrtf(sum_x2/32);
for (int l = 0; l < 32; ++l) weights[l] = av_x + fabsf(x[32*j + l]);
scales[j] = make_qkx2_quants(32, 15, x + 32*j, weights, L + 32*j, &mins[j], Laux, -1.f, 0.1f, 20, false);
float scale = scales[j];
if (scale > max_scale) {
max_scale = scale;
}
float min = mins[j];
if (min > max_min) {
max_min = min;
}
}
#if QK_K == 256
float inv_scale = max_scale > 0 ? 63.f/max_scale : 0.f;
float inv_min = max_min > 0 ? 63.f/max_min : 0.f;
for (int j = 0; j < QK_K/32; ++j) {
uint8_t ls = nearest_int(inv_scale*scales[j]);
uint8_t lm = nearest_int(inv_min*mins[j]);
ls = MIN(63, ls);
lm = MIN(63, lm);
if (j < 4) {
y[i].scales[j] = ls;
y[i].scales[j+4] = lm;
} else {
y[i].scales[j+4] = (ls & 0xF) | ((lm & 0xF) << 4);
y[i].scales[j-4] |= ((ls >> 4) << 6);
y[i].scales[j-0] |= ((lm >> 4) << 6);
}
}
y[i].d = ggml_fp32_to_fp16(max_scale/63.f);
y[i].dmin = ggml_fp32_to_fp16(max_min/63.f);
uint8_t sc, m;
for (int j = 0; j < QK_K/32; ++j) {
get_scale_min_k4(j, y[i].scales, &sc, &m);
const float d = ggml_fp16_to_fp32(y[i].d) * sc;
if (!d) continue;
const float dm = ggml_fp16_to_fp32(y[i].dmin) * m;
for (int ii = 0; ii < 32; ++ii) {
int l = nearest_int((x[32*j + ii] + dm)/d);
l = MAX(0, MIN(15, l));
L[32*j + ii] = l;
}
}
#else
const float s_factor = 15.f;
float inv_scale = max_scale > 0 ? s_factor/max_scale : 0.f;
float inv_min = max_min > 0 ? s_factor/max_min : 0.f;
int d1 = nearest_int(inv_scale*scales[0]);
int m1 = nearest_int(inv_min*mins[0]);
int d2 = nearest_int(inv_scale*scales[1]);
int m2 = nearest_int(inv_min*mins[1]);
y[i].scales[0] = d1 | (m1 << 4);
y[i].scales[1] = d2 | (m2 << 4);
y[i].d[0] = ggml_fp32_to_fp16(max_scale/s_factor);
y[i].d[1] = ggml_fp32_to_fp16(max_min/s_factor);
float sumlx = 0;
int suml2 = 0;
for (int j = 0; j < QK_K/32; ++j) {
const uint8_t sd = y[i].scales[j] & 0xF;
const uint8_t sm = y[i].scales[j] >> 4;
const float d = ggml_fp16_to_fp32(y[i].d[0]) * sd;
if (!d) continue;
const float m = ggml_fp16_to_fp32(y[i].d[1]) * sm;
for (int ii = 0; ii < 32; ++ii) {
int l = nearest_int((x[32*j + ii] + m)/d);
l = MAX(0, MIN(15, l));
L[32*j + ii] = l;
sumlx += (x[32*j + ii] + m)*l*sd;
suml2 += l*l*sd*sd;
}
}
if (suml2) {
y[i].d[0] = ggml_fp32_to_fp16(sumlx/suml2);
}
#endif
uint8_t * q = y[i].qs;
for (int j = 0; j < QK_K; j += 64) {
for (int l = 0; l < 32; ++l) q[l] = L[j + l] | (L[j + l + 32] << 4);
q += 32;
}
x += QK_K;
}
}
void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int k) {
assert(k % QK_K == 0);
const int nb = k / QK_K;
for (int i = 0; i < nb; i++) {
const uint8_t * q = x[i].qs;
#if QK_K == 256
const float d = ggml_fp16_to_fp32(x[i].d);
const float min = ggml_fp16_to_fp32(x[i].dmin);
int is = 0;
uint8_t sc, m;
for (int j = 0; j < QK_K; j += 64) {
get_scale_min_k4(is + 0, x[i].scales, &sc, &m);
const float d1 = d * sc; const float m1 = min * m;
get_scale_min_k4(is + 1, x[i].scales, &sc, &m);
const float d2 = d * sc; const float m2 = min * m;
for (int l = 0; l < 32; ++l) *y++ = d1 * (q[l] & 0xF) - m1;
for (int l = 0; l < 32; ++l) *y++ = d2 * (q[l] >> 4) - m2;
q += 32; is += 2;
}
#else
const float dall = ggml_fp16_to_fp32(x[i].d[0]);
const float mall = ggml_fp16_to_fp32(x[i].d[1]);
const float d1 = dall * (x[i].scales[0] & 0xF), m1 = mall * (x[i].scales[0] >> 4);
const float d2 = dall * (x[i].scales[1] & 0xF), m2 = mall * (x[i].scales[1] >> 4);
for (int l = 0; l < 32; ++l) {
y[l+ 0] = d1 * (q[l] & 0xF) - m1;
y[l+32] = d2 * (q[l] >> 4) - m2;
}
y += QK_K;
#endif
}
}
void quantize_row_q4_K(const float * restrict x, void * restrict vy, int k) {
assert(k % QK_K == 0);
block_q4_K * restrict y = vy;
quantize_row_q4_K_reference(x, y, k);
}
size_t ggml_quantize_q4_K(const float * restrict src, void * restrict dst, int n, int k, int64_t * restrict hist) {
assert(k % QK_K == 0);
(void)hist; // TODO: collect histograms
for (int j = 0; j < n; j += k) {
block_q4_K * restrict y = (block_q4_K *)dst + j/QK_K;
quantize_row_q4_K_reference(src + j, y, k);
}
return (n/QK_K*sizeof(block_q4_K));
}
// ====================== 5-bit (de)-quantization
void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict y, int k) {
assert(k % QK_K == 0);
const int nb = k / QK_K;
#if QK_K == 256
uint8_t L[QK_K];
float mins[QK_K/32];
float scales[QK_K/32];
float weights[32];
uint8_t Laux[32];
#else
int8_t L[QK_K];
float scales[QK_K/16];
#endif
for (int i = 0; i < nb; i++) {
#if QK_K == 256
float max_scale = 0; // as we are deducting the min, scales are always positive
float max_min = 0;
for (int j = 0; j < QK_K/32; ++j) {
//scales[j] = make_qkx1_quants(32, 31, x + 32*j, L + 32*j, &mins[j], 9, 0.5f);
float sum_x2 = 0;
for (int l = 0; l < 32; ++l) sum_x2 += x[32*j + l] * x[32*j + l];
float av_x = sqrtf(sum_x2/32);
for (int l = 0; l < 32; ++l) weights[l] = av_x + fabsf(x[32*j + l]);
scales[j] = make_qkx2_quants(32, 31, x + 32*j, weights, L + 32*j, &mins[j], Laux, -0.5f, 0.1f, 15, false);
float scale = scales[j];
if (scale > max_scale) {
max_scale = scale;
}
float min = mins[j];
if (min > max_min) {
max_min = min;
}
}
float inv_scale = max_scale > 0 ? 63.f/max_scale : 0.f;
float inv_min = max_min > 0 ? 63.f/max_min : 0.f;
for (int j = 0; j < QK_K/32; ++j) {
uint8_t ls = nearest_int(inv_scale*scales[j]);
uint8_t lm = nearest_int(inv_min*mins[j]);
ls = MIN(63, ls);
lm = MIN(63, lm);
if (j < 4) {
y[i].scales[j] = ls;
y[i].scales[j+4] = lm;
} else {
y[i].scales[j+4] = (ls & 0xF) | ((lm & 0xF) << 4);
y[i].scales[j-4] |= ((ls >> 4) << 6);
y[i].scales[j-0] |= ((lm >> 4) << 6);
}
}
y[i].d = ggml_fp32_to_fp16(max_scale/63.f);
y[i].dmin = ggml_fp32_to_fp16(max_min/63.f);
uint8_t sc, m;
for (int j = 0; j < QK_K/32; ++j) {
get_scale_min_k4(j, y[i].scales, &sc, &m);
const float d = ggml_fp16_to_fp32(y[i].d) * sc;
if (!d) continue;
const float dm = ggml_fp16_to_fp32(y[i].dmin) * m;
for (int ii = 0; ii < 32; ++ii) {
int l = nearest_int((x[32*j + ii] + dm)/d);
l = MAX(0, MIN(31, l));
L[32*j + ii] = l;
}
}
uint8_t * restrict qh = y[i].qh;
uint8_t * restrict ql = y[i].qs;
memset(qh, 0, QK_K/8);
uint8_t m1 = 1, m2 = 2;
for (int n = 0; n < QK_K; n += 64) {
for (int j = 0; j < 32; ++j) {
int l1 = L[n + j];
if (l1 > 15) {
l1 -= 16; qh[j] |= m1;
}
int l2 = L[n + j + 32];
if (l2 > 15) {
l2 -= 16; qh[j] |= m2;
}
ql[j] = l1 | (l2 << 4);
}
m1 <<= 2; m2 <<= 2;
ql += 32;
}
#else
float max_scale = 0, amax = 0;
for (int j = 0; j < QK_K/16; ++j) {
scales[j] = make_qx_quants(16, 16, x + 16*j, L + 16*j, 1);
float abs_scale = fabsf(scales[j]);
if (abs_scale > amax) {
amax = abs_scale;
max_scale = scales[j];
}
}
float iscale = -128.f/max_scale;
for (int j = 0; j < QK_K/16; ++j) {
int l = nearest_int(iscale*scales[j]);
y[i].scales[j] = MAX(-128, MIN(127, l));
}
y[i].d = ggml_fp32_to_fp16(1/iscale);
for (int j = 0; j < QK_K/16; ++j) {
const float d = ggml_fp16_to_fp32(y[i].d) * y[i].scales[j];
if (!d) continue;
for (int ii = 0; ii < 16; ++ii) {
int l = nearest_int(x[16*j + ii]/d);
l = MAX(-16, MIN(15, l));
L[16*j + ii] = l + 16;
}
}
uint8_t * restrict qh = y[i].qh;
uint8_t * restrict ql = y[i].qs;
memset(qh, 0, QK_K/8);
for (int j = 0; j < 32; ++j) {
int jm = j%8;
int is = j/8;
int l1 = L[j];
if (l1 > 15) {
l1 -= 16; qh[jm] |= (1 << is);
}
int l2 = L[j + 32];
if (l2 > 15) {
l2 -= 16; qh[jm] |= (1 << (4 + is));
}
ql[j] = l1 | (l2 << 4);
}
#endif
x += QK_K;
}
}
void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int k) {
assert(k % QK_K == 0);
const int nb = k / QK_K;
for (int i = 0; i < nb; i++) {
const uint8_t * ql = x[i].qs;
const uint8_t * qh = x[i].qh;
#if QK_K == 256
const float d = ggml_fp16_to_fp32(x[i].d);
const float min = ggml_fp16_to_fp32(x[i].dmin);
int is = 0;
uint8_t sc, m;
uint8_t u1 = 1, u2 = 2;
for (int j = 0; j < QK_K; j += 64) {
get_scale_min_k4(is + 0, x[i].scales, &sc, &m);
const float d1 = d * sc; const float m1 = min * m;
get_scale_min_k4(is + 1, x[i].scales, &sc, &m);
const float d2 = d * sc; const float m2 = min * m;
for (int l = 0; l < 32; ++l) *y++ = d1 * ((ql[l] & 0xF) + (qh[l] & u1 ? 16 : 0)) - m1;
for (int l = 0; l < 32; ++l) *y++ = d2 * ((ql[l] >> 4) + (qh[l] & u2 ? 16 : 0)) - m2;
ql += 32; is += 2;
u1 <<= 2; u2 <<= 2;
}
#else
float d = ggml_fp16_to_fp32(x[i].d);
const int8_t * restrict s = x[i].scales;
for (int l = 0; l < 8; ++l) {
y[l+ 0] = d * s[0] * ((ql[l+ 0] & 0xF) - (qh[l] & 0x01 ? 0 : 16));
y[l+ 8] = d * s[0] * ((ql[l+ 8] & 0xF) - (qh[l] & 0x02 ? 0 : 16));
y[l+16] = d * s[1] * ((ql[l+16] & 0xF) - (qh[l] & 0x04 ? 0 : 16));
y[l+24] = d * s[1] * ((ql[l+24] & 0xF) - (qh[l] & 0x08 ? 0 : 16));
y[l+32] = d * s[2] * ((ql[l+ 0] >> 4) - (qh[l] & 0x10 ? 0 : 16));
y[l+40] = d * s[2] * ((ql[l+ 8] >> 4) - (qh[l] & 0x20 ? 0 : 16));
y[l+48] = d * s[3] * ((ql[l+16] >> 4) - (qh[l] & 0x40 ? 0 : 16));
y[l+56] = d * s[3] * ((ql[l+24] >> 4) - (qh[l] & 0x80 ? 0 : 16));
}
y += QK_K;
#endif
}
}
void quantize_row_q5_K(const float * restrict x, void * restrict vy, int k) {
assert(k % QK_K == 0);
block_q5_K * restrict y = vy;
quantize_row_q5_K_reference(x, y, k);
}
size_t ggml_quantize_q5_K(const float * restrict src, void * restrict dst, int n, int k, int64_t * restrict hist) {
assert(k % QK_K == 0);
(void)hist; // TODO: collect histograms
for (int j = 0; j < n; j += k) {
block_q5_K * restrict y = (block_q5_K *)dst + j/QK_K;
quantize_row_q5_K_reference(src + j, y, k);
}
return (n/QK_K*sizeof(block_q5_K));
}
// ====================== 6-bit (de)-quantization
void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int k) {
assert(k % QK_K == 0);
const int nb = k / QK_K;
int8_t L[QK_K];
float scales[QK_K/16];
for (int i = 0; i < nb; i++) {
float max_scale = 0;
float max_abs_scale = 0;
for (int ib = 0; ib < QK_K/16; ++ib) {
const float scale = make_qx_quants(16, 32, x + 16*ib, L + 16*ib, 1);
scales[ib] = scale;
const float abs_scale = fabsf(scale);
if (abs_scale > max_abs_scale) {
max_abs_scale = abs_scale;
max_scale = scale;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
if (!max_abs_scale) {
memset(&y[i], 0, sizeof(block_q6_K));
y[i].d = ggml_fp32_to_fp16(0.f);
x += QK_K;
continue;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
float iscale = -128.f/max_scale;
y[i].d = ggml_fp32_to_fp16(1/iscale);
for (int ib = 0; ib < QK_K/16; ++ib) {
y[i].scales[ib] = MIN(127, nearest_int(iscale*scales[ib]));
Quantization imrovements for k_quants (#2707) * Improve LLaMA-2 2-, 3- and 4-bit quantization * Q3_K_S: use Q5_K for 1st 2 layers of attention.wv and feed_forward.w2 * Q4_K_S: use Q6_K for 1st 2 layers of attention.wv and feed_forward.w2 * Q2_K and Q3_K_M: use Q5_K instead of Q4_K for 1st 2 layers of attention.wv and feed_forward.w2 This leads to a slight model sized increase as follows: Q2_K : 2.684G vs 2.670G Q3_K_S: 2.775G vs 2.745G Q3_K_M: 3.071G vs 3.057G Q4_K_S: 3.592G vs 3.563G LLaMA-2 PPL for context 512 changes as follows: Q2_K : 6.6691 vs 6.8201 Q3_K_S: 6.2129 vs 6.2584 Q3_K_M: 6.0387 vs 6.1371 Q4_K_S: 5.9138 vs 6.0041 There are improvements for LLaMA-1 as well, but they are way smaller than the above. * Minor 4-bit quantization improvement For the same model size as previus commit, we get PPL = 5.9069 vs 5.9138. * Some more fine tuning * Adding make_qkx2_quants With it, we get PPL = 5.8828 for L2-7B Q4_K_S. * Another minor improvement * Q2_K improvement Smaller model, lower perplexity. 7B: file size = 2.632G, PPL = 6.3772 vs original 2.670G PPL = 6.8201 12B: file size = 5.056G, PPL = 5.4577 vs original 5.130G PPL = 5.7178 It is mostly Q3_K except for tok_embeddings, attention.wq, attention.wk, which are Q2_K * Iterating * Revert Q5_K back to make_qkx1_quants * Better Q6_K * make_qkx2_quants is better for Q5_K after all * Fix after rebasing on master * Fix for changed tensor names --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-08-22 16:14:09 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
for (int j = 0; j < QK_K/16; ++j) {
float d = ggml_fp16_to_fp32(y[i].d) * y[i].scales[j];
if (!d) {
continue;
Quantization imrovements for k_quants (#2707) * Improve LLaMA-2 2-, 3- and 4-bit quantization * Q3_K_S: use Q5_K for 1st 2 layers of attention.wv and feed_forward.w2 * Q4_K_S: use Q6_K for 1st 2 layers of attention.wv and feed_forward.w2 * Q2_K and Q3_K_M: use Q5_K instead of Q4_K for 1st 2 layers of attention.wv and feed_forward.w2 This leads to a slight model sized increase as follows: Q2_K : 2.684G vs 2.670G Q3_K_S: 2.775G vs 2.745G Q3_K_M: 3.071G vs 3.057G Q4_K_S: 3.592G vs 3.563G LLaMA-2 PPL for context 512 changes as follows: Q2_K : 6.6691 vs 6.8201 Q3_K_S: 6.2129 vs 6.2584 Q3_K_M: 6.0387 vs 6.1371 Q4_K_S: 5.9138 vs 6.0041 There are improvements for LLaMA-1 as well, but they are way smaller than the above. * Minor 4-bit quantization improvement For the same model size as previus commit, we get PPL = 5.9069 vs 5.9138. * Some more fine tuning * Adding make_qkx2_quants With it, we get PPL = 5.8828 for L2-7B Q4_K_S. * Another minor improvement * Q2_K improvement Smaller model, lower perplexity. 7B: file size = 2.632G, PPL = 6.3772 vs original 2.670G PPL = 6.8201 12B: file size = 5.056G, PPL = 5.4577 vs original 5.130G PPL = 5.7178 It is mostly Q3_K except for tok_embeddings, attention.wq, attention.wk, which are Q2_K * Iterating * Revert Q5_K back to make_qkx1_quants * Better Q6_K * make_qkx2_quants is better for Q5_K after all * Fix after rebasing on master * Fix for changed tensor names --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-08-22 16:14:09 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
for (int ii = 0; ii < 16; ++ii) {
int l = nearest_int(x[16*j + ii]/d);
l = MAX(-32, MIN(31, l));
L[16*j + ii] = l + 32;
Quantization imrovements for k_quants (#2707) * Improve LLaMA-2 2-, 3- and 4-bit quantization * Q3_K_S: use Q5_K for 1st 2 layers of attention.wv and feed_forward.w2 * Q4_K_S: use Q6_K for 1st 2 layers of attention.wv and feed_forward.w2 * Q2_K and Q3_K_M: use Q5_K instead of Q4_K for 1st 2 layers of attention.wv and feed_forward.w2 This leads to a slight model sized increase as follows: Q2_K : 2.684G vs 2.670G Q3_K_S: 2.775G vs 2.745G Q3_K_M: 3.071G vs 3.057G Q4_K_S: 3.592G vs 3.563G LLaMA-2 PPL for context 512 changes as follows: Q2_K : 6.6691 vs 6.8201 Q3_K_S: 6.2129 vs 6.2584 Q3_K_M: 6.0387 vs 6.1371 Q4_K_S: 5.9138 vs 6.0041 There are improvements for LLaMA-1 as well, but they are way smaller than the above. * Minor 4-bit quantization improvement For the same model size as previus commit, we get PPL = 5.9069 vs 5.9138. * Some more fine tuning * Adding make_qkx2_quants With it, we get PPL = 5.8828 for L2-7B Q4_K_S. * Another minor improvement * Q2_K improvement Smaller model, lower perplexity. 7B: file size = 2.632G, PPL = 6.3772 vs original 2.670G PPL = 6.8201 12B: file size = 5.056G, PPL = 5.4577 vs original 5.130G PPL = 5.7178 It is mostly Q3_K except for tok_embeddings, attention.wq, attention.wk, which are Q2_K * Iterating * Revert Q5_K back to make_qkx1_quants * Better Q6_K * make_qkx2_quants is better for Q5_K after all * Fix after rebasing on master * Fix for changed tensor names --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-08-22 16:14:09 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
}
uint8_t * restrict ql = y[i].ql;
uint8_t * restrict qh = y[i].qh;
#if QK_K == 256
for (int j = 0; j < QK_K; j += 128) {
for (int l = 0; l < 32; ++l) {
const uint8_t q1 = L[j + l + 0] & 0xF;
const uint8_t q2 = L[j + l + 32] & 0xF;
const uint8_t q3 = L[j + l + 64] & 0xF;
const uint8_t q4 = L[j + l + 96] & 0xF;
ql[l+ 0] = q1 | (q3 << 4);
ql[l+32] = q2 | (q4 << 4);
qh[l] = (L[j + l] >> 4) | ((L[j + l + 32] >> 4) << 2) | ((L[j + l + 64] >> 4) << 4) | ((L[j + l + 96] >> 4) << 6);
Quantization imrovements for k_quants (#2707) * Improve LLaMA-2 2-, 3- and 4-bit quantization * Q3_K_S: use Q5_K for 1st 2 layers of attention.wv and feed_forward.w2 * Q4_K_S: use Q6_K for 1st 2 layers of attention.wv and feed_forward.w2 * Q2_K and Q3_K_M: use Q5_K instead of Q4_K for 1st 2 layers of attention.wv and feed_forward.w2 This leads to a slight model sized increase as follows: Q2_K : 2.684G vs 2.670G Q3_K_S: 2.775G vs 2.745G Q3_K_M: 3.071G vs 3.057G Q4_K_S: 3.592G vs 3.563G LLaMA-2 PPL for context 512 changes as follows: Q2_K : 6.6691 vs 6.8201 Q3_K_S: 6.2129 vs 6.2584 Q3_K_M: 6.0387 vs 6.1371 Q4_K_S: 5.9138 vs 6.0041 There are improvements for LLaMA-1 as well, but they are way smaller than the above. * Minor 4-bit quantization improvement For the same model size as previus commit, we get PPL = 5.9069 vs 5.9138. * Some more fine tuning * Adding make_qkx2_quants With it, we get PPL = 5.8828 for L2-7B Q4_K_S. * Another minor improvement * Q2_K improvement Smaller model, lower perplexity. 7B: file size = 2.632G, PPL = 6.3772 vs original 2.670G PPL = 6.8201 12B: file size = 5.056G, PPL = 5.4577 vs original 5.130G PPL = 5.7178 It is mostly Q3_K except for tok_embeddings, attention.wq, attention.wk, which are Q2_K * Iterating * Revert Q5_K back to make_qkx1_quants * Better Q6_K * make_qkx2_quants is better for Q5_K after all * Fix after rebasing on master * Fix for changed tensor names --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-08-22 16:14:09 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
ql += 64;
qh += 32;
}
#else
for (int l = 0; l < 32; ++l) {
const uint8_t q1 = L[l + 0] & 0xF;
const uint8_t q2 = L[l + 32] & 0xF;
ql[l] = q1 | (q2 << 4);
}
for (int l = 0; l < 16; ++l) {
qh[l] = (L[l] >> 4) | ((L[l + 16] >> 4) << 2) | ((L[l + 32] >> 4) << 4) | ((L[l + 48] >> 4) << 6);
Quantization imrovements for k_quants (#2707) * Improve LLaMA-2 2-, 3- and 4-bit quantization * Q3_K_S: use Q5_K for 1st 2 layers of attention.wv and feed_forward.w2 * Q4_K_S: use Q6_K for 1st 2 layers of attention.wv and feed_forward.w2 * Q2_K and Q3_K_M: use Q5_K instead of Q4_K for 1st 2 layers of attention.wv and feed_forward.w2 This leads to a slight model sized increase as follows: Q2_K : 2.684G vs 2.670G Q3_K_S: 2.775G vs 2.745G Q3_K_M: 3.071G vs 3.057G Q4_K_S: 3.592G vs 3.563G LLaMA-2 PPL for context 512 changes as follows: Q2_K : 6.6691 vs 6.8201 Q3_K_S: 6.2129 vs 6.2584 Q3_K_M: 6.0387 vs 6.1371 Q4_K_S: 5.9138 vs 6.0041 There are improvements for LLaMA-1 as well, but they are way smaller than the above. * Minor 4-bit quantization improvement For the same model size as previus commit, we get PPL = 5.9069 vs 5.9138. * Some more fine tuning * Adding make_qkx2_quants With it, we get PPL = 5.8828 for L2-7B Q4_K_S. * Another minor improvement * Q2_K improvement Smaller model, lower perplexity. 7B: file size = 2.632G, PPL = 6.3772 vs original 2.670G PPL = 6.8201 12B: file size = 5.056G, PPL = 5.4577 vs original 5.130G PPL = 5.7178 It is mostly Q3_K except for tok_embeddings, attention.wq, attention.wk, which are Q2_K * Iterating * Revert Q5_K back to make_qkx1_quants * Better Q6_K * make_qkx2_quants is better for Q5_K after all * Fix after rebasing on master * Fix for changed tensor names --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-08-22 16:14:09 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
#endif
x += QK_K;
Quantization imrovements for k_quants (#2707) * Improve LLaMA-2 2-, 3- and 4-bit quantization * Q3_K_S: use Q5_K for 1st 2 layers of attention.wv and feed_forward.w2 * Q4_K_S: use Q6_K for 1st 2 layers of attention.wv and feed_forward.w2 * Q2_K and Q3_K_M: use Q5_K instead of Q4_K for 1st 2 layers of attention.wv and feed_forward.w2 This leads to a slight model sized increase as follows: Q2_K : 2.684G vs 2.670G Q3_K_S: 2.775G vs 2.745G Q3_K_M: 3.071G vs 3.057G Q4_K_S: 3.592G vs 3.563G LLaMA-2 PPL for context 512 changes as follows: Q2_K : 6.6691 vs 6.8201 Q3_K_S: 6.2129 vs 6.2584 Q3_K_M: 6.0387 vs 6.1371 Q4_K_S: 5.9138 vs 6.0041 There are improvements for LLaMA-1 as well, but they are way smaller than the above. * Minor 4-bit quantization improvement For the same model size as previus commit, we get PPL = 5.9069 vs 5.9138. * Some more fine tuning * Adding make_qkx2_quants With it, we get PPL = 5.8828 for L2-7B Q4_K_S. * Another minor improvement * Q2_K improvement Smaller model, lower perplexity. 7B: file size = 2.632G, PPL = 6.3772 vs original 2.670G PPL = 6.8201 12B: file size = 5.056G, PPL = 5.4577 vs original 5.130G PPL = 5.7178 It is mostly Q3_K except for tok_embeddings, attention.wq, attention.wk, which are Q2_K * Iterating * Revert Q5_K back to make_qkx1_quants * Better Q6_K * make_qkx2_quants is better for Q5_K after all * Fix after rebasing on master * Fix for changed tensor names --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-08-22 16:14:09 +00:00
}
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int k) {
assert(k % QK_K == 0);
const int nb = k / QK_K;
for (int i = 0; i < nb; i++) {
const float d = ggml_fp16_to_fp32(x[i].d);
const uint8_t * restrict ql = x[i].ql;
const uint8_t * restrict qh = x[i].qh;
const int8_t * restrict sc = x[i].scales;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
#if QK_K == 256
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
for (int n = 0; n < QK_K; n += 128) {
for (int l = 0; l < 32; ++l) {
int is = l/16;
const int8_t q1 = (int8_t)((ql[l + 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
const int8_t q2 = (int8_t)((ql[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
const int8_t q3 = (int8_t)((ql[l + 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32;
const int8_t q4 = (int8_t)((ql[l + 32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32;
y[l + 0] = d * sc[is + 0] * q1;
y[l + 32] = d * sc[is + 2] * q2;
y[l + 64] = d * sc[is + 4] * q3;
y[l + 96] = d * sc[is + 6] * q4;
}
y += 128;
ql += 64;
qh += 32;
sc += 8;
}
#else
for (int l = 0; l < 16; ++l) {
const int8_t q1 = (int8_t)((ql[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
const int8_t q2 = (int8_t)((ql[l+16] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
const int8_t q3 = (int8_t)((ql[l+ 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32;
const int8_t q4 = (int8_t)((ql[l+16] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32;
y[l+ 0] = d * sc[0] * q1;
y[l+16] = d * sc[1] * q2;
y[l+32] = d * sc[2] * q3;
y[l+48] = d * sc[3] * q4;
}
y += 64;
#endif
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
void quantize_row_q6_K(const float * restrict x, void * restrict vy, int k) {
assert(k % QK_K == 0);
block_q6_K * restrict y = vy;
quantize_row_q6_K_reference(x, y, k);
}
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist) {
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
assert(k % QK_K == 0);
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
(void)hist; // TODO: collect histograms
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
for (int j = 0; j < n; j += k) {
block_q6_K * restrict y = (block_q6_K *)dst + j/QK_K;
quantize_row_q6_K_reference(src + j, y, k);
}
return (n/QK_K*sizeof(block_q6_K));
}
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
//===================================== Q8_K ==============================================
void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k) {
assert(k % QK_K == 0);
const int nb = k / QK_K;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
for (int i = 0; i < nb; i++) {
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
float max = 0;
float amax = 0;
for (int j = 0; j < QK_K; ++j) {
float ax = fabsf(x[j]);
if (ax > amax) {
amax = ax; max = x[j];
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
if (!amax) {
y[i].d = 0;
memset(y[i].qs, 0, QK_K);
x += QK_K;
continue;
}
const float iscale = -128.f/max;
for (int j = 0; j < QK_K; ++j) {
int v = nearest_int(iscale*x[j]);
y[i].qs[j] = MIN(127, v);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
for (int j = 0; j < QK_K/16; ++j) {
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
int sum = 0;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
for (int ii = 0; ii < 16; ++ii) {
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
sum += y[i].qs[j*16 + ii];
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
y[i].bsums[j] = sum;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
y[i].d = 1/iscale;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
x += QK_K;
}
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k) {
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
assert(k % QK_K == 0);
const int nb = k / QK_K;
for (int i = 0; i < nb; i++) {
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
for (int j = 0; j < QK_K; ++j) {
*y++ = x[i].d * x[i].qs[j];
}
}
}
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
void quantize_row_q8_K(const float * restrict x, void * restrict y, int k) {
quantize_row_q8_K_reference(x, y, k);
}
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
//===================================== Dot ptoducts =================================
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
//
// Helper functions
//
#if __AVX__ || __AVX2__ || __AVX512F__
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// shuffles to pick the required scales in dot products
static inline __m256i get_scale_shuffle_q3k(int i) {
static const uint8_t k_shuffle[128] = {
0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,
12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13, 14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,
};
return _mm256_loadu_si256((const __m256i*)k_shuffle + i);
}
static inline __m256i get_scale_shuffle_k4(int i) {
static const uint8_t k_shuffle[256] = {
0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,
2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5,
6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9,
10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,
12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,
14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15
};
return _mm256_loadu_si256((const __m256i*)k_shuffle + i);
}
static inline __m128i get_scale_shuffle(int i) {
static const uint8_t k_shuffle[128] = {
0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1,
2, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3,
4, 4, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 5, 5,
6, 6, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 7, 7,
8, 8, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 9, 9,
10,10,10,10,10,10,10,10, 11,11,11,11,11,11,11,11,
12,12,12,12,12,12,12,12, 13,13,13,13,13,13,13,13,
14,14,14,14,14,14,14,14, 15,15,15,15,15,15,15,15
};
return _mm_loadu_si128((const __m128i*)k_shuffle + i);
}
#endif
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
const int qk = QK8_0;
const int nb = n / qk;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
assert(n % qk == 0);
const block_q4_0 * restrict x = vx;
const block_q8_0 * restrict y = vy;
#if defined(__ARM_NEON)
float32x4_t sumv0 = vdupq_n_f32(0.0f);
float32x4_t sumv1 = vdupq_n_f32(0.0f);
assert(nb % 2 == 0); // TODO: handle odd nb
for (int i = 0; i < nb; i += 2) {
const block_q4_0 * restrict x0 = &x[i + 0];
const block_q4_0 * restrict x1 = &x[i + 1];
const block_q8_0 * restrict y0 = &y[i + 0];
const block_q8_0 * restrict y1 = &y[i + 1];
const uint8x16_t m4b = vdupq_n_u8(0x0F);
const int8x16_t s8b = vdupq_n_s8(0x8);
const uint8x16_t v0_0 = vld1q_u8(x0->qs);
const uint8x16_t v0_1 = vld1q_u8(x1->qs);
// 4-bit -> 8-bit
const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b));
const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b));
const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
// sub 8
const int8x16_t v0_0ls = vsubq_s8(v0_0l, s8b);
const int8x16_t v0_0hs = vsubq_s8(v0_0h, s8b);
const int8x16_t v0_1ls = vsubq_s8(v0_1l, s8b);
const int8x16_t v0_1hs = vsubq_s8(v0_1h, s8b);
// load y
const int8x16_t v1_0l = vld1q_s8(y0->qs);
const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
const int8x16_t v1_1l = vld1q_s8(y1->qs);
const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
#if defined(__ARM_FEATURE_DOTPROD)
// dot product into int32x4_t
const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0l), v0_0hs, v1_0h);
const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1l), v0_1hs, v1_1h);
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), ggml_fp16_to_fp32(x0->d)*ggml_fp16_to_fp32(y0->d));
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), ggml_fp16_to_fp32(x1->d)*ggml_fp16_to_fp32(y1->d));
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
#else
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0ls), vget_low_s8 (v1_0l));
const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0ls), vget_high_s8(v1_0l));
const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hs), vget_low_s8 (v1_0h));
const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hs), vget_high_s8(v1_0h));
const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1ls), vget_low_s8 (v1_1l));
const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1ls), vget_high_s8(v1_1l));
const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hs), vget_low_s8 (v1_1h));
const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hs), vget_high_s8(v1_1h));
const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h));
const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h));
const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h));
const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h));
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), ggml_fp16_to_fp32(x0->d)*ggml_fp16_to_fp32(y0->d));
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), ggml_fp16_to_fp32(x1->d)*ggml_fp16_to_fp32(y1->d));
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
#endif
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
*s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
#elif defined(__AVX2__)
// Initialize accumulator with zeros
__m256 acc = _mm256_setzero_ps();
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// Main loop
for (int i = 0; i < nb; ++i) {
/* Compute combined scale for the block */
const __m256 d = _mm256_set1_ps( ggml_fp16_to_fp32(x[i].d) * ggml_fp16_to_fp32(y[i].d) );
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
__m256i bx = bytes_from_nibbles_32(x[i].qs);
// Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval.
const __m256i off = _mm256_set1_epi8( 8 );
bx = _mm256_sub_epi8( bx, off );
__m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
const __m256 q = mul_sum_i8_pairs_float(bx, by);
/* Multiply q with scale and accumulate */
acc = _mm256_fmadd_ps( d, q, acc );
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
*s = hsum_float_8(acc);
#elif defined(__AVX__)
// Initialize accumulator with zeros
__m256 acc = _mm256_setzero_ps();
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// Main loop
for (int i = 0; i < nb; ++i) {
// Compute combined scale for the block
const __m256 d = _mm256_set1_ps( ggml_fp16_to_fp32(x[i].d) * ggml_fp16_to_fp32(y[i].d) );
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
const __m128i lowMask = _mm_set1_epi8(0xF);
const __m128i off = _mm_set1_epi8(8);
const __m128i tmp = _mm_loadu_si128((const __m128i *)x[i].qs);
__m128i bx = _mm_and_si128(lowMask, tmp);
__m128i by = _mm_loadu_si128((const __m128i *)y[i].qs);
bx = _mm_sub_epi8(bx, off);
const __m128i i32_0 = mul_sum_i8_pairs(bx, by);
bx = _mm_and_si128(lowMask, _mm_srli_epi64(tmp, 4));
by = _mm_loadu_si128((const __m128i *)(y[i].qs + 16));
bx = _mm_sub_epi8(bx, off);
const __m128i i32_1 = mul_sum_i8_pairs(bx, by);
// Convert int32_t to float
__m256 p = _mm256_cvtepi32_ps(MM256_SET_M128I(i32_0, i32_1));
// Apply the scale, and accumulate
acc = _mm256_add_ps(_mm256_mul_ps( d, p ), acc);
}
*s = hsum_float_8(acc);
#elif defined(__SSSE3__)
// set constants
const __m128i lowMask = _mm_set1_epi8(0xF);
const __m128i off = _mm_set1_epi8(8);
// Initialize accumulator with zeros
__m128 acc_0 = _mm_setzero_ps();
__m128 acc_1 = _mm_setzero_ps();
__m128 acc_2 = _mm_setzero_ps();
__m128 acc_3 = _mm_setzero_ps();
// First round without accumulation
{
_mm_prefetch(&x[0] + sizeof(block_q4_0), _MM_HINT_T0);
_mm_prefetch(&y[0] + sizeof(block_q8_0), _MM_HINT_T0);
// Compute combined scale for the block 0 and 1
const __m128 d_0_1 = _mm_set1_ps( ggml_fp16_to_fp32(x[0].d) * ggml_fp16_to_fp32(y[0].d) );
const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[0].qs);
__m128i bx_0 = _mm_and_si128(lowMask, tmp_0_1);
__m128i by_0 = _mm_loadu_si128((const __m128i *)y[0].qs);
bx_0 = _mm_sub_epi8(bx_0, off);
const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0);
__m128i bx_1 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_0_1, 4));
__m128i by_1 = _mm_loadu_si128((const __m128i *)(y[0].qs + 16));
bx_1 = _mm_sub_epi8(bx_1, off);
const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1);
_mm_prefetch(&x[1] + sizeof(block_q4_0), _MM_HINT_T0);
_mm_prefetch(&y[1] + sizeof(block_q8_0), _MM_HINT_T0);
// Compute combined scale for the block 2 and 3
const __m128 d_2_3 = _mm_set1_ps( ggml_fp16_to_fp32(x[1].d) * ggml_fp16_to_fp32(y[1].d) );
const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[1].qs);
__m128i bx_2 = _mm_and_si128(lowMask, tmp_2_3);
__m128i by_2 = _mm_loadu_si128((const __m128i *)y[1].qs);
bx_2 = _mm_sub_epi8(bx_2, off);
const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2);
__m128i bx_3 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_2_3, 4));
__m128i by_3 = _mm_loadu_si128((const __m128i *)(y[1].qs + 16));
bx_3 = _mm_sub_epi8(bx_3, off);
const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3);
// Convert int32_t to float
__m128 p0 = _mm_cvtepi32_ps(i32_0);
__m128 p1 = _mm_cvtepi32_ps(i32_1);
__m128 p2 = _mm_cvtepi32_ps(i32_2);
__m128 p3 = _mm_cvtepi32_ps(i32_3);
// Apply the scale
acc_0 = _mm_mul_ps( d_0_1, p0 );
acc_1 = _mm_mul_ps( d_0_1, p1 );
acc_2 = _mm_mul_ps( d_2_3, p2 );
acc_3 = _mm_mul_ps( d_2_3, p3 );
}
assert(nb % 2 == 0); // TODO: handle odd nb
// Main loop
for (int i = 2; i < nb; i+=2) {
_mm_prefetch(&x[i] + sizeof(block_q4_0), _MM_HINT_T0);
_mm_prefetch(&y[i] + sizeof(block_q8_0), _MM_HINT_T0);
// Compute combined scale for the block 0 and 1
const __m128 d_0_1 = _mm_set1_ps( ggml_fp16_to_fp32(x[i].d) * ggml_fp16_to_fp32(y[i].d) );
const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[i].qs);
__m128i bx_0 = _mm_and_si128(lowMask, tmp_0_1);
__m128i by_0 = _mm_loadu_si128((const __m128i *)y[i].qs);
bx_0 = _mm_sub_epi8(bx_0, off);
const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0);
__m128i bx_1 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_0_1, 4));
__m128i by_1 = _mm_loadu_si128((const __m128i *)(y[i].qs + 16));
bx_1 = _mm_sub_epi8(bx_1, off);
const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1);
_mm_prefetch(&x[i] + 2 * sizeof(block_q4_0), _MM_HINT_T0);
_mm_prefetch(&y[i] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
// Compute combined scale for the block 2 and 3
const __m128 d_2_3 = _mm_set1_ps( ggml_fp16_to_fp32(x[i + 1].d) * ggml_fp16_to_fp32(y[i + 1].d) );
const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[i + 1].qs);
__m128i bx_2 = _mm_and_si128(lowMask, tmp_2_3);
__m128i by_2 = _mm_loadu_si128((const __m128i *)y[i + 1].qs);
bx_2 = _mm_sub_epi8(bx_2, off);
const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2);
__m128i bx_3 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_2_3, 4));
__m128i by_3 = _mm_loadu_si128((const __m128i *)(y[i + 1].qs + 16));
bx_3 = _mm_sub_epi8(bx_3, off);
const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3);
// Convert int32_t to float
__m128 p0 = _mm_cvtepi32_ps(i32_0);
__m128 p1 = _mm_cvtepi32_ps(i32_1);
__m128 p2 = _mm_cvtepi32_ps(i32_2);
__m128 p3 = _mm_cvtepi32_ps(i32_3);
// Apply the scale
__m128 p0_d = _mm_mul_ps( d_0_1, p0 );
__m128 p1_d = _mm_mul_ps( d_0_1, p1 );
__m128 p2_d = _mm_mul_ps( d_2_3, p2 );
__m128 p3_d = _mm_mul_ps( d_2_3, p3 );
// Acummulate
acc_0 = _mm_add_ps(p0_d, acc_0);
acc_1 = _mm_add_ps(p1_d, acc_1);
acc_2 = _mm_add_ps(p2_d, acc_2);
acc_3 = _mm_add_ps(p3_d, acc_3);
}
*s = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3);
#elif defined(__riscv_v_intrinsic)
float sumf = 0.0;
size_t vl = __riscv_vsetvl_e8m1(qk/2);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
for (int i = 0; i < nb; i++) {
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// load elements
vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[i].qs, vl);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[i].qs, vl);
vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[i].qs+16, vl);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// mask and store lower part of x, and then upper part
vuint8mf2_t x_a = __riscv_vand_vx_u8mf2(tx, 0x0F, vl);
vuint8mf2_t x_l = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
vint8mf2_t x_ai = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a);
vint8mf2_t x_li = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l);
// subtract offset
vint8mf2_t v0 = __riscv_vsub_vx_i8mf2(x_ai, 8, vl);
vint8mf2_t v1 = __riscv_vsub_vx_i8mf2(x_li, 8, vl);
vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl);
vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl);
vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl);
vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl);
vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl);
int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
sumf += sumi*ggml_fp16_to_fp32(x[i].d)*ggml_fp16_to_fp32(y[i].d);
}
*s = sumf;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
#else
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// scalar
float sumf = 0.0;
for (int i = 0; i < nb; i++) {
int sumi = 0;
for (int j = 0; j < qk/2; ++j) {
const int v0 = (x[i].qs[j] & 0x0F) - 8;
const int v1 = (x[i].qs[j] >> 4) - 8;
sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]);
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
}
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
sumf += sumi*ggml_fp16_to_fp32(x[i].d)*ggml_fp16_to_fp32(y[i].d);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
*s = sumf;
#endif
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
const int qk = QK8_1;
const int nb = n / qk;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
assert(n % qk == 0);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
const block_q4_1 * restrict x = vx;
const block_q8_1 * restrict y = vy;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// TODO: add WASM SIMD
#if defined(__ARM_NEON)
float32x4_t sumv0 = vdupq_n_f32(0.0f);
float32x4_t sumv1 = vdupq_n_f32(0.0f);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
float summs = 0;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
assert(nb % 2 == 0); // TODO: handle odd nb
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
for (int i = 0; i < nb; i += 2) {
const block_q4_1 * restrict x0 = &x[i + 0];
const block_q4_1 * restrict x1 = &x[i + 1];
const block_q8_1 * restrict y0 = &y[i + 0];
const block_q8_1 * restrict y1 = &y[i + 1];
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
summs += ggml_fp16_to_fp32(x0->m) * y0->s + ggml_fp16_to_fp32(x1->m) * y1->s;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
const uint8x16_t m4b = vdupq_n_u8(0x0F);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
const uint8x16_t v0_0 = vld1q_u8(x0->qs);
const uint8x16_t v0_1 = vld1q_u8(x1->qs);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// 4-bit -> 8-bit
const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b));
const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b));
const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// load y
const int8x16_t v1_0l = vld1q_s8(y0->qs);
const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
const int8x16_t v1_1l = vld1q_s8(y1->qs);
const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
#if defined(__ARM_FEATURE_DOTPROD)
// dot product into int32x4_t
const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0l), v0_0h, v1_0h);
const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1l), v0_1h, v1_1h);
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), ggml_fp16_to_fp32(x0->d)*y0->d);
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), ggml_fp16_to_fp32(x1->d)*y1->d);
#else
const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0l), vget_low_s8 (v1_0l));
const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0l), vget_high_s8(v1_0l));
const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0h), vget_low_s8 (v1_0h));
const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0h), vget_high_s8(v1_0h));
const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1l), vget_low_s8 (v1_1l));
const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1l), vget_high_s8(v1_1l));
const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1h), vget_low_s8 (v1_1h));
const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1h), vget_high_s8(v1_1h));
const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h));
const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h));
const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h));
const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h));
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), ggml_fp16_to_fp32(x0->d)*y0->d);
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), ggml_fp16_to_fp32(x1->d)*y1->d);
#endif
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
*s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs;
#elif defined(__AVX2__) || defined(__AVX__)
// Initialize accumulator with zeros
__m256 acc = _mm256_setzero_ps();
float summs = 0;
// Main loop
for (int i = 0; i < nb; ++i) {
const float d0 = ggml_fp16_to_fp32(x[i].d);
const float d1 = y[i].d;
summs += ggml_fp16_to_fp32(x[i].m) * y[i].s;
const __m256 d0v = _mm256_set1_ps( d0 );
const __m256 d1v = _mm256_set1_ps( d1 );
// Compute combined scales
const __m256 d0d1 = _mm256_mul_ps( d0v, d1v );
// Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes
const __m256i bx = bytes_from_nibbles_32(x[i].qs);
const __m256i by = _mm256_loadu_si256( (const __m256i *)y[i].qs );
const __m256 xy = mul_sum_us8_pairs_float(bx, by);
// Accumulate d0*d1*x*y
#if defined(__AVX2__)
acc = _mm256_fmadd_ps( d0d1, xy, acc );
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
#else
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
acc = _mm256_add_ps( _mm256_mul_ps( d0d1, xy ), acc );
#endif
}
*s = hsum_float_8(acc) + summs;
#elif defined(__riscv_v_intrinsic)
float sumf = 0.0;
size_t vl = __riscv_vsetvl_e8m1(qk/2);
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
for (int i = 0; i < nb; i++) {
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// load elements
vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[i].qs, vl);
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[i].qs, vl);
vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[i].qs+16, vl);
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// mask and store lower part of x, and then upper part
vuint8mf2_t x_a = __riscv_vand_vx_u8mf2(tx, 0x0F, vl);
vuint8mf2_t x_l = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl);
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
vint8mf2_t v0 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a);
vint8mf2_t v1 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l);
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl);
vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl);
vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl);
vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl);
vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl);
int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
sumf += (ggml_fp16_to_fp32(x[i].d)*y[i].d)*sumi + ggml_fp16_to_fp32(x[i].m)*y[i].s;
}
*s = sumf;
#else
// scalar
float sumf = 0.0;
for (int i = 0; i < nb; i++) {
int sumi = 0;
for (int j = 0; j < qk/2; ++j) {
const int v0 = (x[i].qs[j] & 0x0F);
const int v1 = (x[i].qs[j] >> 4);
sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]);
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
sumf += (ggml_fp16_to_fp32(x[i].d)*y[i].d)*sumi + ggml_fp16_to_fp32(x[i].m)*y[i].s;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
}
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
*s = sumf;
#endif
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
const int qk = QK8_0;
const int nb = n / qk;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
assert(n % qk == 0);
assert(qk == QK5_0);
const block_q5_0 * restrict x = vx;
const block_q8_0 * restrict y = vy;
#if defined(__ARM_NEON)
float32x4_t sumv0 = vdupq_n_f32(0.0f);
float32x4_t sumv1 = vdupq_n_f32(0.0f);
uint32_t qh0;
uint32_t qh1;
uint64_t tmp0[4];
uint64_t tmp1[4];
assert(nb % 2 == 0); // TODO: handle odd nb
for (int i = 0; i < nb; i += 2) {
const block_q5_0 * restrict x0 = &x[i];
const block_q5_0 * restrict x1 = &x[i + 1];
const block_q8_0 * restrict y0 = &y[i];
const block_q8_0 * restrict y1 = &y[i + 1];
const uint8x16_t m4b = vdupq_n_u8(0x0F);
// extract the 5th bit via lookup table ((!b) << 4)
memcpy(&qh0, x0->qh, sizeof(qh0));
memcpy(&qh1, x1->qh, sizeof(qh1));
tmp0[0] = table_b2b_1[(qh0 >> 0) & 0xFF];
tmp0[1] = table_b2b_1[(qh0 >> 8) & 0xFF];
tmp0[2] = table_b2b_1[(qh0 >> 16) & 0xFF];
tmp0[3] = table_b2b_1[(qh0 >> 24) ];
tmp1[0] = table_b2b_1[(qh1 >> 0) & 0xFF];
tmp1[1] = table_b2b_1[(qh1 >> 8) & 0xFF];
tmp1[2] = table_b2b_1[(qh1 >> 16) & 0xFF];
tmp1[3] = table_b2b_1[(qh1 >> 24) ];
const int8x16_t qhl0 = vld1q_s8((const int8_t *)(tmp0 + 0));
const int8x16_t qhh0 = vld1q_s8((const int8_t *)(tmp0 + 2));
const int8x16_t qhl1 = vld1q_s8((const int8_t *)(tmp1 + 0));
const int8x16_t qhh1 = vld1q_s8((const int8_t *)(tmp1 + 2));
const uint8x16_t v0_0 = vld1q_u8(x0->qs);
const uint8x16_t v0_1 = vld1q_u8(x1->qs);
// 4-bit -> 8-bit
int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b));
int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b));
int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
// add high bit and sub 16 (equivalent to sub 0x10 when bit is zero)
const int8x16_t v0_0lf = vsubq_s8(v0_0l, qhl0);
const int8x16_t v0_0hf = vsubq_s8(v0_0h, qhh0);
const int8x16_t v0_1lf = vsubq_s8(v0_1l, qhl1);
const int8x16_t v0_1hf = vsubq_s8(v0_1h, qhh1);
// load y
const int8x16_t v1_0l = vld1q_s8(y0->qs);
const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
const int8x16_t v1_1l = vld1q_s8(y1->qs);
const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
#if defined(__ARM_FEATURE_DOTPROD)
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), ggml_fp16_to_fp32(x0->d)*ggml_fp16_to_fp32(y0->d));
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), ggml_fp16_to_fp32(x1->d)*ggml_fp16_to_fp32(y1->d));
#else
const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lf), vget_low_s8 (v1_0l));
const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lf), vget_high_s8(v1_0l));
const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hf), vget_low_s8 (v1_0h));
const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hf), vget_high_s8(v1_0h));
const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lf), vget_low_s8 (v1_1l));
const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lf), vget_high_s8(v1_1l));
const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hf), vget_low_s8 (v1_1h));
const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hf), vget_high_s8(v1_1h));
const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h));
const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h));
const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h));
const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h));
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), ggml_fp16_to_fp32(x0->d)*ggml_fp16_to_fp32(y0->d));
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), ggml_fp16_to_fp32(x1->d)*ggml_fp16_to_fp32(y1->d));
#endif
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
*s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
#elif defined(__wasm_simd128__)
v128_t sumv = wasm_f32x4_splat(0.0f);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
uint32_t qh;
uint64_t tmp[4];
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// TODO: check if unrolling this is better
for (int i = 0; i < nb; ++i) {
const block_q5_0 * restrict x0 = &x[i];
const block_q8_0 * restrict y0 = &y[i];
const v128_t m4b = wasm_i8x16_splat(0x0F);
// extract the 5th bit
memcpy(&qh, x0->qh, sizeof(qh));
tmp[0] = table_b2b_1[(qh >> 0) & 0xFF];
tmp[1] = table_b2b_1[(qh >> 8) & 0xFF];
tmp[2] = table_b2b_1[(qh >> 16) & 0xFF];
tmp[3] = table_b2b_1[(qh >> 24) ];
const v128_t qhl = wasm_v128_load(tmp + 0);
const v128_t qhh = wasm_v128_load(tmp + 2);
const v128_t v0 = wasm_v128_load(x0->qs);
// 4-bit -> 8-bit
const v128_t v0l = wasm_v128_and (v0, m4b);
const v128_t v0h = wasm_u8x16_shr(v0, 4);
// add high bit and sub 16 (equivalent to sub 0x10 when bit is zero)
const v128_t v0lf = wasm_i8x16_sub(v0l, qhl);
const v128_t v0hf = wasm_i8x16_sub(v0h, qhh);
// load y
const v128_t v1l = wasm_v128_load(y0->qs);
const v128_t v1h = wasm_v128_load(y0->qs + 16);
// int8x16 -> int16x8
const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf);
const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf);
const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf);
const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf);
const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l);
const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l);
const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h);
const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h);
// dot product
sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(
wasm_i32x4_add(
wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll),
wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
wasm_f32x4_splat(ggml_fp16_to_fp32(x0->d) * ggml_fp16_to_fp32(y0->d))));
}
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
*s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3);
#elif defined(__AVX2__)
// Initialize accumulator with zeros
__m256 acc = _mm256_setzero_ps();
// Main loop
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
for (int i = 0; i < nb; i++) {
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
/* Compute combined scale for the block */
const __m256 d = _mm256_set1_ps(ggml_fp16_to_fp32(x[i].d) * ggml_fp16_to_fp32(y[i].d));
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
__m256i bx = bytes_from_nibbles_32(x[i].qs);
__m256i bxhi = bytes_from_bits_32(x[i].qh);
bxhi = _mm256_andnot_si256(bxhi, _mm256_set1_epi8((char)0xF0));
bx = _mm256_or_si256(bx, bxhi);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
__m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
const __m256 q = mul_sum_i8_pairs_float(bx, by);
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
/* Multiply q with scale and accumulate */
acc = _mm256_fmadd_ps(d, q, acc);
}
*s = hsum_float_8(acc);
#elif defined(__AVX__)
// Initialize accumulator with zeros
__m256 acc = _mm256_setzero_ps();
__m128i mask = _mm_set1_epi8((char)0xF0);
// Main loop
for (int i = 0; i < nb; i++) {
/* Compute combined scale for the block */
const __m256 d = _mm256_set1_ps(ggml_fp16_to_fp32(x[i].d) * ggml_fp16_to_fp32(y[i].d));
__m256i bx = bytes_from_nibbles_32(x[i].qs);
const __m256i bxhi = bytes_from_bits_32(x[i].qh);
__m128i bxhil = _mm256_castsi256_si128(bxhi);
__m128i bxhih = _mm256_extractf128_si256(bxhi, 1);
bxhil = _mm_andnot_si128(bxhil, mask);
bxhih = _mm_andnot_si128(bxhih, mask);
__m128i bxl = _mm256_castsi256_si128(bx);
__m128i bxh = _mm256_extractf128_si256(bx, 1);
bxl = _mm_or_si128(bxl, bxhil);
bxh = _mm_or_si128(bxh, bxhih);
bx = MM256_SET_M128I(bxh, bxl);
const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
const __m256 q = mul_sum_i8_pairs_float(bx, by);
/* Multiply q with scale and accumulate */
acc = _mm256_add_ps(_mm256_mul_ps(d, q), acc);
}
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
*s = hsum_float_8(acc);
#elif defined(__riscv_v_intrinsic)
float sumf = 0.0;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
uint32_t qh;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
size_t vl = __riscv_vsetvl_e8m1(qk/2);
// These tempory registers are for masking and shift operations
vuint32m2_t vt_1 = __riscv_vid_v_u32m2(vl);
vuint32m2_t vt_2 = __riscv_vsll_vv_u32m2(__riscv_vmv_v_x_u32m2(1, vl), vt_1, vl);
vuint32m2_t vt_3 = __riscv_vsll_vx_u32m2(vt_2, 16, vl);
vuint32m2_t vt_4 = __riscv_vadd_vx_u32m2(vt_1, 12, vl);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
for (int i = 0; i < nb; i++) {
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
memcpy(&qh, x[i].qh, sizeof(uint32_t));
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
vuint32m2_t xha_0 = __riscv_vand_vx_u32m2(vt_2, qh, vl);
vuint32m2_t xhr_0 = __riscv_vsrl_vv_u32m2(xha_0, vt_1, vl);
vuint32m2_t xhl_0 = __riscv_vsll_vx_u32m2(xhr_0, 4, vl);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// ((qh & (1u << (j + 16))) >> (j + 12));
vuint32m2_t xha_1 = __riscv_vand_vx_u32m2(vt_3, qh, vl);
vuint32m2_t xhl_1 = __riscv_vsrl_vv_u32m2(xha_1, vt_4, vl);
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// narrowing
vuint16m1_t xhc_0 = __riscv_vncvt_x_x_w_u16m1(xhl_0, vl);
vuint8mf2_t xh_0 = __riscv_vncvt_x_x_w_u8mf2(xhc_0, vl);
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
vuint16m1_t xhc_1 = __riscv_vncvt_x_x_w_u16m1(xhl_1, vl);
vuint8mf2_t xh_1 = __riscv_vncvt_x_x_w_u8mf2(xhc_1, vl);
// load
vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[i].qs, vl);
vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[i].qs, vl);
vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[i].qs+16, vl);
vuint8mf2_t x_at = __riscv_vand_vx_u8mf2(tx, 0x0F, vl);
vuint8mf2_t x_lt = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl);
vuint8mf2_t x_a = __riscv_vor_vv_u8mf2(x_at, xh_0, vl);
vuint8mf2_t x_l = __riscv_vor_vv_u8mf2(x_lt, xh_1, vl);
vint8mf2_t x_ai = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a);
vint8mf2_t x_li = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l);
vint8mf2_t v0 = __riscv_vsub_vx_i8mf2(x_ai, 16, vl);
vint8mf2_t v1 = __riscv_vsub_vx_i8mf2(x_li, 16, vl);
vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl);
vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl);
vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl);
vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl);
vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl);
int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
sumf += (ggml_fp16_to_fp32(x[i].d)*ggml_fp16_to_fp32(y[i].d)) * sumi;
}
*s = sumf;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
#else
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// scalar
float sumf = 0.0;
for (int i = 0; i < nb; i++) {
uint32_t qh;
memcpy(&qh, x[i].qh, sizeof(qh));
int sumi = 0;
for (int j = 0; j < qk/2; ++j) {
const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
const int32_t x0 = ((x[i].qs[j] & 0x0F) | xh_0) - 16;
const int32_t x1 = ((x[i].qs[j] >> 4) | xh_1) - 16;
sumi += (x0 * y[i].qs[j]) + (x1 * y[i].qs[j + qk/2]);
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
}
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
sumf += (ggml_fp16_to_fp32(x[i].d)*ggml_fp16_to_fp32(y[i].d)) * sumi;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
*s = sumf;
#endif
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
const int qk = QK8_1;
const int nb = n / qk;
k-quants : fix quantization ranges (#3646)
2023-10-17 06:19:28 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
assert(n % qk == 0);
assert(qk == QK5_1);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
const block_q5_1 * restrict x = vx;
const block_q8_1 * restrict y = vy;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
#if defined(__ARM_NEON)
float32x4_t sumv0 = vdupq_n_f32(0.0f);
float32x4_t sumv1 = vdupq_n_f32(0.0f);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
float summs0 = 0.0f;
float summs1 = 0.0f;
uint32_t qh0;
uint32_t qh1;
uint64_t tmp0[4];
uint64_t tmp1[4];
assert(nb % 2 == 0); // TODO: handle odd nb
for (int i = 0; i < nb; i += 2) {
const block_q5_1 * restrict x0 = &x[i];
const block_q5_1 * restrict x1 = &x[i + 1];
const block_q8_1 * restrict y0 = &y[i];
const block_q8_1 * restrict y1 = &y[i + 1];
const uint8x16_t m4b = vdupq_n_u8(0x0F);
summs0 += ggml_fp16_to_fp32(x0->m) * y0->s;
summs1 += ggml_fp16_to_fp32(x1->m) * y1->s;
// extract the 5th bit via lookup table ((b) << 4)
memcpy(&qh0, x0->qh, sizeof(qh0));
memcpy(&qh1, x1->qh, sizeof(qh1));
tmp0[0] = table_b2b_0[(qh0 >> 0) & 0xFF];
tmp0[1] = table_b2b_0[(qh0 >> 8) & 0xFF];
tmp0[2] = table_b2b_0[(qh0 >> 16) & 0xFF];
tmp0[3] = table_b2b_0[(qh0 >> 24) ];
tmp1[0] = table_b2b_0[(qh1 >> 0) & 0xFF];
tmp1[1] = table_b2b_0[(qh1 >> 8) & 0xFF];
tmp1[2] = table_b2b_0[(qh1 >> 16) & 0xFF];
tmp1[3] = table_b2b_0[(qh1 >> 24) ];
const int8x16_t qhl0 = vld1q_s8((const int8_t *)(tmp0 + 0));
const int8x16_t qhh0 = vld1q_s8((const int8_t *)(tmp0 + 2));
const int8x16_t qhl1 = vld1q_s8((const int8_t *)(tmp1 + 0));
const int8x16_t qhh1 = vld1q_s8((const int8_t *)(tmp1 + 2));
const uint8x16_t v0_0 = vld1q_u8(x0->qs);
const uint8x16_t v0_1 = vld1q_u8(x1->qs);
// 4-bit -> 8-bit
const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b));
const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b));
const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
// add high bit
const int8x16_t v0_0lf = vorrq_s8(v0_0l, qhl0);
const int8x16_t v0_0hf = vorrq_s8(v0_0h, qhh0);
const int8x16_t v0_1lf = vorrq_s8(v0_1l, qhl1);
const int8x16_t v0_1hf = vorrq_s8(v0_1h, qhh1);
// load y
const int8x16_t v1_0l = vld1q_s8(y0->qs);
const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
const int8x16_t v1_1l = vld1q_s8(y1->qs);
const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
#if defined(__ARM_FEATURE_DOTPROD)
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), ggml_fp16_to_fp32(x0->d)*y0->d);
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), ggml_fp16_to_fp32(x1->d)*y1->d);
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
#else
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lf), vget_low_s8 (v1_0l));
const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lf), vget_high_s8(v1_0l));
const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hf), vget_low_s8 (v1_0h));
const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hf), vget_high_s8(v1_0h));
const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lf), vget_low_s8 (v1_1l));
const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lf), vget_high_s8(v1_1l));
const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hf), vget_low_s8 (v1_1h));
const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hf), vget_high_s8(v1_1h));
const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h));
const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h));
const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h));
const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h));
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), ggml_fp16_to_fp32(x0->d)*y0->d);
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), ggml_fp16_to_fp32(x1->d)*y1->d);
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
#endif
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
}
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
*s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs0 + summs1;
#elif defined(__wasm_simd128__)
v128_t sumv = wasm_f32x4_splat(0.0f);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
float summs = 0.0f;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
uint32_t qh;
uint64_t tmp[4];
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// TODO: check if unrolling this is better
for (int i = 0; i < nb; ++i) {
const block_q5_1 * restrict x0 = &x[i];
const block_q8_1 * restrict y0 = &y[i];
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
summs += ggml_fp16_to_fp32(x0->m) * y0->s;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
const v128_t m4b = wasm_i8x16_splat(0x0F);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// extract the 5th bit
memcpy(&qh, x0->qh, sizeof(qh));
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
tmp[0] = table_b2b_0[(qh >> 0) & 0xFF];
tmp[1] = table_b2b_0[(qh >> 8) & 0xFF];
tmp[2] = table_b2b_0[(qh >> 16) & 0xFF];
tmp[3] = table_b2b_0[(qh >> 24) ];
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
const v128_t qhl = wasm_v128_load(tmp + 0);
const v128_t qhh = wasm_v128_load(tmp + 2);
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
const v128_t v0 = wasm_v128_load(x0->qs);
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// 4-bit -> 8-bit
const v128_t v0l = wasm_v128_and (v0, m4b);
const v128_t v0h = wasm_u8x16_shr(v0, 4);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// add high bit
const v128_t v0lf = wasm_v128_or(v0l, qhl);
const v128_t v0hf = wasm_v128_or(v0h, qhh);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// load y
const v128_t v1l = wasm_v128_load(y0->qs);
const v128_t v1h = wasm_v128_load(y0->qs + 16);
// int8x16 -> int16x8
const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf);
const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf);
const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf);
const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf);
const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l);
const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l);
const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h);
const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h);
// dot product
sumv = wasm_f32x4_add(sumv,
wasm_f32x4_mul(wasm_f32x4_convert_i32x4(wasm_i32x4_add(
wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll),
wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
wasm_f32x4_splat(ggml_fp16_to_fp32(x0->d) * y0->d)));
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
*s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3) + summs;
#elif defined(__AVX2__)
// Initialize accumulator with zeros
__m256 acc = _mm256_setzero_ps();
float summs = 0.0f;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// Main loop
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
for (int i = 0; i < nb; i++) {
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
const __m256 dx = _mm256_set1_ps(ggml_fp16_to_fp32(x[i].d));
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
summs += ggml_fp16_to_fp32(x[i].m) * y[i].s;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
__m256i bx = bytes_from_nibbles_32(x[i].qs);
__m256i bxhi = bytes_from_bits_32(x[i].qh);
bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10));
bx = _mm256_or_si256(bx, bxhi);
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
const __m256 dy = _mm256_set1_ps(y[i].d);
const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
const __m256 q = mul_sum_us8_pairs_float(bx, by);
acc = _mm256_fmadd_ps(q, _mm256_mul_ps(dx, dy), acc);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
*s = hsum_float_8(acc) + summs;
#elif defined(__AVX__)
// Initialize accumulator with zeros
__m256 acc = _mm256_setzero_ps();
__m128i mask = _mm_set1_epi8(0x10);
float summs = 0.0f;
// Main loop
for (int i = 0; i < nb; i++) {
const __m256 dx = _mm256_set1_ps(ggml_fp16_to_fp32(x[i].d));
summs += ggml_fp16_to_fp32(x[i].m) * y[i].s;
__m256i bx = bytes_from_nibbles_32(x[i].qs);
const __m256i bxhi = bytes_from_bits_32(x[i].qh);
__m128i bxhil = _mm256_castsi256_si128(bxhi);
__m128i bxhih = _mm256_extractf128_si256(bxhi, 1);
bxhil = _mm_and_si128(bxhil, mask);
bxhih = _mm_and_si128(bxhih, mask);
__m128i bxl = _mm256_castsi256_si128(bx);
__m128i bxh = _mm256_extractf128_si256(bx, 1);
bxl = _mm_or_si128(bxl, bxhil);
bxh = _mm_or_si128(bxh, bxhih);
bx = MM256_SET_M128I(bxh, bxl);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
const __m256 dy = _mm256_set1_ps(y[i].d);
const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
k-quants : fix quantization ranges (#3646)
2023-10-17 06:19:28 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
const __m256 q = mul_sum_us8_pairs_float(bx, by);
acc = _mm256_add_ps(_mm256_mul_ps(q, _mm256_mul_ps(dx, dy)), acc);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
*s = hsum_float_8(acc) + summs;
#elif defined(__riscv_v_intrinsic)
float sumf = 0.0;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
uint32_t qh;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
size_t vl = __riscv_vsetvl_e8m1(qk/2);
// temporary registers for shift operations
vuint32m2_t vt_1 = __riscv_vid_v_u32m2(vl);
vuint32m2_t vt_2 = __riscv_vadd_vx_u32m2(vt_1, 12, vl);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
for (int i = 0; i < nb; i++) {
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
memcpy(&qh, x[i].qh, sizeof(uint32_t));
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// load qh
vuint32m2_t vqh = __riscv_vmv_v_x_u32m2(qh, vl);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// ((qh >> (j + 0)) << 4) & 0x10;
vuint32m2_t xhr_0 = __riscv_vsrl_vv_u32m2(vqh, vt_1, vl);
vuint32m2_t xhl_0 = __riscv_vsll_vx_u32m2(xhr_0, 4, vl);
vuint32m2_t xha_0 = __riscv_vand_vx_u32m2(xhl_0, 0x10, vl);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// ((qh >> (j + 12)) ) & 0x10;
vuint32m2_t xhr_1 = __riscv_vsrl_vv_u32m2(vqh, vt_2, vl);
vuint32m2_t xha_1 = __riscv_vand_vx_u32m2(xhr_1, 0x10, vl);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// narrowing
vuint16m1_t xhc_0 = __riscv_vncvt_x_x_w_u16m1(xha_0, vl);
vuint8mf2_t xh_0 = __riscv_vncvt_x_x_w_u8mf2(xhc_0, vl);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
vuint16m1_t xhc_1 = __riscv_vncvt_x_x_w_u16m1(xha_1, vl);
vuint8mf2_t xh_1 = __riscv_vncvt_x_x_w_u8mf2(xhc_1, vl);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// load
vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[i].qs, vl);
Guard against all weights in a super-block being zero (#3010) * Guard against all weights in a super-block being zero * Also guard against extremely small weights Closes #2982 --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-09-05 07:55:33 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[i].qs, vl);
vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[i].qs+16, vl);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
vuint8mf2_t x_at = __riscv_vand_vx_u8mf2(tx, 0x0F, vl);
vuint8mf2_t x_lt = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
vuint8mf2_t x_a = __riscv_vor_vv_u8mf2(x_at, xh_0, vl);
vuint8mf2_t x_l = __riscv_vor_vv_u8mf2(x_lt, xh_1, vl);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
vint8mf2_t v0 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a);
vint8mf2_t v1 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l);
vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl);
vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl);
vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl);
vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl);
vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
sumf += (ggml_fp16_to_fp32(x[i].d)*y[i].d)*sumi + ggml_fp16_to_fp32(x[i].m)*y[i].s;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
*s = sumf;
#else
// scalar
float sumf = 0.0;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
for (int i = 0; i < nb; i++) {
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
uint32_t qh;
memcpy(&qh, x[i].qh, sizeof(qh));
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
int sumi = 0;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
for (int j = 0; j < qk/2; ++j) {
const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10;
const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
const int32_t x0 = (x[i].qs[j] & 0xF) | xh_0;
const int32_t x1 = (x[i].qs[j] >> 4) | xh_1;
sumi += (x0 * y[i].qs[j]) + (x1 * y[i].qs[j + qk/2]);
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
}
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
sumf += (ggml_fp16_to_fp32(x[i].d)*y[i].d)*sumi + ggml_fp16_to_fp32(x[i].m)*y[i].s;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
*s = sumf;
#endif
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
const int qk = QK8_0;
const int nb = n / qk;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
assert(n % qk == 0);
const block_q8_0 * restrict x = vx;
const block_q8_0 * restrict y = vy;
#if defined(__ARM_NEON)
float32x4_t sumv0 = vdupq_n_f32(0.0f);
float32x4_t sumv1 = vdupq_n_f32(0.0f);
assert(nb % 2 == 0); // TODO: handle odd nb
for (int i = 0; i < nb; i += 2) {
const block_q8_0 * restrict x0 = &x[i + 0];
const block_q8_0 * restrict x1 = &x[i + 1];
const block_q8_0 * restrict y0 = &y[i + 0];
const block_q8_0 * restrict y1 = &y[i + 1];
const int8x16_t x0_0 = vld1q_s8(x0->qs);
const int8x16_t x0_1 = vld1q_s8(x0->qs + 16);
const int8x16_t x1_0 = vld1q_s8(x1->qs);
const int8x16_t x1_1 = vld1q_s8(x1->qs + 16);
// load y
const int8x16_t y0_0 = vld1q_s8(y0->qs);
const int8x16_t y0_1 = vld1q_s8(y0->qs + 16);
const int8x16_t y1_0 = vld1q_s8(y1->qs);
const int8x16_t y1_1 = vld1q_s8(y1->qs + 16);
#if defined(__ARM_FEATURE_DOTPROD)
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
vdotq_s32(vdupq_n_s32(0), x0_0, y0_0),
vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), ggml_fp16_to_fp32(x0->d)*ggml_fp16_to_fp32(y0->d));
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
vdotq_s32(vdupq_n_s32(0), x1_0, y1_0),
vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), ggml_fp16_to_fp32(x1->d)*ggml_fp16_to_fp32(y1->d));
#else
const int16x8_t p0_0 = vmull_s8(vget_low_s8 (x0_0), vget_low_s8 (y0_0));
const int16x8_t p0_1 = vmull_s8(vget_high_s8(x0_0), vget_high_s8(y0_0));
const int16x8_t p0_2 = vmull_s8(vget_low_s8 (x0_1), vget_low_s8 (y0_1));
const int16x8_t p0_3 = vmull_s8(vget_high_s8(x0_1), vget_high_s8(y0_1));
const int16x8_t p1_0 = vmull_s8(vget_low_s8 (x1_0), vget_low_s8 (y1_0));
const int16x8_t p1_1 = vmull_s8(vget_high_s8(x1_0), vget_high_s8(y1_0));
const int16x8_t p1_2 = vmull_s8(vget_low_s8 (x1_1), vget_low_s8 (y1_1));
const int16x8_t p1_3 = vmull_s8(vget_high_s8(x1_1), vget_high_s8(y1_1));
const int32x4_t p0 = vaddq_s32(vpaddlq_s16(p0_0), vpaddlq_s16(p0_1));
const int32x4_t p1 = vaddq_s32(vpaddlq_s16(p0_2), vpaddlq_s16(p0_3));
const int32x4_t p2 = vaddq_s32(vpaddlq_s16(p1_0), vpaddlq_s16(p1_1));
const int32x4_t p3 = vaddq_s32(vpaddlq_s16(p1_2), vpaddlq_s16(p1_3));
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(p0, p1)), ggml_fp16_to_fp32(x0->d)*ggml_fp16_to_fp32(y0->d));
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(p2, p3)), ggml_fp16_to_fp32(x1->d)*ggml_fp16_to_fp32(y1->d));
#endif
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
*s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
#elif defined(__AVX2__) || defined(__AVX__)
// Initialize accumulator with zeros
__m256 acc = _mm256_setzero_ps();
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// Main loop
for (int i = 0; i < nb; ++i) {
// Compute combined scale for the block
const __m256 d = _mm256_set1_ps(ggml_fp16_to_fp32(x[i].d) * ggml_fp16_to_fp32(y[i].d));
__m256i bx = _mm256_loadu_si256((const __m256i *)x[i].qs);
__m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
const __m256 q = mul_sum_i8_pairs_float(bx, by);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// Multiply q with scale and accumulate
#if defined(__AVX2__)
acc = _mm256_fmadd_ps( d, q, acc );
#else
acc = _mm256_add_ps( _mm256_mul_ps( d, q ), acc );
#endif
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
*s = hsum_float_8(acc);
#elif defined(__riscv_v_intrinsic)
float sumf = 0.0;
size_t vl = __riscv_vsetvl_e8m1(qk);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
for (int i = 0; i < nb; i++) {
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
// load elements
vint8m1_t bx = __riscv_vle8_v_i8m1(x[i].qs, vl);
vint8m1_t by = __riscv_vle8_v_i8m1(y[i].qs, vl);
vint16m2_t vw_mul = __riscv_vwmul_vv_i16m2(bx, by, vl);
vint32m1_t v_zero = __riscv_vmv_v_x_i32m1(0, vl);
vint32m1_t v_sum = __riscv_vwredsum_vs_i16m2_i32m1(vw_mul, v_zero, vl);
int sumi = __riscv_vmv_x_s_i32m1_i32(v_sum);
sumf += sumi*(ggml_fp16_to_fp32(x[i].d)*ggml_fp16_to_fp32(y[i].d));
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
*s = sumf;
#else
// scalar
float sumf = 0.0;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
for (int i = 0; i < nb; i++) {
int sumi = 0;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
for (int j = 0; j < qk; j++) {
sumi += x[i].qs[j]*y[i].qs[j];
}
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
sumf += sumi*(ggml_fp16_to_fp32(x[i].d)*ggml_fp16_to_fp32(y[i].d));
}
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
*s = sumf;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
#endif
ggml : factor all quantization code in ggml-quants ggml-ci
2023-10-28 14:05:07 +00:00
}
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
#if QK_K == 256
k-quants : allow to optionally disable at compile time (#1734) * k-quants : put behind optional compile flag LLAMA_K_QUANTS * build : enable k-quants by default
2023-06-07 07:59:52 +00:00
void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
k-quants : allow to optionally disable at compile time (#1734) * k-quants : put behind optional compile flag LLAMA_K_QUANTS * build : enable k-quants by default
2023-06-07 07:59:52 +00:00
const block_q2_K * restrict x = vx;
const block_q8_K * restrict y = vy;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
const int nb = n / QK_K;
#ifdef __ARM_NEON
const uint8x16_t m3 = vdupq_n_u8(0x3);
const uint8x16_t m4 = vdupq_n_u8(0xF);
k-quants : fix build on armv7 (android only) (#2920) * k-quants : fix build on armv7 * ggml : cleanup unused arm32 specific impl * k-quants : avoid some unused vzero / mzero define * ggml-alloc : use 4g for MEASURE_MAX_SIZE in 32-bit arm
2023-09-02 12:23:45 +00:00
#if defined(__ARM_FEATURE_DOTPROD)
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
const int32x4_t vzero = vdupq_n_s32(0);
k-quants : fix build on armv7 (android only) (#2920) * k-quants : fix build on armv7 * ggml : cleanup unused arm32 specific impl * k-quants : avoid some unused vzero / mzero define * ggml-alloc : use 4g for MEASURE_MAX_SIZE in 32-bit arm
2023-09-02 12:23:45 +00:00
#endif
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
int8x16x2_t q2bytes;
uint8_t aux[16];
float sum = 0;
for (int i = 0; i < nb; ++i) {
const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin);
const uint8_t * restrict q2 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
const uint8_t * restrict sc = x[i].scales;
const uint8x16_t mins_and_scales = vld1q_u8(sc);
const uint8x16_t scales = vandq_u8(mins_and_scales, m4);
vst1q_u8(aux, scales);
const uint8x16_t mins = vshrq_n_u8(mins_and_scales, 4);
const int16x8x2_t q8sums = vld1q_s16_x2(y[i].bsums);
const int16x8x2_t mins16 = {vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(mins))), vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(mins)))};
const int32x4_t s0 = vaddq_s32(vmull_s16(vget_low_s16 (mins16.val[0]), vget_low_s16 (q8sums.val[0])),
vmull_s16(vget_high_s16(mins16.val[0]), vget_high_s16(q8sums.val[0])));
const int32x4_t s1 = vaddq_s32(vmull_s16(vget_low_s16 (mins16.val[1]), vget_low_s16 (q8sums.val[1])),
vmull_s16(vget_high_s16(mins16.val[1]), vget_high_s16(q8sums.val[1])));
sum += dmin * vaddvq_s32(vaddq_s32(s0, s1));
int isum = 0;
int is = 0;
// We use this macro instead of a function call because for some reason
// the code runs 2-3% slower, even if the function is declared inline
#if defined(__ARM_FEATURE_DOTPROD)
#define MULTIPLY_ACCUM_WITH_SCALE(index)\
isum += vaddvq_s32(vdotq_s32(vzero, q2bytes.val[0], q8bytes.val[0])) * aux[is+(index)];\
isum += vaddvq_s32(vdotq_s32(vzero, q2bytes.val[1], q8bytes.val[1])) * aux[is+1+(index)];
#else
#define MULTIPLY_ACCUM_WITH_SCALE(index)\
{\
const int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q2bytes.val[0]), vget_low_s8 (q8bytes.val[0])),\
vmull_s8(vget_high_s8(q2bytes.val[0]), vget_high_s8(q8bytes.val[0])));\
const int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q2bytes.val[1]), vget_low_s8 (q8bytes.val[1])),\
vmull_s8(vget_high_s8(q2bytes.val[1]), vget_high_s8(q8bytes.val[1])));\
isum += vaddvq_s16(p1) * aux[is+(index)] + vaddvq_s16(p2) * aux[is+1+(index)];\
}
#endif
#define SHIFT_MULTIPLY_ACCUM_WITH_SCALE(shift, index)\
q8bytes = vld1q_s8_x2(q8); q8 += 32;\
q2bytes.val[0] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits.val[0], (shift)), m3));\
q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits.val[1], (shift)), m3));\
MULTIPLY_ACCUM_WITH_SCALE((index));
for (int j = 0; j < QK_K/128; ++j) {
const uint8x16x2_t q2bits = vld1q_u8_x2(q2); q2 += 32;
int8x16x2_t q8bytes = vld1q_s8_x2(q8); q8 += 32;
q2bytes.val[0] = vreinterpretq_s8_u8(vandq_u8(q2bits.val[0], m3));
q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(q2bits.val[1], m3));
MULTIPLY_ACCUM_WITH_SCALE(0);
SHIFT_MULTIPLY_ACCUM_WITH_SCALE(2, 2);
SHIFT_MULTIPLY_ACCUM_WITH_SCALE(4, 4);
SHIFT_MULTIPLY_ACCUM_WITH_SCALE(6, 6);
is += 8;
}
sum += d * isum;
}
*s = sum;
#elif defined __AVX2__
const __m256i m3 = _mm256_set1_epi8(3);
const __m128i m4 = _mm_set1_epi8(0xF);
__m256 acc = _mm256_setzero_ps();
for (int i = 0; i < nb; ++i) {
const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin);
const uint8_t * restrict q2 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
const __m128i mins_and_scales = _mm_loadu_si128((const __m128i*)x[i].scales);
const __m128i scales8 = _mm_and_si128(mins_and_scales, m4);
const __m128i mins8 = _mm_and_si128(_mm_srli_epi16(mins_and_scales, 4), m4);
const __m256i mins = _mm256_cvtepi8_epi16(mins8);
const __m256i prod = _mm256_madd_epi16(mins, _mm256_loadu_si256((const __m256i*)y[i].bsums));
acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&dmin), _mm256_cvtepi32_ps(prod), acc);
const __m256i all_scales = _mm256_cvtepi8_epi16(scales8);
const __m128i l_scales = _mm256_extracti128_si256(all_scales, 0);
const __m128i h_scales = _mm256_extracti128_si256(all_scales, 1);
ggml : workaround for missing _mm256_setr_m128i in GCC < 8 in k_quants.c (#2405)
2023-07-28 18:17:45 +00:00
const __m256i scales[2] = {MM256_SET_M128I(l_scales, l_scales), MM256_SET_M128I(h_scales, h_scales)};
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
__m256i sumi = _mm256_setzero_si256();
for (int j = 0; j < QK_K/128; ++j) {
const __m256i q2bits = _mm256_loadu_si256((const __m256i*)q2); q2 += 32;
const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
const __m256i q8_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
const __m256i q8_3 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
const __m256i q2_0 = _mm256_and_si256(q2bits, m3);
const __m256i q2_1 = _mm256_and_si256(_mm256_srli_epi16(q2bits, 2), m3);
const __m256i q2_2 = _mm256_and_si256(_mm256_srli_epi16(q2bits, 4), m3);
const __m256i q2_3 = _mm256_and_si256(_mm256_srli_epi16(q2bits, 6), m3);
__m256i p0 = _mm256_maddubs_epi16(q2_0, q8_0);
__m256i p1 = _mm256_maddubs_epi16(q2_1, q8_1);
__m256i p2 = _mm256_maddubs_epi16(q2_2, q8_2);
__m256i p3 = _mm256_maddubs_epi16(q2_3, q8_3);
p0 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(0)), p0);
p1 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(1)), p1);
p2 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(2)), p2);
p3 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(3)), p3);
p0 = _mm256_add_epi32(p0, p1);
p2 = _mm256_add_epi32(p2, p3);
sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p0, p2));
}
acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
}
*s = hsum_float_8(acc);
k-quants : add AVX support to dot functions (#1916) * k_quants : add AVX support * k_quants : apply review comments
2023-06-26 16:46:07 +00:00
#elif defined __AVX__
const __m128i m3 = _mm_set1_epi8(0x3);
const __m128i m4 = _mm_set1_epi8(0xF);
const __m128i m2 = _mm_set1_epi8(0x2);
__m256 acc = _mm256_setzero_ps();
for (int i = 0; i < nb; ++i) {
const float dall = y[i].d * ggml_fp16_to_fp32(x[i].d);
const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin);
const uint8_t * restrict q2 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
// load mins and scales from block_q2_K.scales[QK_K/16]
const __m128i mins_and_scales = _mm_loadu_si128((const __m128i*)x[i].scales);
const __m128i scales16 = _mm_and_si128(mins_and_scales, m4);
const __m128i mins16 = _mm_and_si128(_mm_srli_epi16(mins_and_scales, 4), m4);
const __m128i mins_0 = _mm_cvtepi8_epi16(mins16);
const __m128i mins_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(mins16, mins16));
// summs = y[i].bsums * (x[i].scales >> 4) in 16bits*8*2 to 32bits*4*2
const __m128i summs_0 = _mm_madd_epi16(mins_0, _mm_loadu_si128((const __m128i*)&y[i].bsums[0]));
const __m128i summs_1 = _mm_madd_epi16(mins_1, _mm_loadu_si128((const __m128i*)&y[i].bsums[8]));
// sumf += -dmin * summs in 32bits*8
ggml : workaround for missing _mm256_setr_m128i in GCC < 8 in k_quants.c (#2405)
2023-07-28 18:17:45 +00:00
acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dmin), _mm256_cvtepi32_ps(MM256_SET_M128I(summs_1, summs_0))), acc);
k-quants : add AVX support to dot functions (#1916) * k_quants : add AVX support * k_quants : apply review comments
2023-06-26 16:46:07 +00:00
const __m128i scales_0 = _mm_cvtepi8_epi16(scales16);
const __m128i scales_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(scales16, scales16));
const __m128i scales[2] = { scales_0, scales_1 };
__m128i sumi_0 = _mm_setzero_si128();
__m128i sumi_1 = _mm_setzero_si128();
for (int j = 0; j < QK_K/128; ++j) {
// load Q8 quants int8*16*8 from block_q8_K.qs[QK_K]
const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
// load 2bits*16*8 from block_q2_K.qs[QK_K/4]
__m128i q2bits = _mm_loadu_si128((const __m128i*)q2); q2 += 16;
const __m128i q2_0 = _mm_and_si128(q2bits, m3);
const __m128i q2_2 = _mm_and_si128(_mm_srli_epi16(q2bits, 2), m3);
const __m128i q2_4 = _mm_and_si128(_mm_srli_epi16(q2bits, 4), m3);
const __m128i q2_6 = _mm_and_si128(_mm_srli_epi16(q2bits, 6), m3);
q2bits = _mm_loadu_si128((const __m128i*)q2); q2 += 16;
const __m128i q2_1 = _mm_and_si128(q2bits, m3);
const __m128i q2_3 = _mm_and_si128(_mm_srli_epi16(q2bits, 2), m3);
const __m128i q2_5 = _mm_and_si128(_mm_srli_epi16(q2bits, 4), m3);
const __m128i q2_7 = _mm_and_si128(_mm_srli_epi16(q2bits, 6), m3);
// isuml = q8[l] * ((q2[l] >> shift) & 3) in 8bits*16*8 to 16bits*8*8
__m128i p0 = _mm_maddubs_epi16(q2_0, q8_0);
__m128i p1 = _mm_maddubs_epi16(q2_1, q8_1);
__m128i p2 = _mm_maddubs_epi16(q2_2, q8_2);
__m128i p3 = _mm_maddubs_epi16(q2_3, q8_3);
__m128i p4 = _mm_maddubs_epi16(q2_4, q8_4);
__m128i p5 = _mm_maddubs_epi16(q2_5, q8_5);
__m128i p6 = _mm_maddubs_epi16(q2_6, q8_6);
__m128i p7 = _mm_maddubs_epi16(q2_7, q8_7);
// isum += (x[i].scales[is++] & 0xF) * isuml in 16bits*8*8 to 32bits*4*8
__m128i shuffle = _mm_set1_epi16(0x0100);
p0 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p0);
shuffle = _mm_add_epi16(shuffle, m2);
p1 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p1);
shuffle = _mm_add_epi16(shuffle, m2);
p2 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p2);
shuffle = _mm_add_epi16(shuffle, m2);
p3 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p3);
shuffle = _mm_add_epi16(shuffle, m2);
p4 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p4);
shuffle = _mm_add_epi16(shuffle, m2);
p5 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p5);
shuffle = _mm_add_epi16(shuffle, m2);
p6 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p6);
shuffle = _mm_add_epi16(shuffle, m2);
p7 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p7);
p0 = _mm_add_epi32(p0, p1);
p2 = _mm_add_epi32(p2, p3);
p4 = _mm_add_epi32(p4, p5);
p6 = _mm_add_epi32(p6, p7);
// isum in 32bits*4*2
sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p0, p2));
sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p4, p6));
}
// sumf += dall * isum - dmin * summs in 32bits
ggml : workaround for missing _mm256_setr_m128i in GCC < 8 in k_quants.c (#2405)
2023-07-28 18:17:45 +00:00
__m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
k-quants : add AVX support to dot functions (#1916) * k_quants : add AVX support * k_quants : apply review comments
2023-06-26 16:46:07 +00:00
acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dall), _mm256_cvtepi32_ps(sumi)), acc);
}
*s = hsum_float_8(acc);
ggml : add RISC-V Vector Support for K-Quants and improved the existing intrinsics (#3453) * Added RVV intrinsics support for Q8 quantize row and also improved the existing dot product function for risc-v. The RVV intrinsics is added for the following quantize row functions quantize_row_q8_0 quantize_row_q8_1 The following dot product functions have also been optimized by using LMUL = 1/2 instead of LMUL = 1 ggml_vec_dot_q4_0_q8_0 ggml_vec_dot_q4_1_q8_1 ggml_vec_dot_q5_0_q8_0 ggml_vec_dot_q5_1_q8_1 And vector initialization in Q5 by temporary array is also replaced by the vid intrinsics Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai> * Added RVV intrinsics support for k_quants This adds RISC-V Vector intrinsics support for the following K_quants functions for both QKK = 256 and QKK = 64 ggml_vec_dot_q2_K_q8_K ggml_vec_dot_q3_K_q8_K ggml_vec_dot_q4_K_q8_K ggml_vec_dot_q5_K_q8_K ggml_vec_dot_q6_K_q8_K Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai> --------- Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai>
2023-10-03 18:38:19 +00:00
#elif defined __riscv_v_intrinsic
float sumf = 0;
uint8_t temp_01[32] = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1};
for (int i = 0; i < nb; ++i) {
const uint8_t * q2 = x[i].qs;
const int8_t * q8 = y[i].qs;
const uint8_t * sc = x[i].scales;
const float dall = y[i].d * ggml_fp16_to_fp32(x[i].d);
const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin);
size_t vl = 16;
vuint8m1_t scales = __riscv_vle8_v_u8m1(sc, vl);
vuint8m1_t aux = __riscv_vand_vx_u8m1(scales, 0x0F, vl);
vint16m1_t q8sums = __riscv_vle16_v_i16m1(y[i].bsums, vl);
vuint8mf2_t scales_2 = __riscv_vle8_v_u8mf2(sc, vl);
vuint8mf2_t mins8 = __riscv_vsrl_vx_u8mf2(scales_2, 0x4, vl);
vint16m1_t mins = __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(mins8, vl));
vint32m2_t prod = __riscv_vwmul_vv_i32m2(q8sums, mins, vl);
vint32m1_t vsums = __riscv_vredsum_vs_i32m2_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl);
sumf += dmin * __riscv_vmv_x_s_i32m1_i32(vsums);
vl = 32;
vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
vuint8m1_t v_b = __riscv_vle8_v_u8m1(temp_01, vl);
uint8_t is=0;
int isum=0;
for (int j = 0; j < QK_K/128; ++j) {
// load Q2
vuint8m1_t q2_x = __riscv_vle8_v_u8m1(q2, vl);
vuint8m1_t q2_0 = __riscv_vand_vx_u8m1(q2_x, 0x03, vl);
vuint8m1_t q2_1 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q2_x, 0x2, vl), 0x03 , vl);
vuint8m1_t q2_2 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q2_x, 0x4, vl), 0x03 , vl);
vuint8m1_t q2_3 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q2_x, 0x6, vl), 0x03 , vl);
// duplicate scale elements for product
vuint8m1_t sc0 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 0+is, vl), vl);
vuint8m1_t sc1 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 2+is, vl), vl);
vuint8m1_t sc2 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 4+is, vl), vl);
vuint8m1_t sc3 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 6+is, vl), vl);
vint16m2_t p0 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_0, sc0, vl));
vint16m2_t p1 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_1, sc1, vl));
vint16m2_t p2 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_2, sc2, vl));
vint16m2_t p3 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_3, sc3, vl));
// load Q8
vint8m1_t q8_0 = __riscv_vle8_v_i8m1(q8, vl);
vint8m1_t q8_1 = __riscv_vle8_v_i8m1(q8+32, vl);
vint8m1_t q8_2 = __riscv_vle8_v_i8m1(q8+64, vl);
vint8m1_t q8_3 = __riscv_vle8_v_i8m1(q8+96, vl);
vint32m4_t s0 = __riscv_vwmul_vv_i32m4(p0, __riscv_vwcvt_x_x_v_i16m2(q8_0, vl), vl);
vint32m4_t s1 = __riscv_vwmul_vv_i32m4(p1, __riscv_vwcvt_x_x_v_i16m2(q8_1, vl), vl);
vint32m4_t s2 = __riscv_vwmul_vv_i32m4(p2, __riscv_vwcvt_x_x_v_i16m2(q8_2, vl), vl);
vint32m4_t s3 = __riscv_vwmul_vv_i32m4(p3, __riscv_vwcvt_x_x_v_i16m2(q8_3, vl), vl);
vint32m1_t isum0 = __riscv_vredsum_vs_i32m4_i32m1(__riscv_vadd_vv_i32m4(s0, s1, vl), vzero, vl);
vint32m1_t isum1 = __riscv_vredsum_vs_i32m4_i32m1(__riscv_vadd_vv_i32m4(s2, s3, vl), isum0, vl);
isum += __riscv_vmv_x_s_i32m1_i32(isum1);
q2+=32; q8+=128; is=8;
}
sumf += dall * isum;
}
*s = sumf;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
#else
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * q2 = x[i].qs;
const int8_t * q8 = y[i].qs;
const uint8_t * sc = x[i].scales;
int summs = 0;
for (int j = 0; j < 16; ++j) {
summs += y[i].bsums[j] * (sc[j] >> 4);
}
const float dall = y[i].d * ggml_fp16_to_fp32(x[i].d);
const float dmin = y[i].d * ggml_fp16_to_fp32(x[i].dmin);
int isum = 0;
int is = 0;
int d;
for (int k = 0; k < QK_K/128; ++k) {
int shift = 0;
for (int j = 0; j < 4; ++j) {
d = sc[is++] & 0xF;
int isuml = 0;
for (int l = 0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
isum += d * isuml;
d = sc[is++] & 0xF;
isuml = 0;
for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
isum += d * isuml;
shift += 2;
q8 += 32;
}
q2 += 32;
}
sumf += dall * isum - dmin * summs;
}
*s = sumf;
#endif
}
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
#else
void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
const block_q2_K * restrict x = vx;
const block_q8_K * restrict y = vy;
const int nb = n / QK_K;
#ifdef __ARM_NEON
const uint8x16_t m3 = vdupq_n_u8(0x3);
k-quants : fix build on armv7 (android only) (#2920) * k-quants : fix build on armv7 * ggml : cleanup unused arm32 specific impl * k-quants : avoid some unused vzero / mzero define * ggml-alloc : use 4g for MEASURE_MAX_SIZE in 32-bit arm
2023-09-02 12:23:45 +00:00
#if defined(__ARM_FEATURE_DOTPROD)
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
const int32x4_t vzero = vdupq_n_s32(0);
k-quants : fix build on armv7 (android only) (#2920) * k-quants : fix build on armv7 * ggml : cleanup unused arm32 specific impl * k-quants : avoid some unused vzero / mzero define * ggml-alloc : use 4g for MEASURE_MAX_SIZE in 32-bit arm
2023-09-02 12:23:45 +00:00
#endif
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
int8x16x4_t q2bytes;
uint32_t aux32[2];
const uint8_t * scales = (const uint8_t *)aux32;
float sum = 0;
for (int i = 0; i < nb; ++i) {
const float d = y[i].d * (float)x[i].d;
const float dmin = -y[i].d * (float)x[i].dmin;
const uint8_t * restrict q2 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
const uint32_t * restrict sc = (const uint32_t *)x[i].scales;
aux32[0] = sc[0] & 0x0f0f0f0f;
aux32[1] = (sc[0] >> 4) & 0x0f0f0f0f;
sum += dmin * (scales[4] * y[i].bsums[0] + scales[5] * y[i].bsums[1] + scales[6] * y[i].bsums[2] + scales[7] * y[i].bsums[3]);
int isum1 = 0, isum2 = 0;
const uint8x16_t q2bits = vld1q_u8(q2);
const int8x16x4_t q8bytes = vld1q_s8_x4(q8);
q2bytes.val[0] = vreinterpretq_s8_u8(vandq_u8(q2bits, m3));
q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits, 2), m3));
q2bytes.val[2] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits, 4), m3));
q2bytes.val[3] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits, 6), m3));
#if defined(__ARM_FEATURE_DOTPROD)
isum1 += vaddvq_s32(vdotq_s32(vzero, q2bytes.val[0], q8bytes.val[0])) * scales[0];
isum2 += vaddvq_s32(vdotq_s32(vzero, q2bytes.val[1], q8bytes.val[1])) * scales[1];
isum1 += vaddvq_s32(vdotq_s32(vzero, q2bytes.val[2], q8bytes.val[2])) * scales[2];
isum2 += vaddvq_s32(vdotq_s32(vzero, q2bytes.val[3], q8bytes.val[3])) * scales[3];
#else
const int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q2bytes.val[0]), vget_low_s8 (q8bytes.val[0])),
vmull_s8(vget_high_s8(q2bytes.val[0]), vget_high_s8(q8bytes.val[0])));
const int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q2bytes.val[1]), vget_low_s8 (q8bytes.val[1])),
vmull_s8(vget_high_s8(q2bytes.val[1]), vget_high_s8(q8bytes.val[1])));
isum1 += vaddvq_s16(p1) * scales[0];
isum2 += vaddvq_s16(p2) * scales[1];
const int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q2bytes.val[2]), vget_low_s8 (q8bytes.val[2])),
vmull_s8(vget_high_s8(q2bytes.val[2]), vget_high_s8(q8bytes.val[2])));
const int16x8_t p4 = vaddq_s16(vmull_s8(vget_low_s8 (q2bytes.val[3]), vget_low_s8 (q8bytes.val[3])),
vmull_s8(vget_high_s8(q2bytes.val[3]), vget_high_s8(q8bytes.val[3])));
isum1 += vaddvq_s16(p3) * scales[2];
isum2 += vaddvq_s16(p4) * scales[3];
#endif
sum += d * (isum1 + isum2);
}
*s = sum;
#elif defined __AVX2__
const __m256i m3 = _mm256_set1_epi8(3);
__m256 acc = _mm256_setzero_ps();
uint32_t ud, um;
const uint8_t * restrict db = (const uint8_t *)&ud;
const uint8_t * restrict mb = (const uint8_t *)&um;
float summs = 0;
// TODO: optimize this
for (int i = 0; i < nb; ++i) {
const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin);
const uint8_t * restrict q2 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
const uint32_t * restrict sc = (const uint32_t *)x[i].scales;
ud = (sc[0] >> 0) & 0x0f0f0f0f;
um = (sc[0] >> 4) & 0x0f0f0f0f;
int32_t smin = mb[0] * y[i].bsums[0] + mb[1] * y[i].bsums[1] + mb[2] * y[i].bsums[2] + mb[3] * y[i].bsums[3];
summs += dmin * smin;
const __m128i q2bits = _mm_loadu_si128((const __m128i*)q2);
ggml : workaround for missing _mm256_setr_m128i in GCC < 8 in k_quants.c (#2405)
2023-07-28 18:17:45 +00:00
const __m256i q2_0 = _mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(q2bits, 2), q2bits), m3);
const __m256i q2_1 = _mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(q2bits, 6), _mm_srli_epi16(q2bits, 4)), m3);
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
const __m256i p0 = _mm256_maddubs_epi16(q2_0, q8_0);
const __m256i p1 = _mm256_maddubs_epi16(q2_1, q8_1);
const __m256i p_0 = _mm256_cvtepi16_epi32(_mm256_extracti128_si256(p0, 0));
const __m256i p_1 = _mm256_cvtepi16_epi32(_mm256_extracti128_si256(p0, 1));
const __m256i p_2 = _mm256_cvtepi16_epi32(_mm256_extracti128_si256(p1, 0));
const __m256i p_3 = _mm256_cvtepi16_epi32(_mm256_extracti128_si256(p1, 1));
acc = _mm256_fmadd_ps(_mm256_set1_ps(d * db[0]), _mm256_cvtepi32_ps(p_0), acc);
acc = _mm256_fmadd_ps(_mm256_set1_ps(d * db[1]), _mm256_cvtepi32_ps(p_1), acc);
acc = _mm256_fmadd_ps(_mm256_set1_ps(d * db[2]), _mm256_cvtepi32_ps(p_2), acc);
acc = _mm256_fmadd_ps(_mm256_set1_ps(d * db[3]), _mm256_cvtepi32_ps(p_3), acc);
}
*s = hsum_float_8(acc) + summs;
k_quants : add AVX support to dot functions with QK_K as 64 (#2339) * add AVX to ggml_vec_dot_q2_K_q8_K() * add AVX to ggml_vec_dot_q3_K_q8_K() * add AVX to ggml_vec_dot_q4_K_q8_K() * add AVX to ggml_vec_dot_q5_K_q8_K() * add AVX to ggml_vec_dot_q6_K_q8_K() * refactor AVX code in ggml_vec_dot_q6_K_q8_K()
2023-07-25 12:13:41 +00:00
#elif defined __AVX__
const __m128i m3 = _mm_set1_epi8(3);
__m256 acc = _mm256_setzero_ps();
uint32_t ud, um;
const uint8_t * restrict db = (const uint8_t *)&ud;
const uint8_t * restrict mb = (const uint8_t *)&um;
float summs = 0;
// TODO: optimize this
for (int i = 0; i < nb; ++i) {
const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin);
const uint8_t * restrict q2 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
const uint32_t * restrict sc = (const uint32_t *)x[i].scales;
ud = (sc[0] >> 0) & 0x0f0f0f0f;
um = (sc[0] >> 4) & 0x0f0f0f0f;
int32_t smin = mb[0] * y[i].bsums[0] + mb[1] * y[i].bsums[1] + mb[2] * y[i].bsums[2] + mb[3] * y[i].bsums[3];
summs += dmin * smin;
const __m128i q2bits = _mm_loadu_si128((const __m128i*)q2);
const __m128i q2_0 = _mm_and_si128(q2bits, m3);
const __m128i q2_1 = _mm_and_si128(_mm_srli_epi16(q2bits, 2), m3);
const __m128i q2_2 = _mm_and_si128(_mm_srli_epi16(q2bits, 4), m3);
const __m128i q2_3 = _mm_and_si128(_mm_srli_epi16(q2bits, 6), m3);
const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
const __m128i p0 = _mm_maddubs_epi16(q2_0, _mm256_extractf128_si256(q8_0, 0));
const __m128i p1 = _mm_maddubs_epi16(q2_1, _mm256_extractf128_si256(q8_0, 1));
const __m128i p2 = _mm_maddubs_epi16(q2_2, _mm256_extractf128_si256(q8_1, 0));
const __m128i p3 = _mm_maddubs_epi16(q2_3, _mm256_extractf128_si256(q8_1, 1));
ggml : workaround for missing _mm256_setr_m128i in GCC < 8 in k_quants.c (#2405)
2023-07-28 18:17:45 +00:00
const __m256i p_0 = MM256_SET_M128I(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p0, p0)), _mm_cvtepi16_epi32(p0));
const __m256i p_1 = MM256_SET_M128I(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p1, p1)), _mm_cvtepi16_epi32(p1));
const __m256i p_2 = MM256_SET_M128I(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p2, p2)), _mm_cvtepi16_epi32(p2));
const __m256i p_3 = MM256_SET_M128I(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p3, p3)), _mm_cvtepi16_epi32(p3));
k_quants : add AVX support to dot functions with QK_K as 64 (#2339) * add AVX to ggml_vec_dot_q2_K_q8_K() * add AVX to ggml_vec_dot_q3_K_q8_K() * add AVX to ggml_vec_dot_q4_K_q8_K() * add AVX to ggml_vec_dot_q5_K_q8_K() * add AVX to ggml_vec_dot_q6_K_q8_K() * refactor AVX code in ggml_vec_dot_q6_K_q8_K()
2023-07-25 12:13:41 +00:00
acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[0]), _mm256_cvtepi32_ps(p_0)), acc);
acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[1]), _mm256_cvtepi32_ps(p_1)), acc);
acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[2]), _mm256_cvtepi32_ps(p_2)), acc);
acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[3]), _mm256_cvtepi32_ps(p_3)), acc);
}
*s = hsum_float_8(acc) + summs;
ggml : add RISC-V Vector Support for K-Quants and improved the existing intrinsics (#3453) * Added RVV intrinsics support for Q8 quantize row and also improved the existing dot product function for risc-v. The RVV intrinsics is added for the following quantize row functions quantize_row_q8_0 quantize_row_q8_1 The following dot product functions have also been optimized by using LMUL = 1/2 instead of LMUL = 1 ggml_vec_dot_q4_0_q8_0 ggml_vec_dot_q4_1_q8_1 ggml_vec_dot_q5_0_q8_0 ggml_vec_dot_q5_1_q8_1 And vector initialization in Q5 by temporary array is also replaced by the vid intrinsics Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai> * Added RVV intrinsics support for k_quants This adds RISC-V Vector intrinsics support for the following K_quants functions for both QKK = 256 and QKK = 64 ggml_vec_dot_q2_K_q8_K ggml_vec_dot_q3_K_q8_K ggml_vec_dot_q4_K_q8_K ggml_vec_dot_q5_K_q8_K ggml_vec_dot_q6_K_q8_K Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai> --------- Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai>
2023-10-03 18:38:19 +00:00
#elif defined __riscv_v_intrinsic
uint32_t aux32[2];
const uint8_t * scales = (const uint8_t *)aux32;
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const float d = y[i].d * (float)x[i].d;
const float dmin = -y[i].d * (float)x[i].dmin;
const uint8_t * restrict q2 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
const uint32_t * restrict sc = (const uint32_t *)x[i].scales;
aux32[0] = sc[0] & 0x0f0f0f0f;
aux32[1] = (sc[0] >> 4) & 0x0f0f0f0f;
sumf += dmin * (scales[4] * y[i].bsums[0] + scales[5] * y[i].bsums[1] + scales[6] * y[i].bsums[2] + scales[7] * y[i].bsums[3]);
int isum1 = 0;
int isum2 = 0;
size_t vl = 16;
vint16m1_t vzero = __riscv_vmv_v_x_i16m1(0, 1);
// load Q2
vuint8mf2_t q2_x = __riscv_vle8_v_u8mf2(q2, vl);
vint8mf2_t q2_0 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(q2_x, 0x03, vl));
vint8mf2_t q2_1 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(q2_x, 0x2, vl), 0x03 , vl));
vint8mf2_t q2_2 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(q2_x, 0x4, vl), 0x03 , vl));
vint8mf2_t q2_3 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(q2_x, 0x6, vl), 0x03 , vl));
// load Q8, and take product with Q2
vint16m1_t p0 = __riscv_vwmul_vv_i16m1(q2_0, __riscv_vle8_v_i8mf2(q8, vl), vl);
vint16m1_t p1 = __riscv_vwmul_vv_i16m1(q2_1, __riscv_vle8_v_i8mf2(q8+16, vl), vl);
vint16m1_t p2 = __riscv_vwmul_vv_i16m1(q2_2, __riscv_vle8_v_i8mf2(q8+32, vl), vl);
vint16m1_t p3 = __riscv_vwmul_vv_i16m1(q2_3, __riscv_vle8_v_i8mf2(q8+48, vl), vl);
vint16m1_t vs_0 = __riscv_vredsum_vs_i16m1_i16m1(p0, vzero, vl);
vint16m1_t vs_1 = __riscv_vredsum_vs_i16m1_i16m1(p1, vzero, vl);
vint16m1_t vs_2 = __riscv_vredsum_vs_i16m1_i16m1(p2, vzero, vl);
vint16m1_t vs_3 = __riscv_vredsum_vs_i16m1_i16m1(p3, vzero, vl);
isum1 += __riscv_vmv_x_s_i16m1_i16(vs_0) * scales[0];
isum2 += __riscv_vmv_x_s_i16m1_i16(vs_1) * scales[1];
isum1 += __riscv_vmv_x_s_i16m1_i16(vs_2) * scales[2];
isum2 += __riscv_vmv_x_s_i16m1_i16(vs_3) * scales[3];
sumf += d * (isum1 + isum2);
}
*s = sumf;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
#else
float sumf = 0;
int isum[4];
for (int i = 0; i < nb; ++i) {
const uint8_t * q2 = x[i].qs;
const int8_t * q8 = y[i].qs;
const uint8_t * sc = x[i].scales;
int summs = 0;
for (int j = 0; j < QK_K/16; ++j) {
summs += y[i].bsums[j] * (sc[j] >> 4);
}
const float dall = y[i].d * ggml_fp16_to_fp32(x[i].d);
const float dmin = y[i].d * ggml_fp16_to_fp32(x[i].dmin);
isum[0] = isum[1] = isum[2] = isum[3] = 0;
for (int l = 0; l < 16; ++l) {
isum[0] += q8[l+ 0] * ((q2[l] >> 0) & 3);
isum[1] += q8[l+16] * ((q2[l] >> 2) & 3);
isum[2] += q8[l+32] * ((q2[l] >> 4) & 3);
isum[3] += q8[l+48] * ((q2[l] >> 6) & 3);
}
for (int l = 0; l < 4; ++l) {
isum[l] *= (sc[l] & 0xF);
}
sumf += dall * (isum[0] + isum[1] + isum[2] + isum[3]) - dmin * summs;
}
*s = sumf;
#endif
}
#endif
#if QK_K == 256
k-quants : allow to optionally disable at compile time (#1734) * k-quants : put behind optional compile flag LLAMA_K_QUANTS * build : enable k-quants by default
2023-06-07 07:59:52 +00:00
void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
assert(n % QK_K == 0);
const uint32_t kmask1 = 0x03030303;
const uint32_t kmask2 = 0x0f0f0f0f;
k-quants : allow to optionally disable at compile time (#1734) * k-quants : put behind optional compile flag LLAMA_K_QUANTS * build : enable k-quants by default
2023-06-07 07:59:52 +00:00
const block_q3_K * restrict x = vx;
const block_q8_K * restrict y = vy;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
const int nb = n / QK_K;
#ifdef __ARM_NEON
uint32_t aux[3];
uint32_t utmp[4];
const uint8x16_t m3b = vdupq_n_u8(0x3);
#ifdef __ARM_FEATURE_DOTPROD
const int32x4_t vzero = vdupq_n_s32(0);
#endif
const uint8x16_t m0 = vdupq_n_u8(1);
const uint8x16_t m1 = vshlq_n_u8(m0, 1);
const uint8x16_t m2 = vshlq_n_u8(m0, 2);
const uint8x16_t m3 = vshlq_n_u8(m0, 3);
const int8_t m32 = 32;
int8x16x4_t q3bytes;
float sum = 0;
for (int i = 0; i < nb; ++i) {
const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
const uint8_t * restrict q3 = x[i].qs;
const uint8_t * restrict qh = x[i].hmask;
const int8_t * restrict q8 = y[i].qs;
uint8x16x2_t qhbits = vld1q_u8_x2(qh);
uint8x16x4_t q3h;
int32_t isum = 0;
// Set up scales
memcpy(aux, x[i].scales, 12);
utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4);
utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4);
utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4);
utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4);
int8_t * scale = (int8_t *)utmp;
for (int j = 0; j < 16; ++j) scale[j] -= m32;
for (int j = 0; j < QK_K/128; ++j) {
const uint8x16x2_t q3bits = vld1q_u8_x2(q3); q3 += 32;
Revert "ggml : load data into int8x16x4_t using vld4q_s8 on arm64 (#1738)" This reverts commit 8432d4d9f716b25133e3ed671d91e21f6f3be867.
2023-06-08 17:48:14 +00:00
const int8x16x4_t q8bytes_1 = vld1q_s8_x4(q8); q8 += 64;
const int8x16x4_t q8bytes_2 = vld1q_s8_x4(q8); q8 += 64;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
q3h.val[0] = vshlq_n_u8(vbicq_u8(m0, qhbits.val[0]), 2);
q3h.val[1] = vshlq_n_u8(vbicq_u8(m0, qhbits.val[1]), 2);
q3h.val[2] = vshlq_n_u8(vbicq_u8(m1, qhbits.val[0]), 1);
q3h.val[3] = vshlq_n_u8(vbicq_u8(m1, qhbits.val[1]), 1);
q3bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(q3bits.val[0], m3b)), vreinterpretq_s8_u8(q3h.val[0]));
q3bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(q3bits.val[1], m3b)), vreinterpretq_s8_u8(q3h.val[1]));
q3bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[0], 2), m3b)), vreinterpretq_s8_u8(q3h.val[2]));
q3bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[1], 2), m3b)), vreinterpretq_s8_u8(q3h.val[3]));
#if defined(__ARM_FEATURE_DOTPROD)
isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[0], q8bytes_1.val[0])) * scale[0];
isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[1], q8bytes_1.val[1])) * scale[1];
isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[2], q8bytes_1.val[2])) * scale[2];
isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[3], q8bytes_1.val[3])) * scale[3];
#else
int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[0]), vget_low_s8 (q8bytes_1.val[0])),
vmull_s8(vget_high_s8(q3bytes.val[0]), vget_high_s8(q8bytes_1.val[0])));
int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[1]), vget_low_s8 (q8bytes_1.val[1])),
vmull_s8(vget_high_s8(q3bytes.val[1]), vget_high_s8(q8bytes_1.val[1])));
int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[2]), vget_low_s8 (q8bytes_1.val[2])),
vmull_s8(vget_high_s8(q3bytes.val[2]), vget_high_s8(q8bytes_1.val[2])));
int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[3]), vget_low_s8 (q8bytes_1.val[3])),
vmull_s8(vget_high_s8(q3bytes.val[3]), vget_high_s8(q8bytes_1.val[3])));
isum += vaddvq_s16(p0) * scale[0] + vaddvq_s16(p1) * scale[1] + vaddvq_s16(p2) * scale[2] + vaddvq_s16(p3) * scale[3];
#endif
scale += 4;
q3h.val[0] = vbicq_u8(m2, qhbits.val[0]);
q3h.val[1] = vbicq_u8(m2, qhbits.val[1]);
q3h.val[2] = vshrq_n_u8(vbicq_u8(m3, qhbits.val[0]), 1);
q3h.val[3] = vshrq_n_u8(vbicq_u8(m3, qhbits.val[1]), 1);
q3bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[0], 4), m3b)), vreinterpretq_s8_u8(q3h.val[0]));
q3bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[1], 4), m3b)), vreinterpretq_s8_u8(q3h.val[1]));
q3bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[0], 6), m3b)), vreinterpretq_s8_u8(q3h.val[2]));
q3bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[1], 6), m3b)), vreinterpretq_s8_u8(q3h.val[3]));
#if defined(__ARM_FEATURE_DOTPROD)
isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[0], q8bytes_2.val[0])) * scale[0];
isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[1], q8bytes_2.val[1])) * scale[1];
isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[2], q8bytes_2.val[2])) * scale[2];
isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[3], q8bytes_2.val[3])) * scale[3];
#else
p0 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[0]), vget_low_s8 (q8bytes_2.val[0])),
vmull_s8(vget_high_s8(q3bytes.val[0]), vget_high_s8(q8bytes_2.val[0])));
p1 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[1]), vget_low_s8 (q8bytes_2.val[1])),
vmull_s8(vget_high_s8(q3bytes.val[1]), vget_high_s8(q8bytes_2.val[1])));
p2 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[2]), vget_low_s8 (q8bytes_2.val[2])),
vmull_s8(vget_high_s8(q3bytes.val[2]), vget_high_s8(q8bytes_2.val[2])));
p3 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[3]), vget_low_s8 (q8bytes_2.val[3])),
vmull_s8(vget_high_s8(q3bytes.val[3]), vget_high_s8(q8bytes_2.val[3])));
isum += vaddvq_s16(p0) * scale[0] + vaddvq_s16(p1) * scale[1] + vaddvq_s16(p2) * scale[2] + vaddvq_s16(p3) * scale[3];
#endif
scale += 4;
if (j == 0) {
qhbits.val[0] = vshrq_n_u8(qhbits.val[0], 4);
qhbits.val[1] = vshrq_n_u8(qhbits.val[1], 4);
}
}
sum += d * isum;
}
*s = sum;
#elif defined __AVX2__
const __m256i m3 = _mm256_set1_epi8(3);
const __m256i mone = _mm256_set1_epi8(1);
const __m128i m32 = _mm_set1_epi8(32);
__m256 acc = _mm256_setzero_ps();
uint32_t aux[3];
for (int i = 0; i < nb; ++i) {
const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
const uint8_t * restrict q3 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
// Set up scales
memcpy(aux, x[i].scales, 12);
__m128i scales128 = _mm_set_epi32(
((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4),
((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4),
(aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4),
(aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4));
scales128 = _mm_sub_epi8(scales128, m32);
const __m256i all_scales = _mm256_cvtepi8_epi16(scales128);
const __m128i l_scales = _mm256_extracti128_si256(all_scales, 0);
const __m128i h_scales = _mm256_extracti128_si256(all_scales, 1);
ggml : workaround for missing _mm256_setr_m128i in GCC < 8 in k_quants.c (#2405)
2023-07-28 18:17:45 +00:00
const __m256i scales[2] = {MM256_SET_M128I(l_scales, l_scales), MM256_SET_M128I(h_scales, h_scales)};
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
// high bit
const __m256i hbits = _mm256_loadu_si256((const __m256i*)x[i].hmask);
// integer accumulator
__m256i sumi = _mm256_setzero_si256();
int bit = 0;
int is = 0;
for (int j = 0; j < QK_K/128; ++j) {
// load low 2 bits
const __m256i q3bits = _mm256_loadu_si256((const __m256i*)q3); q3 += 32;
// prepare low and high bits
const __m256i q3l_0 = _mm256_and_si256(q3bits, m3);
const __m256i q3h_0 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
++bit;
const __m256i q3l_1 = _mm256_and_si256(_mm256_srli_epi16(q3bits, 2), m3);
const __m256i q3h_1 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
++bit;
const __m256i q3l_2 = _mm256_and_si256(_mm256_srli_epi16(q3bits, 4), m3);
const __m256i q3h_2 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
++bit;
const __m256i q3l_3 = _mm256_and_si256(_mm256_srli_epi16(q3bits, 6), m3);
const __m256i q3h_3 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
++bit;
// load Q8 quants
const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
const __m256i q8_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
const __m256i q8_3 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
// Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use _mm256_maddubs_epi16,
// and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set,
// and 2 if the high bit was set)
__m256i q8s_0 = _mm256_maddubs_epi16(q3h_0, q8_0);
__m256i q8s_1 = _mm256_maddubs_epi16(q3h_1, q8_1);
__m256i q8s_2 = _mm256_maddubs_epi16(q3h_2, q8_2);
__m256i q8s_3 = _mm256_maddubs_epi16(q3h_3, q8_3);
__m256i p16_0 = _mm256_maddubs_epi16(q3l_0, q8_0);
__m256i p16_1 = _mm256_maddubs_epi16(q3l_1, q8_1);
__m256i p16_2 = _mm256_maddubs_epi16(q3l_2, q8_2);
__m256i p16_3 = _mm256_maddubs_epi16(q3l_3, q8_3);
p16_0 = _mm256_sub_epi16(p16_0, q8s_0);
p16_1 = _mm256_sub_epi16(p16_1, q8s_1);
p16_2 = _mm256_sub_epi16(p16_2, q8s_2);
p16_3 = _mm256_sub_epi16(p16_3, q8s_3);
// multiply with scales
p16_0 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 0)), p16_0);
p16_1 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 1)), p16_1);
p16_2 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 2)), p16_2);
p16_3 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 3)), p16_3);
// accumulate
p16_0 = _mm256_add_epi32(p16_0, p16_1);
p16_2 = _mm256_add_epi32(p16_2, p16_3);
sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_2));
}
// multiply with block scale and accumulate
acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
}
*s = hsum_float_8(acc);
k-quants : add AVX support to dot functions (#1916) * k_quants : add AVX support * k_quants : apply review comments
2023-06-26 16:46:07 +00:00
#elif defined __AVX__
const __m128i m3 = _mm_set1_epi8(3);
const __m128i mone = _mm_set1_epi8(1);
const __m128i m32 = _mm_set1_epi8(32);
const __m128i m2 = _mm_set1_epi8(2);
__m256 acc = _mm256_setzero_ps();
build : fix most gcc and clang warnings (#2861) * fix most gcc and clang warnings * baby-llama : remove commented opt_params_adam * fix some MinGW warnings * fix more MinGW warnings
2023-09-01 13:34:50 +00:00
const uint32_t *aux;
k-quants : add AVX support to dot functions (#1916) * k_quants : add AVX support * k_quants : apply review comments
2023-06-26 16:46:07 +00:00
for (int i = 0; i < nb; ++i) {
const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
const uint8_t * restrict q3 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
// Set up scales
build : fix most gcc and clang warnings (#2861) * fix most gcc and clang warnings * baby-llama : remove commented opt_params_adam * fix some MinGW warnings * fix more MinGW warnings
2023-09-01 13:34:50 +00:00
aux = (const uint32_t *)x[i].scales;
k-quants : add AVX support to dot functions (#1916) * k_quants : add AVX support * k_quants : apply review comments
2023-06-26 16:46:07 +00:00
__m128i scales128 = _mm_set_epi32(
((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4),
((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4),
(aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4),
(aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4));
scales128 = _mm_sub_epi8(scales128, m32);
const __m128i scales_0 = _mm_cvtepi8_epi16(scales128);
const __m128i scales_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(scales128, scales128));
const __m128i scales[2] = { scales_0, scales_1 };
// high bit *128*2 from block_q3_K.hmask[QK_K/8]
const __m128i hbits_0 = _mm_loadu_si128((const __m128i*)&x[i].hmask[0]);
const __m128i hbits_1 = _mm_loadu_si128((const __m128i*)&x[i].hmask[16]);
// integer accumulator
__m128i sumi_0 = _mm_setzero_si128();
__m128i sumi_1 = _mm_setzero_si128();
for (int j = 0; j < QK_K/128; ++j) {
// load low 2 bits *64*2 from block_q3_K.qs[QK_K/4]
const __m128i q3bits_0 = _mm_loadu_si128((const __m128i*)q3); q3 += 16;
const __m128i q3bits_1 = _mm_loadu_si128((const __m128i*)q3); q3 += 16;
// prepare low and high bits
k-quants : fix indentation
2023-06-26 17:10:52 +00:00
const int bit = j << 2;
k-quants : add AVX support to dot functions (#1916) * k_quants : add AVX support * k_quants : apply review comments
2023-06-26 16:46:07 +00:00
const __m128i q3l_0 = _mm_and_si128(q3bits_0, m3);
const __m128i q3l_1 = _mm_and_si128(q3bits_1, m3);
const __m128i q3h_0 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit)), bit), 2);
const __m128i q3h_1 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit)), bit), 2);
const __m128i q3l_2 = _mm_and_si128(_mm_srli_epi16(q3bits_0, 2), m3);
const __m128i q3l_3 = _mm_and_si128(_mm_srli_epi16(q3bits_1, 2), m3);
const __m128i q3h_2 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit+1)), bit+1), 2);
const __m128i q3h_3 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit+1)), bit+1), 2);
const __m128i q3l_4 = _mm_and_si128(_mm_srli_epi16(q3bits_0, 4), m3);
const __m128i q3l_5 = _mm_and_si128(_mm_srli_epi16(q3bits_1, 4), m3);
const __m128i q3h_4 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit+2)), bit+2), 2);
const __m128i q3h_5 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit+2)), bit+2), 2);
const __m128i q3l_6 = _mm_and_si128(_mm_srli_epi16(q3bits_0, 6), m3);
const __m128i q3l_7 = _mm_and_si128(_mm_srli_epi16(q3bits_1, 6), m3);
const __m128i q3h_6 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit+3)), bit+3), 2);
const __m128i q3h_7 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit+3)), bit+3), 2);
// load Q8 quants from block_q8_K.qs[QK_K]
const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
// Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use _mm256_maddubs_epi16,
// and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set,
// and 2 if the high bit was set)
__m128i q8s_0 = _mm_maddubs_epi16(q3h_0, q8_0);
__m128i q8s_1 = _mm_maddubs_epi16(q3h_1, q8_1);
__m128i q8s_2 = _mm_maddubs_epi16(q3h_2, q8_2);
__m128i q8s_3 = _mm_maddubs_epi16(q3h_3, q8_3);
__m128i q8s_4 = _mm_maddubs_epi16(q3h_4, q8_4);
__m128i q8s_5 = _mm_maddubs_epi16(q3h_5, q8_5);
__m128i q8s_6 = _mm_maddubs_epi16(q3h_6, q8_6);
__m128i q8s_7 = _mm_maddubs_epi16(q3h_7, q8_7);
__m128i p16_0 = _mm_maddubs_epi16(q3l_0, q8_0);
__m128i p16_1 = _mm_maddubs_epi16(q3l_1, q8_1);
__m128i p16_2 = _mm_maddubs_epi16(q3l_2, q8_2);
__m128i p16_3 = _mm_maddubs_epi16(q3l_3, q8_3);
__m128i p16_4 = _mm_maddubs_epi16(q3l_4, q8_4);
__m128i p16_5 = _mm_maddubs_epi16(q3l_5, q8_5);
__m128i p16_6 = _mm_maddubs_epi16(q3l_6, q8_6);
__m128i p16_7 = _mm_maddubs_epi16(q3l_7, q8_7);
p16_0 = _mm_sub_epi16(p16_0, q8s_0);
p16_1 = _mm_sub_epi16(p16_1, q8s_1);
p16_2 = _mm_sub_epi16(p16_2, q8s_2);
p16_3 = _mm_sub_epi16(p16_3, q8s_3);
p16_4 = _mm_sub_epi16(p16_4, q8s_4);
p16_5 = _mm_sub_epi16(p16_5, q8s_5);
p16_6 = _mm_sub_epi16(p16_6, q8s_6);
p16_7 = _mm_sub_epi16(p16_7, q8s_7);
// multiply with scales
__m128i shuffle = _mm_set1_epi16(0x0100);
p16_0 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_0);
shuffle = _mm_add_epi16(shuffle, m2);
p16_1 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_1);
shuffle = _mm_add_epi16(shuffle, m2);
p16_2 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_2);
shuffle = _mm_add_epi16(shuffle, m2);
p16_3 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_3);
shuffle = _mm_add_epi16(shuffle, m2);
p16_4 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_4);
shuffle = _mm_add_epi16(shuffle, m2);
p16_5 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_5);
shuffle = _mm_add_epi16(shuffle, m2);
p16_6 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_6);
shuffle = _mm_add_epi16(shuffle, m2);
p16_7 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_7);
// accumulate
p16_0 = _mm_add_epi32(p16_0, p16_1);
p16_2 = _mm_add_epi32(p16_2, p16_3);
p16_4 = _mm_add_epi32(p16_4, p16_5);
p16_6 = _mm_add_epi32(p16_6, p16_7);
sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_4, p16_6));
}
// multiply with block scale and accumulate
ggml : workaround for missing _mm256_setr_m128i in GCC < 8 in k_quants.c (#2405)
2023-07-28 18:17:45 +00:00
__m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
k-quants : add AVX support to dot functions (#1916) * k_quants : add AVX support * k_quants : apply review comments
2023-06-26 16:46:07 +00:00
acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi)), acc);
}
*s = hsum_float_8(acc);
ggml : add RISC-V Vector Support for K-Quants and improved the existing intrinsics (#3453) * Added RVV intrinsics support for Q8 quantize row and also improved the existing dot product function for risc-v. The RVV intrinsics is added for the following quantize row functions quantize_row_q8_0 quantize_row_q8_1 The following dot product functions have also been optimized by using LMUL = 1/2 instead of LMUL = 1 ggml_vec_dot_q4_0_q8_0 ggml_vec_dot_q4_1_q8_1 ggml_vec_dot_q5_0_q8_0 ggml_vec_dot_q5_1_q8_1 And vector initialization in Q5 by temporary array is also replaced by the vid intrinsics Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai> * Added RVV intrinsics support for k_quants This adds RISC-V Vector intrinsics support for the following K_quants functions for both QKK = 256 and QKK = 64 ggml_vec_dot_q2_K_q8_K ggml_vec_dot_q3_K_q8_K ggml_vec_dot_q4_K_q8_K ggml_vec_dot_q5_K_q8_K ggml_vec_dot_q6_K_q8_K Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai> --------- Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai>
2023-10-03 18:38:19 +00:00
#elif defined __riscv_v_intrinsic
uint32_t aux[3];
uint32_t utmp[4];
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * restrict q3 = x[i].qs;
const uint8_t * restrict qh = x[i].hmask;
const int8_t * restrict q8 = y[i].qs;
memcpy(aux, x[i].scales, 12);
utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4);
utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4);
utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4);
utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4);
int8_t * scale = (int8_t *)utmp;
for (int j = 0; j < 16; ++j) scale[j] -= 32;
size_t vl = 32;
uint8_t m = 1;
vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
vuint8m1_t vqh = __riscv_vle8_v_u8m1(qh, vl);
int sum_t = 0;
for (int j = 0; j < QK_K; j += 128) {
vl = 32;
// load Q3
vuint8m1_t q3_x = __riscv_vle8_v_u8m1(q3, vl);
vint8m1_t q3_0 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q3_x, 0x03, vl));
vint8m1_t q3_1 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q3_x, 0x2, vl), 0x03 , vl));
vint8m1_t q3_2 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q3_x, 0x4, vl), 0x03 , vl));
vint8m1_t q3_3 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q3_x, 0x6, vl), 0x03 , vl));
// compute mask for subtraction
vuint8m1_t qh_m0 = __riscv_vand_vx_u8m1(vqh, m, vl);
vbool8_t vmask_0 = __riscv_vmseq_vx_u8m1_b8(qh_m0, 0, vl);
vint8m1_t q3_m0 = __riscv_vsub_vx_i8m1_m(vmask_0, q3_0, 0x4, vl);
m <<= 1;
vuint8m1_t qh_m1 = __riscv_vand_vx_u8m1(vqh, m, vl);
vbool8_t vmask_1 = __riscv_vmseq_vx_u8m1_b8(qh_m1, 0, vl);
vint8m1_t q3_m1 = __riscv_vsub_vx_i8m1_m(vmask_1, q3_1, 0x4, vl);
m <<= 1;
vuint8m1_t qh_m2 = __riscv_vand_vx_u8m1(vqh, m, vl);
vbool8_t vmask_2 = __riscv_vmseq_vx_u8m1_b8(qh_m2, 0, vl);
vint8m1_t q3_m2 = __riscv_vsub_vx_i8m1_m(vmask_2, q3_2, 0x4, vl);
m <<= 1;
vuint8m1_t qh_m3 = __riscv_vand_vx_u8m1(vqh, m, vl);
vbool8_t vmask_3 = __riscv_vmseq_vx_u8m1_b8(qh_m3, 0, vl);
vint8m1_t q3_m3 = __riscv_vsub_vx_i8m1_m(vmask_3, q3_3, 0x4, vl);
m <<= 1;
// load Q8 and take product with Q3
vint16m2_t a0 = __riscv_vwmul_vv_i16m2(q3_m0, __riscv_vle8_v_i8m1(q8, vl), vl);
vint16m2_t a1 = __riscv_vwmul_vv_i16m2(q3_m1, __riscv_vle8_v_i8m1(q8+32, vl), vl);
vint16m2_t a2 = __riscv_vwmul_vv_i16m2(q3_m2, __riscv_vle8_v_i8m1(q8+64, vl), vl);
vint16m2_t a3 = __riscv_vwmul_vv_i16m2(q3_m3, __riscv_vle8_v_i8m1(q8+96, vl), vl);
vl = 16;
// retreive lane to multiply with scale
vint32m2_t aux0_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 0), (scale[0]), vl);
vint32m2_t aux0_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 1), (scale[1]), vl);
vint32m2_t aux1_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a1, 0), (scale[2]), vl);
vint32m2_t aux1_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a1, 1), (scale[3]), vl);
vint32m2_t aux2_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a2, 0), (scale[4]), vl);
vint32m2_t aux2_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a2, 1), (scale[5]), vl);
vint32m2_t aux3_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a3, 0), (scale[6]), vl);
vint32m2_t aux3_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a3, 1), (scale[7]), vl);
vint32m1_t isum0 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux0_0, aux0_1, vl), vzero, vl);
vint32m1_t isum1 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux1_0, aux1_1, vl), isum0, vl);
vint32m1_t isum2 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux2_0, aux2_1, vl), isum1, vl);
vint32m1_t isum3 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux3_0, aux3_1, vl), isum2, vl);
sum_t += __riscv_vmv_x_s_i32m1_i32(isum3);
q3 += 32; q8 += 128; scale += 8;
}
const float d = ggml_fp16_to_fp32(x[i].d) * y[i].d;
sumf += d*sum_t;
}
*s = sumf;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
#else
// scalar version
// This function is written like this so the compiler can manage to vectorize most of it
// Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
// manually vectorized version above. Every other version I tried would run at least 4 times slower.
// The ideal situation would be if we could just write the code once, and the compiler would
// automatically produce the best possible set of machine instructions, instead of us having to manually
// write vectorized versions for AVX, ARM_NEON, etc.
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
uint32_t auxs[4];
const int8_t * scales = (const int8_t*)auxs;
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * restrict q3 = x[i].qs;
const uint8_t * restrict hm = x[i].hmask;
const int8_t * restrict q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * restrict a = aux8;
uint8_t m = 1;
for (int j = 0; j < QK_K; j += 128) {
for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
a += 32; m <<= 1;
q3 += 32;
}
a = aux8;
memcpy(auxs, x[i].scales, 12);
uint32_t tmp = auxs[2];
auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
for (int j = 0; j < QK_K/16; ++j) {
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
q8 += 8; a += 8;
}
const float d = ggml_fp16_to_fp32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
#endif
}
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
#else
void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
assert(n % QK_K == 0);
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
const block_q3_K * restrict x = vx;
k-quants : allow to optionally disable at compile time (#1734) * k-quants : put behind optional compile flag LLAMA_K_QUANTS * build : enable k-quants by default
2023-06-07 07:59:52 +00:00
const block_q8_K * restrict y = vy;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
const int nb = n / QK_K;
#ifdef __ARM_NEON
#ifdef __ARM_FEATURE_DOTPROD
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
const int32x4_t vzero = vdupq_n_s32(0);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
#endif
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
const uint8x16_t m3b = vdupq_n_u8(0x3);
const uint8x16_t mh = vdupq_n_u8(4);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
int8x16x4_t q3bytes;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
uint16_t aux16[2];
int8_t * scales = (int8_t *)aux16;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
float sum = 0;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
for (int i = 0; i < nb; ++i) {
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
uint8x16x4_t q3h;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
const uint8x8_t hbits = vld1_u8(x[i].hmask);
const uint8x16_t q3bits = vld1q_u8(x[i].qs);
const int8x16x4_t q8bytes = vld1q_s8_x4(y[i].qs);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
const uint16_t a = *(const uint16_t *)x[i].scales;
aux16[0] = a & 0x0f0f;
aux16[1] = (a >> 4) & 0x0f0f;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
for (int j = 0; j < 4; ++j) scales[j] -= 8;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
int32_t isum = -4*(scales[0] * y[i].bsums[0] + scales[2] * y[i].bsums[1] + scales[1] * y[i].bsums[2] + scales[3] * y[i].bsums[3]);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
const float d = y[i].d * (float)x[i].d;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
const uint8x16_t htmp = vcombine_u8(hbits, vshr_n_u8(hbits, 1));
q3h.val[0] = vandq_u8(mh, vshlq_n_u8(htmp, 2));
q3h.val[1] = vandq_u8(mh, htmp);
q3h.val[2] = vandq_u8(mh, vshrq_n_u8(htmp, 2));
q3h.val[3] = vandq_u8(mh, vshrq_n_u8(htmp, 4));
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
q3bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q3bits, m3b), q3h.val[0]));
q3bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(vshrq_n_u8(q3bits, 2), m3b), q3h.val[1]));
q3bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(vshrq_n_u8(q3bits, 4), m3b), q3h.val[2]));
q3bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q3bits, 6), q3h.val[3]));
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
#if defined(__ARM_FEATURE_DOTPROD)
isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[0], q8bytes.val[0])) * scales[0];
isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[1], q8bytes.val[1])) * scales[2];
isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[2], q8bytes.val[2])) * scales[1];
isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[3], q8bytes.val[3])) * scales[3];
#else
const int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[0]), vget_low_s8 (q8bytes.val[0])),
vmull_s8(vget_high_s8(q3bytes.val[0]), vget_high_s8(q8bytes.val[0])));
const int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[1]), vget_low_s8 (q8bytes.val[1])),
vmull_s8(vget_high_s8(q3bytes.val[1]), vget_high_s8(q8bytes.val[1])));
const int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[2]), vget_low_s8 (q8bytes.val[2])),
vmull_s8(vget_high_s8(q3bytes.val[2]), vget_high_s8(q8bytes.val[2])));
const int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[3]), vget_low_s8 (q8bytes.val[3])),
vmull_s8(vget_high_s8(q3bytes.val[3]), vget_high_s8(q8bytes.val[3])));
isum += vaddvq_s16(p0) * scales[0] + vaddvq_s16(p1) * scales[2] + vaddvq_s16(p2) * scales[1] + vaddvq_s16(p3) * scales[3];
#endif
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
sum += d * isum;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
}
*s = sum;
#elif defined __AVX2__
const __m256i m3 = _mm256_set1_epi8(3);
const __m256i m1 = _mm256_set1_epi8(1);
__m256 acc = _mm256_setzero_ps();
uint64_t aux64;
uint16_t aux16[2];
const int8_t * aux8 = (const int8_t *)aux16;
for (int i = 0; i < nb; ++i) {
const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
const uint8_t * restrict q3 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
const uint16_t a = *(const uint16_t *)x[i].scales;
aux16[0] = a & 0x0f0f;
aux16[1] = (a >> 4) & 0x0f0f;
ggml : workaround for missing _mm256_setr_m128i in GCC < 8 in k_quants.c (#2405)
2023-07-28 18:17:45 +00:00
const __m256i scale_0 = MM256_SET_M128I(_mm_set1_epi16(aux8[2] - 8), _mm_set1_epi16(aux8[0] - 8));
const __m256i scale_1 = MM256_SET_M128I(_mm_set1_epi16(aux8[3] - 8), _mm_set1_epi16(aux8[1] - 8));
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
memcpy(&aux64, x[i].hmask, 8);
const __m128i haux = _mm_set_epi64x(aux64 >> 1, aux64 >> 0);
ggml : workaround for missing _mm256_setr_m128i in GCC < 8 in k_quants.c (#2405)
2023-07-28 18:17:45 +00:00
__m256i q3h_0 = MM256_SET_M128I(_mm_srli_epi16(haux, 2), haux);
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
__m256i q3h_1 = _mm256_srli_epi16(q3h_0, 4);
q3h_0 = _mm256_slli_epi16(_mm256_andnot_si256(q3h_0, m1), 2);
q3h_1 = _mm256_slli_epi16(_mm256_andnot_si256(q3h_1, m1), 2);
// load low 2 bits
const __m128i q3bits = _mm_loadu_si128((const __m128i*)q3);
// prepare low and high bits
ggml : workaround for missing _mm256_setr_m128i in GCC < 8 in k_quants.c (#2405)
2023-07-28 18:17:45 +00:00
const __m256i q3aux = MM256_SET_M128I(_mm_srli_epi16(q3bits, 2), q3bits);
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
const __m256i q3l_0 = _mm256_and_si256(q3aux, m3);
const __m256i q3l_1 = _mm256_and_si256(_mm256_srli_epi16(q3aux, 4), m3);
// load Q8 quants
const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
// Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use _mm256_maddubs_epi16,
// and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set,
// and 2 if the high bit was set)
const __m256i q8s_0 = _mm256_maddubs_epi16(q3h_0, q8_0);
const __m256i q8s_1 = _mm256_maddubs_epi16(q3h_1, q8_1);
__m256i p16_0 = _mm256_maddubs_epi16(q3l_0, q8_0);
__m256i p16_1 = _mm256_maddubs_epi16(q3l_1, q8_1);
p16_0 = _mm256_sub_epi16(p16_0, q8s_0);
p16_1 = _mm256_sub_epi16(p16_1, q8s_1);
// multiply with scales
p16_0 = _mm256_madd_epi16(scale_0, p16_0);
p16_1 = _mm256_madd_epi16(scale_1, p16_1);
p16_0 = _mm256_add_epi32(p16_0, p16_1);
// multiply with block scale and accumulate
acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(p16_0), acc);
}
*s = hsum_float_8(acc);
k_quants : add AVX support to dot functions with QK_K as 64 (#2339) * add AVX to ggml_vec_dot_q2_K_q8_K() * add AVX to ggml_vec_dot_q3_K_q8_K() * add AVX to ggml_vec_dot_q4_K_q8_K() * add AVX to ggml_vec_dot_q5_K_q8_K() * add AVX to ggml_vec_dot_q6_K_q8_K() * refactor AVX code in ggml_vec_dot_q6_K_q8_K()
2023-07-25 12:13:41 +00:00
#elif defined __AVX__
const __m128i m3 = _mm_set1_epi8(3);
const __m128i m1 = _mm_set1_epi8(1);
__m256 acc = _mm256_setzero_ps();
uint64_t aux64;
uint16_t aux16[2];
const int8_t * aux8 = (const int8_t *)aux16;
for (int i = 0; i < nb; ++i) {
const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
const uint8_t * restrict q3 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
const uint16_t a = *(const uint16_t *)x[i].scales;
aux16[0] = a & 0x0f0f;
aux16[1] = (a >> 4) & 0x0f0f;
const __m128i scale_0 = _mm_set1_epi16(aux8[0] - 8);
const __m128i scale_1 = _mm_set1_epi16(aux8[2] - 8);
const __m128i scale_2 = _mm_set1_epi16(aux8[1] - 8);
const __m128i scale_3 = _mm_set1_epi16(aux8[3] - 8);
memcpy(&aux64, x[i].hmask, 8);
__m128i q3h_0 = _mm_set_epi64x(aux64 >> 1, aux64 >> 0);
__m128i q3h_1 = _mm_srli_epi16(q3h_0, 2);
__m128i q3h_2 = _mm_srli_epi16(q3h_0, 4);
__m128i q3h_3 = _mm_srli_epi16(q3h_0, 6);
q3h_0 = _mm_slli_epi16(_mm_andnot_si128(q3h_0, m1), 2);
q3h_1 = _mm_slli_epi16(_mm_andnot_si128(q3h_1, m1), 2);
q3h_2 = _mm_slli_epi16(_mm_andnot_si128(q3h_2, m1), 2);
q3h_3 = _mm_slli_epi16(_mm_andnot_si128(q3h_3, m1), 2);
// load low 2 bits
const __m128i q3bits = _mm_loadu_si128((const __m128i*)q3);
// prepare low and high bits
const __m128i q3l_0 = _mm_and_si128(q3bits, m3);
const __m128i q3l_1 = _mm_and_si128(_mm_srli_epi16(q3bits, 2), m3);
const __m128i q3l_2 = _mm_and_si128(_mm_srli_epi16(q3bits, 4), m3);
const __m128i q3l_3 = _mm_and_si128(_mm_srli_epi16(q3bits, 6), m3);
// load Q8 quants
const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
// Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use _mm_maddubs_epi16,
// and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set,
// and 2 if the high bit was set)
const __m128i q8s_0 = _mm_maddubs_epi16(q3h_0, _mm256_extractf128_si256(q8_0, 0));
const __m128i q8s_1 = _mm_maddubs_epi16(q3h_1, _mm256_extractf128_si256(q8_0, 1));
const __m128i q8s_2 = _mm_maddubs_epi16(q3h_2, _mm256_extractf128_si256(q8_1, 0));
const __m128i q8s_3 = _mm_maddubs_epi16(q3h_3, _mm256_extractf128_si256(q8_1, 1));
__m128i p16_0 = _mm_maddubs_epi16(q3l_0, _mm256_extractf128_si256(q8_0, 0));
__m128i p16_1 = _mm_maddubs_epi16(q3l_1, _mm256_extractf128_si256(q8_0, 1));
__m128i p16_2 = _mm_maddubs_epi16(q3l_2, _mm256_extractf128_si256(q8_1, 0));
__m128i p16_3 = _mm_maddubs_epi16(q3l_3, _mm256_extractf128_si256(q8_1, 1));
p16_0 = _mm_sub_epi16(p16_0, q8s_0);
p16_1 = _mm_sub_epi16(p16_1, q8s_1);
p16_2 = _mm_sub_epi16(p16_2, q8s_2);
p16_3 = _mm_sub_epi16(p16_3, q8s_3);
// multiply with scales
p16_0 = _mm_madd_epi16(scale_0, p16_0);
p16_1 = _mm_madd_epi16(scale_1, p16_1);
p16_2 = _mm_madd_epi16(scale_2, p16_2);
p16_3 = _mm_madd_epi16(scale_3, p16_3);
p16_0 = _mm_add_epi32(p16_0, p16_2);
p16_1 = _mm_add_epi32(p16_1, p16_3);
ggml : workaround for missing _mm256_setr_m128i in GCC < 8 in k_quants.c (#2405)
2023-07-28 18:17:45 +00:00
__m256i p16 = MM256_SET_M128I(p16_1, p16_0);
k_quants : add AVX support to dot functions with QK_K as 64 (#2339) * add AVX to ggml_vec_dot_q2_K_q8_K() * add AVX to ggml_vec_dot_q3_K_q8_K() * add AVX to ggml_vec_dot_q4_K_q8_K() * add AVX to ggml_vec_dot_q5_K_q8_K() * add AVX to ggml_vec_dot_q6_K_q8_K() * refactor AVX code in ggml_vec_dot_q6_K_q8_K()
2023-07-25 12:13:41 +00:00
// multiply with block scale and accumulate
acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(p16)), acc);
}
*s = hsum_float_8(acc);
ggml : add RISC-V Vector Support for K-Quants and improved the existing intrinsics (#3453) * Added RVV intrinsics support for Q8 quantize row and also improved the existing dot product function for risc-v. The RVV intrinsics is added for the following quantize row functions quantize_row_q8_0 quantize_row_q8_1 The following dot product functions have also been optimized by using LMUL = 1/2 instead of LMUL = 1 ggml_vec_dot_q4_0_q8_0 ggml_vec_dot_q4_1_q8_1 ggml_vec_dot_q5_0_q8_0 ggml_vec_dot_q5_1_q8_1 And vector initialization in Q5 by temporary array is also replaced by the vid intrinsics Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai> * Added RVV intrinsics support for k_quants This adds RISC-V Vector intrinsics support for the following K_quants functions for both QKK = 256 and QKK = 64 ggml_vec_dot_q2_K_q8_K ggml_vec_dot_q3_K_q8_K ggml_vec_dot_q4_K_q8_K ggml_vec_dot_q5_K_q8_K ggml_vec_dot_q6_K_q8_K Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai> --------- Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai>
2023-10-03 18:38:19 +00:00
#elif defined __riscv_v_intrinsic
uint16_t aux16[2];
int8_t * scales = (int8_t *)aux16;
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * restrict q3 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
const uint16_t a = *(const uint16_t *)x[i].scales;
aux16[0] = a & 0x0f0f;
aux16[1] = (a >> 4) & 0x0f0f;
for (int j = 0; j < 4; ++j) scales[j] -= 8;
int32_t isum = -4*(scales[0] * y[i].bsums[0] + scales[2] * y[i].bsums[1] + scales[1] * y[i].bsums[2] + scales[3] * y[i].bsums[3]);
const float d = y[i].d * (float)x[i].d;
vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
// load qh
vuint8mf4_t qh_x1 = __riscv_vle8_v_u8mf4(x[i].hmask, 8);
vuint8mf2_t qh_x2 = __riscv_vlmul_ext_v_u8mf4_u8mf2(__riscv_vsrl_vx_u8mf4(qh_x1, 1, 8));
size_t vl = 16;
// extend and combine both qh_x1 and qh_x2
vuint8mf2_t qh_x = __riscv_vslideup_vx_u8mf2(__riscv_vlmul_ext_v_u8mf4_u8mf2(qh_x1), qh_x2, vl/2, vl);
vuint8mf2_t qh_0 = __riscv_vand_vx_u8mf2(__riscv_vsll_vx_u8mf2(qh_x, 0x2, vl), 0x4, vl);
vuint8mf2_t qh_1 = __riscv_vand_vx_u8mf2(qh_x, 0x4, vl);
vuint8mf2_t qh_2 = __riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(qh_x, 0x2, vl), 0x4, vl);
vuint8mf2_t qh_3 = __riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(qh_x, 0x4, vl), 0x4, vl);
// load Q3
vuint8mf2_t q3_x = __riscv_vle8_v_u8mf2(q3, vl);
vuint8mf2_t q3h_0 = __riscv_vor_vv_u8mf2(__riscv_vand_vx_u8mf2(q3_x, 0x3, vl), qh_0, vl);
vuint8mf2_t q3h_1 = __riscv_vor_vv_u8mf2(__riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(q3_x, 2, vl), 0x3, vl), qh_1, vl);
vuint8mf2_t q3h_2 = __riscv_vor_vv_u8mf2(__riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(q3_x, 4, vl), 0x3, vl), qh_2, vl);
vuint8mf2_t q3h_3 = __riscv_vor_vv_u8mf2(__riscv_vsrl_vx_u8mf2(q3_x, 0x6, vl), qh_3, vl);
vint8mf2_t q3_0 = __riscv_vreinterpret_v_u8mf2_i8mf2(q3h_0);
vint8mf2_t q3_1 = __riscv_vreinterpret_v_u8mf2_i8mf2(q3h_1);
vint8mf2_t q3_2 = __riscv_vreinterpret_v_u8mf2_i8mf2(q3h_2);
vint8mf2_t q3_3 = __riscv_vreinterpret_v_u8mf2_i8mf2(q3h_3);
// load Q8 and take product with Q3
vint16m1_t p0 = __riscv_vwmul_vv_i16m1(q3_0, __riscv_vle8_v_i8mf2(q8, vl), vl);
vint16m1_t p1 = __riscv_vwmul_vv_i16m1(q3_1, __riscv_vle8_v_i8mf2(q8+16, vl), vl);
vint16m1_t p2 = __riscv_vwmul_vv_i16m1(q3_2, __riscv_vle8_v_i8mf2(q8+32, vl), vl);
vint16m1_t p3 = __riscv_vwmul_vv_i16m1(q3_3, __riscv_vle8_v_i8mf2(q8+48, vl), vl);
vint32m1_t vs_0 = __riscv_vwredsum_vs_i16m1_i32m1(p0, vzero, vl);
vint32m1_t vs_1 = __riscv_vwredsum_vs_i16m1_i32m1(p1, vzero, vl);
vint32m1_t vs_2 = __riscv_vwredsum_vs_i16m1_i32m1(p2, vzero, vl);
vint32m1_t vs_3 = __riscv_vwredsum_vs_i16m1_i32m1(p3, vzero, vl);
isum += __riscv_vmv_x_s_i32m1_i32(vs_0) * scales[0];
isum += __riscv_vmv_x_s_i32m1_i32(vs_1) * scales[2];
isum += __riscv_vmv_x_s_i32m1_i32(vs_2) * scales[1];
isum += __riscv_vmv_x_s_i32m1_i32(vs_3) * scales[3];
sumf += d * isum;
}
*s = sumf;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
#else
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
int32_t scales[4];
memset(sums, 0, 8*sizeof(float));
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * restrict q3 = x[i].qs;
const uint8_t * restrict hm = x[i].hmask;
const int8_t * restrict q8 = y[i].qs;
int8_t * restrict a = aux8;
for (int l = 0; l < 8; ++l) {
a[l+ 0] = (int8_t)((q3[l+0] >> 0) & 3) - (hm[l] & 0x01 ? 0 : 4);
a[l+ 8] = (int8_t)((q3[l+8] >> 0) & 3) - (hm[l] & 0x02 ? 0 : 4);
a[l+16] = (int8_t)((q3[l+0] >> 2) & 3) - (hm[l] & 0x04 ? 0 : 4);
a[l+24] = (int8_t)((q3[l+8] >> 2) & 3) - (hm[l] & 0x08 ? 0 : 4);
a[l+32] = (int8_t)((q3[l+0] >> 4) & 3) - (hm[l] & 0x10 ? 0 : 4);
a[l+40] = (int8_t)((q3[l+8] >> 4) & 3) - (hm[l] & 0x20 ? 0 : 4);
a[l+48] = (int8_t)((q3[l+0] >> 6) & 3) - (hm[l] & 0x40 ? 0 : 4);
a[l+56] = (int8_t)((q3[l+8] >> 6) & 3) - (hm[l] & 0x80 ? 0 : 4);
}
scales[0] = (x[i].scales[0] & 0xF) - 8;
scales[1] = (x[i].scales[0] >> 4) - 8;
scales[2] = (x[i].scales[1] & 0xF) - 8;
scales[3] = (x[i].scales[1] >> 4) - 8;
memset(aux32, 0, 8*sizeof(int32_t));
for (int j = 0; j < QK_K/16; ++j) {
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] += q8[l] * a[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux32[l] += scales[j] * aux16[l];
}
const float d = ggml_fp16_to_fp32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
#endif
}
#endif
#if QK_K == 256
void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
assert(n % QK_K == 0);
const block_q4_K * restrict x = vx;
const block_q8_K * restrict y = vy;
const int nb = n / QK_K;
static const uint32_t kmask1 = 0x3f3f3f3f;
static const uint32_t kmask2 = 0x0f0f0f0f;
static const uint32_t kmask3 = 0x03030303;
uint32_t utmp[4];
#ifdef __ARM_NEON
const uint8x16_t m4b = vdupq_n_u8(0xf);
#ifdef __ARM_FEATURE_DOTPROD
const int32x4_t mzero = vdupq_n_s32(0);
#endif
int8x16x2_t q4bytes;
int8x16x2_t q8bytes;
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
const float dmin = y[i].d * ggml_fp16_to_fp32(x[i].dmin);
const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));
memcpy(utmp, x[i].scales, 12);
arm64 support for windows (#3007) Co-authored-by: Cebtenzzre <cebtenzzre@gmail.com>
2023-09-13 01:54:20 +00:00
uint32x2_t mins8 = { 0 };
mins8 = vset_lane_u32(utmp[1] & kmask1, mins8, 0);
mins8 = vset_lane_u32(((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4), mins8, 1);
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
utmp[0] &= kmask1;
const int16x8_t mins = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(mins8)));
const int32x4_t prod = vaddq_s32(vmull_s16(vget_low_s16 (q8sums), vget_low_s16 (mins)),
vmull_s16(vget_high_s16(q8sums), vget_high_s16(mins)));
sumf -= dmin * vaddvq_s32(prod);
const uint8_t * scales = (const uint8_t *)utmp;
const uint8_t * restrict q4 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
int32_t sumi1 = 0;
int32_t sumi2 = 0;
for (int j = 0; j < QK_K/64; ++j) {
const uint8x16x2_t q4bits = vld1q_u8_x2(q4); q4 += 32;
#ifdef __ARM_FEATURE_DOTPROD
q8bytes = vld1q_s8_x2(q8); q8 += 32;
q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[0], m4b));
q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[1], m4b));
const int32x4_t p1 = vdotq_s32(vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]);
sumi1 += vaddvq_s32(p1) * scales[2*j+0];
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
q8bytes = vld1q_s8_x2(q8); q8 += 32;
q4bytes.val[0] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[0], 4));
q4bytes.val[1] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[1], 4));
const int32x4_t p2 = vdotq_s32(vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]);
sumi2 += vaddvq_s32(p2) * scales[2*j+1];
#else
q8bytes = vld1q_s8_x2(q8); q8 += 32;
q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[0], m4b));
q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[1], m4b));
const int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[0]), vget_low_s8 (q8bytes.val[0])),
vmull_s8(vget_high_s8(q4bytes.val[0]), vget_high_s8(q8bytes.val[0])));
const int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[1]), vget_low_s8 (q8bytes.val[1])),
vmull_s8(vget_high_s8(q4bytes.val[1]), vget_high_s8(q8bytes.val[1])));
sumi1 += vaddvq_s16(vaddq_s16(p0, p1)) * scales[2*j+0];
q8bytes = vld1q_s8_x2(q8); q8 += 32;
q4bytes.val[0] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[0], 4));
q4bytes.val[1] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[1], 4));
const int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[0]), vget_low_s8 (q8bytes.val[0])),
vmull_s8(vget_high_s8(q4bytes.val[0]), vget_high_s8(q8bytes.val[0])));
const int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[1]), vget_low_s8 (q8bytes.val[1])),
vmull_s8(vget_high_s8(q4bytes.val[1]), vget_high_s8(q8bytes.val[1])));
sumi2 += vaddvq_s16(vaddq_s16(p2, p3)) * scales[2*j+1];
#endif
}
sumf += d * (sumi1 + sumi2);
}
*s = sumf;
#elif defined __AVX2__
const __m256i m4 = _mm256_set1_epi8(0xF);
__m256 acc = _mm256_setzero_ps();
__m128 acc_m = _mm_setzero_ps();
for (int i = 0; i < nb; ++i) {
const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin);
memcpy(utmp, x[i].scales, 12);
utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
const uint32_t uaux = utmp[1] & kmask1;
utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
utmp[2] = uaux;
utmp[0] &= kmask1;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
const uint8_t * restrict q4 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
const __m256i mins_and_scales = _mm256_cvtepu8_epi16(_mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]));
const __m256i q8sums = _mm256_loadu_si256((const __m256i*)y[i].bsums);
const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1));
const __m128i prod = _mm_madd_epi16(_mm256_extracti128_si256(mins_and_scales, 1), q8s);
acc_m = _mm_fmadd_ps(_mm_set1_ps(dmin), _mm_cvtepi32_ps(prod), acc_m);
const __m128i sc128 = _mm256_extracti128_si256(mins_and_scales, 0);
ggml : workaround for missing _mm256_setr_m128i in GCC < 8 in k_quants.c (#2405)
2023-07-28 18:17:45 +00:00
const __m256i scales = MM256_SET_M128I(sc128, sc128);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
__m256i sumi = _mm256_setzero_si256();
for (int j = 0; j < QK_K/64; ++j) {
const __m256i scale_l = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+0));
const __m256i scale_h = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+1));
const __m256i q4bits = _mm256_loadu_si256((const __m256i*)q4); q4 += 32;
const __m256i q4l = _mm256_and_si256(q4bits, m4);
const __m256i q4h = _mm256_and_si256(_mm256_srli_epi16(q4bits, 4), m4);
const __m256i q8l = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
__m256i p16l = _mm256_maddubs_epi16(q4l, q8l);
p16l = _mm256_madd_epi16(scale_l, p16l);
const __m256i q8h = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
__m256i p16h = _mm256_maddubs_epi16(q4h, q8h);
p16h = _mm256_madd_epi16(scale_h, p16h);
ggml : tiny ggml_vec_dot_q4_K_q8_K AVX2 improvement (#2819)
2023-08-28 12:51:08 +00:00
const __m256i sumj = _mm256_add_epi32(p16l, p16h);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
ggml : tiny ggml_vec_dot_q4_K_q8_K AVX2 improvement (#2819)
2023-08-28 12:51:08 +00:00
sumi = _mm256_add_epi32(sumi, sumj);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
}
__m256 vd = _mm256_set1_ps(d);
acc = _mm256_fmadd_ps(vd, _mm256_cvtepi32_ps(sumi), acc);
}
acc_m = _mm_add_ps(acc_m, _mm_movehl_ps(acc_m, acc_m));
acc_m = _mm_add_ss(acc_m, _mm_movehdup_ps(acc_m));
*s = hsum_float_8(acc) + _mm_cvtss_f32(acc_m);
k-quants : add AVX support to dot functions (#1916) * k_quants : add AVX support * k_quants : apply review comments
2023-06-26 16:46:07 +00:00
#elif defined __AVX__
const __m128i m4 = _mm_set1_epi8(0xF);
const __m128i m2 = _mm_set1_epi8(0x2);
__m256 acc = _mm256_setzero_ps();
__m128 acc_m = _mm_setzero_ps();
for (int i = 0; i < nb; ++i) {
const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin);
const uint8_t * restrict q4 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
memcpy(utmp, x[i].scales, 12);
utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
const uint32_t uaux = utmp[1] & kmask1;
utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
utmp[2] = uaux;
utmp[0] &= kmask1;
const __m128i utmps = _mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]);
const __m128i scales = _mm_cvtepu8_epi16(utmps);
const __m128i mins = _mm_cvtepu8_epi16(_mm_unpackhi_epi64(utmps, utmps));
const __m128i q8sums_0 = _mm_loadu_si128((const __m128i*)&y[i].bsums[0]);
const __m128i q8sums_1 = _mm_loadu_si128((const __m128i*)&y[i].bsums[8]);
const __m128i q8s = _mm_hadd_epi16(q8sums_0, q8sums_1);
const __m128i prod = _mm_madd_epi16(mins, q8s);
acc_m = _mm_add_ps(_mm_mul_ps(_mm_set1_ps(dmin), _mm_cvtepi32_ps(prod)), acc_m);
__m128i sumi_0 = _mm_setzero_si128();
__m128i sumi_1 = _mm_setzero_si128();
__m128i shuffle = _mm_set1_epi16(0x0100);
for (int j = 0; j < QK_K/64; ++j) {
const __m128i scale_l = _mm_shuffle_epi8(scales, shuffle);
shuffle = _mm_add_epi16(shuffle, m2);
const __m128i scale_h = _mm_shuffle_epi8(scales, shuffle);
shuffle = _mm_add_epi16(shuffle, m2);
__m128i q4bits = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
const __m128i q4l_0 = _mm_and_si128(q4bits, m4);
const __m128i q4h_0 = _mm_and_si128(_mm_srli_epi16(q4bits, 4), m4);
q4bits = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
const __m128i q4l_1 = _mm_and_si128(q4bits, m4);
const __m128i q4h_1 = _mm_and_si128(_mm_srli_epi16(q4bits, 4), m4);
const __m128i q8l_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
__m128i p16l = _mm_maddubs_epi16(q4l_0, q8l_0);
p16l = _mm_madd_epi16(scale_l, p16l);
sumi_0 = _mm_add_epi32(sumi_0, p16l);
const __m128i q8l_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
p16l = _mm_maddubs_epi16(q4l_1, q8l_1);
p16l = _mm_madd_epi16(scale_l, p16l);
sumi_1 = _mm_add_epi32(sumi_1, p16l);
const __m128i q8h_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
__m128i p16h = _mm_maddubs_epi16(q4h_0, q8h_0);
p16h = _mm_madd_epi16(scale_h, p16h);
sumi_0 = _mm_add_epi32(sumi_0, p16h);
const __m128i q8h_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
p16h = _mm_maddubs_epi16(q4h_1, q8h_1);
p16h = _mm_madd_epi16(scale_h, p16h);
sumi_1 = _mm_add_epi32(sumi_1, p16h);
}
__m256 vd = _mm256_set1_ps(d);
ggml : workaround for missing _mm256_setr_m128i in GCC < 8 in k_quants.c (#2405)
2023-07-28 18:17:45 +00:00
__m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
k-quants : add AVX support to dot functions (#1916) * k_quants : add AVX support * k_quants : apply review comments
2023-06-26 16:46:07 +00:00
acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(sumi)), acc);
}
acc_m = _mm_add_ps(acc_m, _mm_movehl_ps(acc_m, acc_m));
acc_m = _mm_add_ss(acc_m, _mm_movehdup_ps(acc_m));
*s = hsum_float_8(acc) + _mm_cvtss_f32(acc_m);
ggml : add RISC-V Vector Support for K-Quants and improved the existing intrinsics (#3453) * Added RVV intrinsics support for Q8 quantize row and also improved the existing dot product function for risc-v. The RVV intrinsics is added for the following quantize row functions quantize_row_q8_0 quantize_row_q8_1 The following dot product functions have also been optimized by using LMUL = 1/2 instead of LMUL = 1 ggml_vec_dot_q4_0_q8_0 ggml_vec_dot_q4_1_q8_1 ggml_vec_dot_q5_0_q8_0 ggml_vec_dot_q5_1_q8_1 And vector initialization in Q5 by temporary array is also replaced by the vid intrinsics Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai> * Added RVV intrinsics support for k_quants This adds RISC-V Vector intrinsics support for the following K_quants functions for both QKK = 256 and QKK = 64 ggml_vec_dot_q2_K_q8_K ggml_vec_dot_q3_K_q8_K ggml_vec_dot_q4_K_q8_K ggml_vec_dot_q5_K_q8_K ggml_vec_dot_q6_K_q8_K Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai> --------- Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai>
2023-10-03 18:38:19 +00:00
#elif defined __riscv_v_intrinsic
const uint8_t * scales = (const uint8_t*)&utmp[0];
const uint8_t * mins = (const uint8_t*)&utmp[2];
float sumf = 0;
for (int i = 0; i < nb; ++i) {
size_t vl = 8;
const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
const float dmin = y[i].d * ggml_fp16_to_fp32(x[i].dmin);
vint16mf2_t q8sums_0 = __riscv_vlse16_v_i16mf2(y[i].bsums, 4, vl);
vint16mf2_t q8sums_1 = __riscv_vlse16_v_i16mf2(y[i].bsums+1, 4, vl);
vint16mf2_t q8sums = __riscv_vadd_vv_i16mf2(q8sums_0, q8sums_1, vl);
memcpy(utmp, x[i].scales, 12);
utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
const uint32_t uaux = utmp[1] & kmask1;
utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
utmp[2] = uaux;
utmp[0] &= kmask1;
vuint8mf4_t mins8 = __riscv_vle8_v_u8mf4(mins, vl);
vint16mf2_t v_mins = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vzext_vf2_u16mf2(mins8, vl));
vint32m1_t prod = __riscv_vwmul_vv_i32m1(q8sums, v_mins, vl);
vint32m1_t sumi = __riscv_vredsum_vs_i32m1_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl);
sumf -= dmin * __riscv_vmv_x_s_i32m1_i32(sumi);
const uint8_t * restrict q4 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
vl = 32;
int32_t sum_1 = 0;
int32_t sum_2 = 0;
vint16m1_t vzero = __riscv_vmv_v_x_i16m1(0, 1);
for (int j = 0; j < QK_K/64; ++j) {
// load Q4
vuint8m1_t q4_x = __riscv_vle8_v_u8m1(q4, vl);
// load Q8 and multiply it with lower Q4 nibble
vint8m1_t q8_0 = __riscv_vle8_v_i8m1(q8, vl);
vint8m1_t q4_0 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q4_x, 0x0F, vl));
vint16m2_t qv_0 = __riscv_vwmul_vv_i16m2(q4_0, q8_0, vl);
vint16m1_t vs_0 = __riscv_vredsum_vs_i16m2_i16m1(qv_0, vzero, vl);
sum_1 += __riscv_vmv_x_s_i16m1_i16(vs_0) * scales[2*j+0];
// load Q8 and multiply it with upper Q4 nibble
vint8m1_t q8_1 = __riscv_vle8_v_i8m1(q8+32, vl);
vint8m1_t q4_1 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vsrl_vx_u8m1(q4_x, 0x04, vl));
vint16m2_t qv_1 = __riscv_vwmul_vv_i16m2(q4_1, q8_1, vl);
vint16m1_t vs_1 = __riscv_vredsum_vs_i16m2_i16m1(qv_1, vzero, vl);
sum_2 += __riscv_vmv_x_s_i16m1_i16(vs_1) * scales[2*j+1];
q4 += 32; q8 += 64;
}
sumf += d*(sum_1 + sum_2);
}
*s = sumf;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
#else
const uint8_t * scales = (const uint8_t*)&utmp[0];
const uint8_t * mins = (const uint8_t*)&utmp[2];
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * restrict q4 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * restrict a = aux8;
for (int j = 0; j < QK_K/64; ++j) {
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
a += 32;
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4);
a += 32; q4 += 32;
}
memcpy(utmp, x[i].scales, 12);
utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
const uint32_t uaux = utmp[1] & kmask1;
utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
utmp[2] = uaux;
utmp[0] &= kmask1;
int sumi = 0;
for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
a = aux8;
int is = 0;
for (int j = 0; j < QK_K/32; ++j) {
int32_t scale = scales[is++];
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
}
const float d = ggml_fp16_to_fp32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = ggml_fp16_to_fp32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi;
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
#endif
}
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
#else
void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
assert(n % QK_K == 0);
const block_q4_K * restrict x = vx;
const block_q8_K * restrict y = vy;
const int nb = n / QK_K;
#ifdef __ARM_NEON
const uint8x16_t m4b = vdupq_n_u8(0xf);
#ifdef __ARM_FEATURE_DOTPROD
const int32x4_t mzero = vdupq_n_s32(0);
#endif
float sumf = 0;
int8x16x2_t q4bytes;
int8x16x4_t q8bytes;
float sum_mins = 0.f;
uint16_t aux16[2];
const uint8_t * restrict scales = (const uint8_t *)aux16;
for (int i = 0; i < nb; ++i) {
const uint8_t * restrict q4 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
const uint16_t * restrict a = (const uint16_t *)x[i].scales;
aux16[0] = a[0] & 0x0f0f;
aux16[1] = (a[0] >> 4) & 0x0f0f;
const int32_t summi = scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3]);
sum_mins += y[i].d * (float)x[i].d[1] * summi;
const float d = y[i].d * (float)x[i].d[0];
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
const uint8x16x2_t q4bits = vld1q_u8_x2(q4);
#ifdef __ARM_FEATURE_DOTPROD
q8bytes = vld1q_s8_x4(q8);
q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[0], m4b));
q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[1], m4b));
const int32x4_t p1 = vdotq_s32(vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]);
const int32_t sumi1 = vaddvq_s32(p1) * scales[0];
q4bytes.val[0] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[0], 4));
q4bytes.val[1] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[1], 4));
const int32x4_t p2 = vdotq_s32(vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[2]), q4bytes.val[1], q8bytes.val[3]);
const int32_t sumi2 = vaddvq_s32(p2) * scales[1];
#else
q8bytes = vld1q_s8_x4(q8);
q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[0], m4b));
q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[1], m4b));
const int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[0]), vget_low_s8 (q8bytes.val[0])),
vmull_s8(vget_high_s8(q4bytes.val[0]), vget_high_s8(q8bytes.val[0])));
const int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[1]), vget_low_s8 (q8bytes.val[1])),
vmull_s8(vget_high_s8(q4bytes.val[1]), vget_high_s8(q8bytes.val[1])));
int32_t sumi1 = vaddvq_s16(vaddq_s16(p0, p1)) * scales[0];
q4bytes.val[0] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[0], 4));
q4bytes.val[1] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[1], 4));
const int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[0]), vget_low_s8 (q8bytes.val[2])),
vmull_s8(vget_high_s8(q4bytes.val[0]), vget_high_s8(q8bytes.val[2])));
const int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[1]), vget_low_s8 (q8bytes.val[3])),
vmull_s8(vget_high_s8(q4bytes.val[1]), vget_high_s8(q8bytes.val[3])));
int32_t sumi2 = vaddvq_s16(vaddq_s16(p2, p3)) * scales[1];
#endif
sumf += d * (sumi1 + sumi2);
}
*s = sumf - sum_mins;
#elif defined __AVX2__
const __m256i m4 = _mm256_set1_epi8(0xF);
__m256 acc = _mm256_setzero_ps();
float summs = 0;
uint16_t aux16[2];
const uint8_t * scales = (const uint8_t *)aux16;
for (int i = 0; i < nb; ++i) {
const float d = ggml_fp16_to_fp32(x[i].d[0]) * y[i].d;
const float m = ggml_fp16_to_fp32(x[i].d[1]) * y[i].d;
const __m256 vd = _mm256_set1_ps(d);
const uint16_t * a = (const uint16_t *)x[i].scales;
aux16[0] = a[0] & 0x0f0f;
aux16[1] = (a[0] >> 4) & 0x0f0f;
summs += m * (scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3]));
const uint8_t * restrict q4 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
const __m256i q4bits = _mm256_loadu_si256((const __m256i*)q4);
const __m256i q4l = _mm256_and_si256(q4bits, m4);
const __m256i q4h = _mm256_and_si256(_mm256_srli_epi16(q4bits, 4), m4);
const __m256i q8l = _mm256_loadu_si256((const __m256i*)(q8+ 0));
const __m256i q8h = _mm256_loadu_si256((const __m256i*)(q8+32));
const __m256i p16l = _mm256_maddubs_epi16(q4l, q8l);
const __m256i p16h = _mm256_maddubs_epi16(q4h, q8h);
const __m256i p32l = _mm256_madd_epi16(_mm256_set1_epi16(scales[0]), p16l);
acc = _mm256_fmadd_ps(vd, _mm256_cvtepi32_ps(p32l), acc);
const __m256i p32h = _mm256_madd_epi16(_mm256_set1_epi16(scales[1]), p16h);
acc = _mm256_fmadd_ps(vd, _mm256_cvtepi32_ps(p32h), acc);
}
*s = hsum_float_8(acc) - summs;
k_quants : add AVX support to dot functions with QK_K as 64 (#2339) * add AVX to ggml_vec_dot_q2_K_q8_K() * add AVX to ggml_vec_dot_q3_K_q8_K() * add AVX to ggml_vec_dot_q4_K_q8_K() * add AVX to ggml_vec_dot_q5_K_q8_K() * add AVX to ggml_vec_dot_q6_K_q8_K() * refactor AVX code in ggml_vec_dot_q6_K_q8_K()
2023-07-25 12:13:41 +00:00
#elif defined __AVX__
const __m128i m4 = _mm_set1_epi8(0xF);
__m256 acc = _mm256_setzero_ps();
float summs = 0;
uint16_t aux16[2];
const uint8_t * scales = (const uint8_t *)aux16;
for (int i = 0; i < nb; ++i) {
const float d = ggml_fp16_to_fp32(x[i].d[0]) * y[i].d;
const float m = ggml_fp16_to_fp32(x[i].d[1]) * y[i].d;
const __m256 vd = _mm256_set1_ps(d);
const uint16_t * a = (const uint16_t *)x[i].scales;
aux16[0] = a[0] & 0x0f0f;
aux16[1] = (a[0] >> 4) & 0x0f0f;
summs += m * (scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3]));
const uint8_t * restrict q4 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
const __m256i q4bits = _mm256_loadu_si256((const __m256i*)q4);
const __m128i q4bits_0 = _mm256_extractf128_si256(q4bits, 0);
const __m128i q4bits_1 = _mm256_extractf128_si256(q4bits, 1);
const __m128i q4_0 = _mm_and_si128(q4bits_0, m4);
const __m128i q4_1 = _mm_and_si128(q4bits_1, m4);
const __m128i q4_2 = _mm_and_si128(_mm_srli_epi16(q4bits_0, 4), m4);
const __m128i q4_3 = _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4);
const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
const __m128i p16_0 = _mm_maddubs_epi16(q4_0, _mm256_extractf128_si256(q8_0, 0));
const __m128i p16_1 = _mm_maddubs_epi16(q4_1, _mm256_extractf128_si256(q8_0, 1));
const __m128i p16_2 = _mm_maddubs_epi16(q4_2, _mm256_extractf128_si256(q8_1, 0));
const __m128i p16_3 = _mm_maddubs_epi16(q4_3, _mm256_extractf128_si256(q8_1, 1));
const __m128i p32_0 = _mm_madd_epi16(_mm_set1_epi16(scales[0]), p16_0);
const __m128i p32_1 = _mm_madd_epi16(_mm_set1_epi16(scales[0]), p16_1);
ggml : workaround for missing _mm256_setr_m128i in GCC < 8 in k_quants.c (#2405)
2023-07-28 18:17:45 +00:00
acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(MM256_SET_M128I(p32_1, p32_0))), acc);
k_quants : add AVX support to dot functions with QK_K as 64 (#2339) * add AVX to ggml_vec_dot_q2_K_q8_K() * add AVX to ggml_vec_dot_q3_K_q8_K() * add AVX to ggml_vec_dot_q4_K_q8_K() * add AVX to ggml_vec_dot_q5_K_q8_K() * add AVX to ggml_vec_dot_q6_K_q8_K() * refactor AVX code in ggml_vec_dot_q6_K_q8_K()
2023-07-25 12:13:41 +00:00
const __m128i p32_2 = _mm_madd_epi16(_mm_set1_epi16(scales[1]), p16_2);
const __m128i p32_3 = _mm_madd_epi16(_mm_set1_epi16(scales[1]), p16_3);
ggml : workaround for missing _mm256_setr_m128i in GCC < 8 in k_quants.c (#2405)
2023-07-28 18:17:45 +00:00
acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(MM256_SET_M128I(p32_3, p32_2))), acc);
k_quants : add AVX support to dot functions with QK_K as 64 (#2339) * add AVX to ggml_vec_dot_q2_K_q8_K() * add AVX to ggml_vec_dot_q3_K_q8_K() * add AVX to ggml_vec_dot_q4_K_q8_K() * add AVX to ggml_vec_dot_q5_K_q8_K() * add AVX to ggml_vec_dot_q6_K_q8_K() * refactor AVX code in ggml_vec_dot_q6_K_q8_K()
2023-07-25 12:13:41 +00:00
}
*s = hsum_float_8(acc) - summs;
ggml : add RISC-V Vector Support for K-Quants and improved the existing intrinsics (#3453) * Added RVV intrinsics support for Q8 quantize row and also improved the existing dot product function for risc-v. The RVV intrinsics is added for the following quantize row functions quantize_row_q8_0 quantize_row_q8_1 The following dot product functions have also been optimized by using LMUL = 1/2 instead of LMUL = 1 ggml_vec_dot_q4_0_q8_0 ggml_vec_dot_q4_1_q8_1 ggml_vec_dot_q5_0_q8_0 ggml_vec_dot_q5_1_q8_1 And vector initialization in Q5 by temporary array is also replaced by the vid intrinsics Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai> * Added RVV intrinsics support for k_quants This adds RISC-V Vector intrinsics support for the following K_quants functions for both QKK = 256 and QKK = 64 ggml_vec_dot_q2_K_q8_K ggml_vec_dot_q3_K_q8_K ggml_vec_dot_q4_K_q8_K ggml_vec_dot_q5_K_q8_K ggml_vec_dot_q6_K_q8_K Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai> --------- Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai>
2023-10-03 18:38:19 +00:00
#elif defined __riscv_v_intrinsic
uint16_t s16[2];
const uint8_t * restrict scales = (const uint8_t *)s16;
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * restrict q4 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
const uint16_t * restrict b = (const uint16_t *)x[i].scales;
s16[0] = b[0] & 0x0f0f;
s16[1] = (b[0] >> 4) & 0x0f0f;
sumf -= y[i].d * ggml_fp16_to_fp32(x[i].d[1]) * (scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3]));
const float d = y[i].d * ggml_fp16_to_fp32(x[i].d[0]);
size_t vl = 32;
vint16m1_t vzero = __riscv_vmv_v_x_i16m1(0, 1);
// load Q4
vuint8m1_t q4_x = __riscv_vle8_v_u8m1(q4, vl);
// load Q8 and multiply it with lower Q4 nibble
vint8m1_t q4_a = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q4_x, 0x0F, vl));
vint16m2_t va_0 = __riscv_vwmul_vv_i16m2(q4_a, __riscv_vle8_v_i8m1(q8, vl), vl);
vint16m1_t aux1 = __riscv_vredsum_vs_i16m2_i16m1(va_0, vzero, vl);
sumf += d*scales[0]*__riscv_vmv_x_s_i16m1_i16(aux1);
// load Q8 and multiply it with upper Q4 nibble
vint8m1_t q4_s = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vsrl_vx_u8m1(q4_x, 0x04, vl));
vint16m2_t va_1 = __riscv_vwmul_vv_i16m2(q4_s, __riscv_vle8_v_i8m1(q8+32, vl), vl);
vint16m1_t aux2 = __riscv_vredsum_vs_i16m2_i16m1(va_1, vzero, vl);
sumf += d*scales[1]*__riscv_vmv_x_s_i16m1_i16(aux2);
}
*s = sumf;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
#else
uint8_t aux8[QK_K];
int16_t aux16[16];
float sums [8];
memset(sums, 0, 8*sizeof(float));
uint16_t s16[2];
const uint8_t * restrict scales = (const uint8_t *)s16;
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * restrict q4 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
uint8_t * restrict a = aux8;
for (int l = 0; l < 32; ++l) a[l+ 0] = q4[l] & 0xF;
for (int l = 0; l < 32; ++l) a[l+32] = q4[l] >> 4;
const uint16_t * restrict b = (const uint16_t *)x[i].scales;
s16[0] = b[0] & 0x0f0f;
s16[1] = (b[0] >> 4) & 0x0f0f;
sumf -= y[i].d * ggml_fp16_to_fp32(x[i].d[1]) * (scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3]));
const float d = y[i].d * ggml_fp16_to_fp32(x[i].d[0]);
for (int j = 0; j < QK_K/32; ++j) {
for (int l = 0; l < 16; ++l) aux16[l] = q8[l] * a[l];
q8 += 16; a += 16;
for (int l = 0; l < 16; ++l) aux16[l] += q8[l] * a[l];
q8 += 16; a += 16;
const float dl = d * scales[j];
for (int l = 0; l < 8; ++l) sums[l] += dl * (aux16[l] + aux16[l+8]);
}
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
#endif
}
#endif
#if QK_K == 256
k-quants : allow to optionally disable at compile time (#1734) * k-quants : put behind optional compile flag LLAMA_K_QUANTS * build : enable k-quants by default
2023-06-07 07:59:52 +00:00
void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
assert(n % QK_K == 0);
k-quants : allow to optionally disable at compile time (#1734) * k-quants : put behind optional compile flag LLAMA_K_QUANTS * build : enable k-quants by default
2023-06-07 07:59:52 +00:00
const block_q5_K * restrict x = vx;
const block_q8_K * restrict y = vy;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
const int nb = n / QK_K;
static const uint32_t kmask1 = 0x3f3f3f3f;
static const uint32_t kmask2 = 0x0f0f0f0f;
static const uint32_t kmask3 = 0x03030303;
uint32_t utmp[4];
#ifdef __ARM_NEON
const uint8x16_t m4b = vdupq_n_u8(0xf);
const uint8x16_t mone = vdupq_n_u8(1);
const uint8x16_t mtwo = vdupq_n_u8(2);
k-quants : fix build on armv7 (android only) (#2920) * k-quants : fix build on armv7 * ggml : cleanup unused arm32 specific impl * k-quants : avoid some unused vzero / mzero define * ggml-alloc : use 4g for MEASURE_MAX_SIZE in 32-bit arm
2023-09-02 12:23:45 +00:00
#if defined(__ARM_FEATURE_DOTPROD)
const int32x4_t mzero = vdupq_n_s32(0);
#endif
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
int8x16x4_t q5bytes;
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
const float dmin = y[i].d * ggml_fp16_to_fp32(x[i].dmin);
const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));
memcpy(utmp, x[i].scales, 12);
utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
const uint32_t uaux = utmp[1] & kmask1;
utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
utmp[2] = uaux;
utmp[0] &= kmask1;
const uint8x8_t mins8 = vld1_u8((const uint8_t*)utmp + 8);
const int16x8_t mins = vreinterpretq_s16_u16(vmovl_u8(mins8));
const int32x4_t prod = vaddq_s32(vmull_s16(vget_low_s16 (q8sums), vget_low_s16 (mins)),
vmull_s16(vget_high_s16(q8sums), vget_high_s16(mins)));
int32_t sumi_mins = vaddvq_s32(prod);
const uint8_t * scales = (const uint8_t *)utmp;
const uint8_t * restrict q5 = x[i].qs;
const uint8_t * restrict qh = x[i].qh;
const int8_t * restrict q8 = y[i].qs;
uint8x16x2_t qhbits = vld1q_u8_x2(qh);
uint8x16x4_t q5h;
int32_t sumi = 0;
for (int j = 0; j < QK_K/64; ++j) {
const uint8x16x2_t q5bits = vld1q_u8_x2(q5); q5 += 32;
Revert "ggml : load data into int8x16x4_t using vld4q_s8 on arm64 (#1738)" This reverts commit 8432d4d9f716b25133e3ed671d91e21f6f3be867.
2023-06-08 17:48:14 +00:00
const int8x16x4_t q8bytes = vld1q_s8_x4(q8); q8 += 64;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
q5h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits.val[0]), 4);
q5h.val[1] = vshlq_n_u8(vandq_u8(mone, qhbits.val[1]), 4);
q5h.val[2] = vshlq_n_u8(vandq_u8(mtwo, qhbits.val[0]), 3);
q5h.val[3] = vshlq_n_u8(vandq_u8(mtwo, qhbits.val[1]), 3);
qhbits.val[0] = vshrq_n_u8(qhbits.val[0], 2);
qhbits.val[1] = vshrq_n_u8(qhbits.val[1], 2);
q5bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q5bits.val[0], m4b), q5h.val[0]));
q5bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q5bits.val[1], m4b), q5h.val[1]));
q5bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q5bits.val[0], 4), q5h.val[2]));
q5bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q5bits.val[1], 4), q5h.val[3]));
#if defined(__ARM_FEATURE_DOTPROD)
sumi += vaddvq_s32(vdotq_s32(vdotq_s32(mzero, q5bytes.val[0], q8bytes.val[0]), q5bytes.val[1], q8bytes.val[1])) * *scales++;
sumi += vaddvq_s32(vdotq_s32(vdotq_s32(mzero, q5bytes.val[2], q8bytes.val[2]), q5bytes.val[3], q8bytes.val[3])) * *scales++;
#else
const int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q5bytes.val[0]), vget_low_s8 (q8bytes.val[0])),
vmull_s8(vget_high_s8(q5bytes.val[0]), vget_high_s8(q8bytes.val[0])));
const int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q5bytes.val[1]), vget_low_s8 (q8bytes.val[1])),
vmull_s8(vget_high_s8(q5bytes.val[1]), vget_high_s8(q8bytes.val[1])));
sumi += vaddvq_s16(vaddq_s16(p0, p1)) * *scales++;
const int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q5bytes.val[2]), vget_low_s8 (q8bytes.val[2])),
vmull_s8(vget_high_s8(q5bytes.val[2]), vget_high_s8(q8bytes.val[2])));
const int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q5bytes.val[3]), vget_low_s8 (q8bytes.val[3])),
vmull_s8(vget_high_s8(q5bytes.val[3]), vget_high_s8(q8bytes.val[3])));
sumi += vaddvq_s16(vaddq_s16(p2, p3)) * *scales++;
#endif
}
sumf += d * sumi - dmin * sumi_mins;
}
*s = sumf;
#elif defined __AVX2__
const __m256i m4 = _mm256_set1_epi8(0xF);
const __m128i mzero = _mm_setzero_si128();
const __m256i mone = _mm256_set1_epi8(1);
__m256 acc = _mm256_setzero_ps();
float summs = 0.f;
for (int i = 0; i < nb; ++i) {
const uint8_t * restrict q5 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
#if QK_K == 256
const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
memcpy(utmp, x[i].scales, 12);
utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
const uint32_t uaux = utmp[1] & kmask1;
utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
utmp[2] = uaux;
utmp[0] &= kmask1;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
#else
// TODO
const float d = 0, dmin = 0;
#endif
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
const __m256i mins_and_scales = _mm256_cvtepu8_epi16(_mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]));
const __m256i q8sums = _mm256_loadu_si256((const __m256i*)y[i].bsums);
const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1));
const __m128i prod = _mm_madd_epi16(_mm256_extracti128_si256(mins_and_scales, 1), q8s);
const __m128i hsum = _mm_hadd_epi32(_mm_hadd_epi32(prod, mzero), mzero);
summs += dmin * _mm_extract_epi32(hsum, 0);
const __m128i sc128 = _mm256_extracti128_si256(mins_and_scales, 0);
ggml : workaround for missing _mm256_setr_m128i in GCC < 8 in k_quants.c (#2405)
2023-07-28 18:17:45 +00:00
const __m256i scales = MM256_SET_M128I(sc128, sc128);
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
const __m256i hbits = _mm256_loadu_si256((const __m256i*)x[i].qh);
__m256i hmask = mone;
__m256i sumi = _mm256_setzero_si256();
int bit = 0;
for (int j = 0; j < QK_K/64; ++j) {
const __m256i scale_0 = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+0));
const __m256i scale_1 = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+1));
const __m256i q5bits = _mm256_loadu_si256((const __m256i*)q5); q5 += 32;
const __m256i q5l_0 = _mm256_and_si256(q5bits, m4);
const __m256i q5h_0 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_and_si256(hbits, hmask), bit++), 4);
const __m256i q5_0 = _mm256_add_epi8(q5l_0, q5h_0);
hmask = _mm256_slli_epi16(hmask, 1);
const __m256i q5l_1 = _mm256_and_si256(_mm256_srli_epi16(q5bits, 4), m4);
const __m256i q5h_1 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_and_si256(hbits, hmask), bit++), 4);
const __m256i q5_1 = _mm256_add_epi8(q5l_1, q5h_1);
hmask = _mm256_slli_epi16(hmask, 1);
const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
__m256i p16_0 = _mm256_maddubs_epi16(q5_0, q8_0);
__m256i p16_1 = _mm256_maddubs_epi16(q5_1, q8_1);
p16_0 = _mm256_madd_epi16(scale_0, p16_0);
p16_1 = _mm256_madd_epi16(scale_1, p16_1);
sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_1));
}
__m256 vd = _mm256_set1_ps(d);
acc = _mm256_fmadd_ps(vd, _mm256_cvtepi32_ps(sumi), acc);
}
*s = hsum_float_8(acc) + summs;
k-quants : add AVX support to dot functions (#1916) * k_quants : add AVX support * k_quants : apply review comments
2023-06-26 16:46:07 +00:00
#elif defined __AVX__
const __m128i m4 = _mm_set1_epi8(0xF);
const __m128i mzero = _mm_setzero_si128();
const __m128i mone = _mm_set1_epi8(1);
const __m128i m2 = _mm_set1_epi8(2);
__m256 acc = _mm256_setzero_ps();
float summs = 0.f;
for (int i = 0; i < nb; ++i) {
const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin);
const uint8_t * restrict q5 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
memcpy(utmp, x[i].scales, 12);
utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
const uint32_t uaux = utmp[1] & kmask1;
utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
utmp[2] = uaux;
utmp[0] &= kmask1;
const __m128i utmps = _mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]);
const __m128i scales = _mm_cvtepu8_epi16(utmps);
const __m128i mins = _mm_cvtepu8_epi16(_mm_unpackhi_epi64(utmps, utmps));
const __m128i q8sums_0 = _mm_loadu_si128((const __m128i*)&y[i].bsums[0]);
const __m128i q8sums_1 = _mm_loadu_si128((const __m128i*)&y[i].bsums[8]);
const __m128i q8s = _mm_hadd_epi16(q8sums_0, q8sums_1);
const __m128i prod = _mm_madd_epi16(mins, q8s);
const __m128i hsum = _mm_hadd_epi32(_mm_hadd_epi32(prod, mzero), mzero);
summs += dmin * _mm_extract_epi32(hsum, 0);
const __m128i hbits_0 = _mm_loadu_si128((const __m128i*)&x[i].qh[0]);
const __m128i hbits_1 = _mm_loadu_si128((const __m128i*)&x[i].qh[16]);
__m128i hmask = mone;
__m128i sumi_0 = _mm_setzero_si128();
__m128i sumi_1 = _mm_setzero_si128();
int bit = 0;
__m128i shuffle = _mm_set1_epi16(0x0100);
for (int j = 0; j < QK_K/64; ++j) {
const __m128i scale_0 = _mm_shuffle_epi8(scales, shuffle);
shuffle = _mm_add_epi16(shuffle, m2);
const __m128i scale_1 = _mm_shuffle_epi8(scales, shuffle);
shuffle = _mm_add_epi16(shuffle, m2);
const __m128i q5bits_0 = _mm_loadu_si128((const __m128i*)q5); q5 += 16;
const __m128i q5bits_1 = _mm_loadu_si128((const __m128i*)q5); q5 += 16;
__m128i q5l_0 = _mm_and_si128(q5bits_0, m4);
__m128i q5l_1 = _mm_and_si128(q5bits_1, m4);
__m128i q5h_0 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_0, hmask), bit), 4);
__m128i q5h_1 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_1, hmask), bit++), 4);
__m128i q5_0 = _mm_add_epi8(q5l_0, q5h_0);
__m128i q5_1 = _mm_add_epi8(q5l_1, q5h_1);
hmask = _mm_slli_epi16(hmask, 1);
__m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
__m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
__m128i p16_0 = _mm_maddubs_epi16(q5_0, q8_0);
__m128i p16_1 = _mm_maddubs_epi16(q5_1, q8_1);
p16_0 = _mm_madd_epi16(scale_0, p16_0);
p16_1 = _mm_madd_epi16(scale_0, p16_1);
q5l_0 = _mm_and_si128(_mm_srli_epi16(q5bits_0, 4), m4);
q5l_1 = _mm_and_si128(_mm_srli_epi16(q5bits_1, 4), m4);
q5h_0 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_0, hmask), bit), 4);
q5h_1 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_1, hmask), bit++), 4);
q5_0 = _mm_add_epi8(q5l_0, q5h_0);
q5_1 = _mm_add_epi8(q5l_1, q5h_1);
hmask = _mm_slli_epi16(hmask, 1);
q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
__m128i p16_2 = _mm_maddubs_epi16(q5_0, q8_0);
__m128i p16_3 = _mm_maddubs_epi16(q5_1, q8_1);
p16_2 = _mm_madd_epi16(scale_1, p16_2);
p16_3 = _mm_madd_epi16(scale_1, p16_3);
sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
}
__m256 vd = _mm256_set1_ps(d);
ggml : workaround for missing _mm256_setr_m128i in GCC < 8 in k_quants.c (#2405)
2023-07-28 18:17:45 +00:00
__m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
k-quants : add AVX support to dot functions (#1916) * k_quants : add AVX support * k_quants : apply review comments
2023-06-26 16:46:07 +00:00
acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(sumi)), acc);
}
*s = hsum_float_8(acc) + summs;
ggml : add RISC-V Vector Support for K-Quants and improved the existing intrinsics (#3453) * Added RVV intrinsics support for Q8 quantize row and also improved the existing dot product function for risc-v. The RVV intrinsics is added for the following quantize row functions quantize_row_q8_0 quantize_row_q8_1 The following dot product functions have also been optimized by using LMUL = 1/2 instead of LMUL = 1 ggml_vec_dot_q4_0_q8_0 ggml_vec_dot_q4_1_q8_1 ggml_vec_dot_q5_0_q8_0 ggml_vec_dot_q5_1_q8_1 And vector initialization in Q5 by temporary array is also replaced by the vid intrinsics Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai> * Added RVV intrinsics support for k_quants This adds RISC-V Vector intrinsics support for the following K_quants functions for both QKK = 256 and QKK = 64 ggml_vec_dot_q2_K_q8_K ggml_vec_dot_q3_K_q8_K ggml_vec_dot_q4_K_q8_K ggml_vec_dot_q5_K_q8_K ggml_vec_dot_q6_K_q8_K Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai> --------- Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai>
2023-10-03 18:38:19 +00:00
#elif defined __riscv_v_intrinsic
const uint8_t * scales = (const uint8_t*)&utmp[0];
const uint8_t * mins = (const uint8_t*)&utmp[2];
float sumf = 0;
float sums = 0.0;
size_t vl;
for (int i = 0; i < nb; ++i) {
vl = 8;
const uint8_t * restrict q5 = x[i].qs;
const uint8_t * restrict hm = x[i].qh;
const int8_t * restrict q8 = y[i].qs;
const float d = ggml_fp16_to_fp32(x[i].d) * y[i].d;
const float dmin = ggml_fp16_to_fp32(x[i].dmin) * y[i].d;
vint16mf2_t q8sums_0 = __riscv_vlse16_v_i16mf2(y[i].bsums, 4, vl);
vint16mf2_t q8sums_1 = __riscv_vlse16_v_i16mf2(y[i].bsums+1, 4, vl);
vint16mf2_t q8sums = __riscv_vadd_vv_i16mf2(q8sums_0, q8sums_1, vl);
memcpy(utmp, x[i].scales, 12);
utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
const uint32_t uaux = utmp[1] & kmask1;
utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
utmp[2] = uaux;
utmp[0] &= kmask1;
vuint8mf4_t mins8 = __riscv_vle8_v_u8mf4(mins, vl);
vint16mf2_t v_mins = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vzext_vf2_u16mf2(mins8, vl));
vint32m1_t prod = __riscv_vwmul_vv_i32m1(q8sums, v_mins, vl);
vint32m1_t sumi = __riscv_vredsum_vs_i32m1_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl);
sumf -= dmin * __riscv_vmv_x_s_i32m1_i32(sumi);
vl = 32;
int32_t aux32 = 0;
int is = 0;
uint8_t m = 1;
vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
vuint8m1_t vqh = __riscv_vle8_v_u8m1(hm, vl);
for (int j = 0; j < QK_K/64; ++j) {
// load Q5 and Q8
vuint8m1_t q5_x = __riscv_vle8_v_u8m1(q5, vl);
vint8m1_t q8_y1 = __riscv_vle8_v_i8m1(q8, vl);
vint8m1_t q8_y2 = __riscv_vle8_v_i8m1(q8+32, vl);
// compute mask for addition
vint8m1_t q5_a = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q5_x, 0x0F, vl));
vuint8m1_t qh_m1 = __riscv_vand_vx_u8m1(vqh, m, vl);
vbool8_t vmask_1 = __riscv_vmsne_vx_u8m1_b8(qh_m1, 0, vl);
vint8m1_t q5_m1 = __riscv_vadd_vx_i8m1_m(vmask_1, q5_a, 16, vl);
m <<= 1;
vint8m1_t q5_l = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vsrl_vx_u8m1(q5_x, 0x04, vl));
vuint8m1_t qh_m2 = __riscv_vand_vx_u8m1(vqh, m, vl);
vbool8_t vmask_2 = __riscv_vmsne_vx_u8m1_b8(qh_m2, 0, vl);
vint8m1_t q5_m2 = __riscv_vadd_vx_i8m1_m(vmask_2, q5_l, 16, vl);
m <<= 1;
vint16m2_t v0 = __riscv_vwmul_vv_i16m2(q5_m1, q8_y1, vl);
vint16m2_t v1 = __riscv_vwmul_vv_i16m2(q5_m2, q8_y2, vl);
vint32m4_t vs1 = __riscv_vwmul_vx_i32m4(v0, scales[is++], vl);
vint32m4_t vs2 = __riscv_vwmul_vx_i32m4(v1, scales[is++], vl);
vint32m1_t vacc1 = __riscv_vredsum_vs_i32m4_i32m1(vs1, vzero, vl);
vint32m1_t vacc2 = __riscv_vredsum_vs_i32m4_i32m1(vs2, vzero, vl);
aux32 += __riscv_vmv_x_s_i32m1_i32(vacc1) + __riscv_vmv_x_s_i32m1_i32(vacc2);
q5 += 32; q8 += 64;
}
vfloat32m1_t vaux = __riscv_vfmul_vf_f32m1(__riscv_vfmv_v_f_f32m1(aux32, 1), d, 1);
sums += __riscv_vfmv_f_s_f32m1_f32(vaux);
}
*s = sumf+sums;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
#else
const uint8_t * scales = (const uint8_t*)&utmp[0];
const uint8_t * mins = (const uint8_t*)&utmp[2];
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * restrict q4 = x[i].qs;
const uint8_t * restrict hm = x[i].qh;
const int8_t * restrict q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * restrict a = aux8;
uint8_t m = 1;
for (int j = 0; j < QK_K/64; ++j) {
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4);
for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
a += 32; m <<= 1;
q4 += 32;
}
memcpy(utmp, x[i].scales, 12);
utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
const uint32_t uaux = utmp[1] & kmask1;
utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
utmp[2] = uaux;
utmp[0] &= kmask1;
int sumi = 0;
for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
a = aux8;
int is = 0;
for (int j = 0; j < QK_K/32; ++j) {
int32_t scale = scales[is++];
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
}
const float d = ggml_fp16_to_fp32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
const float dmin = ggml_fp16_to_fp32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi;
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
#endif
}
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
#else
void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
assert(n % QK_K == 0);
const block_q5_K * restrict x = vx;
const block_q8_K * restrict y = vy;
const int nb = n / QK_K;
#ifdef __ARM_NEON
const uint8x16_t m4b = vdupq_n_u8(0xf);
const uint8x16_t mh = vdupq_n_u8(16);
k-quants : fix build on armv7 (android only) (#2920) * k-quants : fix build on armv7 * ggml : cleanup unused arm32 specific impl * k-quants : avoid some unused vzero / mzero define * ggml-alloc : use 4g for MEASURE_MAX_SIZE in 32-bit arm
2023-09-02 12:23:45 +00:00
#if defined(__ARM_FEATURE_DOTPROD)
const int32x4_t mzero = vdupq_n_s32(0);
#endif
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
int8x16x4_t q5bytes;
uint8x16x4_t q5h;
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const float d = y[i].d * (float)x[i].d;
const int8_t * sc = x[i].scales;
const uint8_t * restrict q5 = x[i].qs;
const uint8_t * restrict qh = x[i].qh;
const int8_t * restrict q8 = y[i].qs;
const uint8x8_t qhbits = vld1_u8(qh);
const uint8x16x2_t q5bits = vld1q_u8_x2(q5);
const int8x16x4_t q8bytes = vld1q_s8_x4(q8);
const uint8x16_t htmp = vcombine_u8(qhbits, vshr_n_u8(qhbits, 1));
q5h.val[0] = vbicq_u8(mh, vshlq_n_u8(htmp, 4));
q5h.val[1] = vbicq_u8(mh, vshlq_n_u8(htmp, 2));
q5h.val[2] = vbicq_u8(mh, htmp);
q5h.val[3] = vbicq_u8(mh, vshrq_n_u8(htmp, 2));
q5bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(q5bits.val[0], m4b)), vreinterpretq_s8_u8(q5h.val[0]));
q5bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(q5bits.val[1], m4b)), vreinterpretq_s8_u8(q5h.val[1]));
q5bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vshrq_n_u8(q5bits.val[0], 4)), vreinterpretq_s8_u8(q5h.val[2]));
q5bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vshrq_n_u8(q5bits.val[1], 4)), vreinterpretq_s8_u8(q5h.val[3]));
#if defined(__ARM_FEATURE_DOTPROD)
int32_t sumi1 = sc[0] * vaddvq_s32(vdotq_s32(mzero, q5bytes.val[0], q8bytes.val[0]));
int32_t sumi2 = sc[1] * vaddvq_s32(vdotq_s32(mzero, q5bytes.val[1], q8bytes.val[1]));
int32_t sumi3 = sc[2] * vaddvq_s32(vdotq_s32(mzero, q5bytes.val[2], q8bytes.val[2]));
int32_t sumi4 = sc[3] * vaddvq_s32(vdotq_s32(mzero, q5bytes.val[3], q8bytes.val[3]));
sumf += d * (sumi1 + sumi2 + sumi3 + sumi4);
#else
const int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q5bytes.val[0]), vget_low_s8 (q8bytes.val[0])),
vmull_s8(vget_high_s8(q5bytes.val[0]), vget_high_s8(q8bytes.val[0])));
const int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q5bytes.val[1]), vget_low_s8 (q8bytes.val[1])),
vmull_s8(vget_high_s8(q5bytes.val[1]), vget_high_s8(q8bytes.val[1])));
int32_t sumi = sc[0] * vaddvq_s16(p0) + sc[1] * vaddvq_s16(p1);
const int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q5bytes.val[2]), vget_low_s8 (q8bytes.val[2])),
vmull_s8(vget_high_s8(q5bytes.val[2]), vget_high_s8(q8bytes.val[2])));
const int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q5bytes.val[3]), vget_low_s8 (q8bytes.val[3])),
vmull_s8(vget_high_s8(q5bytes.val[3]), vget_high_s8(q8bytes.val[3])));
sumi += sc[2] * vaddvq_s16(p2) + sc[3] * vaddvq_s16(p3);
sumf += d*sumi;
#endif
}
*s = sumf;
#elif defined __AVX2__
const __m256i m4 = _mm256_set1_epi8(0xF);
const __m256i mone = _mm256_set1_epi8(1);
__m256 acc = _mm256_setzero_ps();
for (int i = 0; i < nb; ++i) {
const uint8_t * restrict q5 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
const __m256i q5bits = _mm256_loadu_si256((const __m256i*)q5);
ggml : workaround for missing _mm256_setr_m128i in GCC < 8 in k_quants.c (#2405)
2023-07-28 18:17:45 +00:00
const __m256i scale_l = MM256_SET_M128I(_mm_set1_epi16(x[i].scales[1]), _mm_set1_epi16(x[i].scales[0]));
const __m256i scale_h = MM256_SET_M128I(_mm_set1_epi16(x[i].scales[3]), _mm_set1_epi16(x[i].scales[2]));
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
int64_t aux64;
memcpy(&aux64, x[i].qh, 8);
const __m128i haux128 = _mm_set_epi64x(aux64 >> 1, aux64);
ggml : workaround for missing _mm256_setr_m128i in GCC < 8 in k_quants.c (#2405)
2023-07-28 18:17:45 +00:00
const __m256i haux256 = MM256_SET_M128I(_mm_srli_epi16(haux128, 2), haux128);
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
const __m256i q5h_0 = _mm256_slli_epi16(_mm256_andnot_si256(haux256, mone), 4);
const __m256i q5h_1 = _mm256_slli_epi16(_mm256_andnot_si256(_mm256_srli_epi16(haux256, 4), mone), 4);
const __m256i q5l_0 = _mm256_and_si256(q5bits, m4);
const __m256i q5l_1 = _mm256_and_si256(_mm256_srli_epi16(q5bits, 4), m4);
const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
const __m256i p16_0 = _mm256_madd_epi16(scale_l, _mm256_maddubs_epi16(q5l_0, q8_0));
const __m256i p16_1 = _mm256_madd_epi16(scale_h, _mm256_maddubs_epi16(q5l_1, q8_1));
const __m256i s16_0 = _mm256_madd_epi16(scale_l, _mm256_maddubs_epi16(q5h_0, q8_0));
const __m256i s16_1 = _mm256_madd_epi16(scale_h, _mm256_maddubs_epi16(q5h_1, q8_1));
const __m256i dot = _mm256_sub_epi32(_mm256_add_epi32(p16_0, p16_1), _mm256_add_epi32(s16_0, s16_1));
acc = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(dot), acc);
}
*s = hsum_float_8(acc);
k_quants : add AVX support to dot functions with QK_K as 64 (#2339) * add AVX to ggml_vec_dot_q2_K_q8_K() * add AVX to ggml_vec_dot_q3_K_q8_K() * add AVX to ggml_vec_dot_q4_K_q8_K() * add AVX to ggml_vec_dot_q5_K_q8_K() * add AVX to ggml_vec_dot_q6_K_q8_K() * refactor AVX code in ggml_vec_dot_q6_K_q8_K()
2023-07-25 12:13:41 +00:00
#elif defined __AVX__
const __m128i m4 = _mm_set1_epi8(0xF);
const __m128i mone = _mm_set1_epi8(1);
__m256 acc = _mm256_setzero_ps();
for (int i = 0; i < nb; ++i) {
const uint8_t * restrict q5 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
const __m256i q5bits = _mm256_loadu_si256((const __m256i*)q5);
const __m128i scale_0 = _mm_set1_epi16(x[i].scales[0]);
const __m128i scale_1 = _mm_set1_epi16(x[i].scales[1]);
const __m128i scale_2 = _mm_set1_epi16(x[i].scales[2]);
const __m128i scale_3 = _mm_set1_epi16(x[i].scales[3]);
int64_t aux64;
memcpy(&aux64, x[i].qh, 8);
const __m128i haux128_0 = _mm_set_epi64x(aux64 >> 1, aux64);
const __m128i haux128_1 = _mm_srli_epi16(haux128_0, 2);
const __m128i q5h_0 = _mm_slli_epi16(_mm_andnot_si128(haux128_0, mone), 4);
const __m128i q5h_1 = _mm_slli_epi16(_mm_andnot_si128(haux128_1, mone), 4);
const __m128i q5h_2 = _mm_slli_epi16(_mm_andnot_si128(_mm_srli_epi16(haux128_0, 4), mone), 4);
const __m128i q5h_3 = _mm_slli_epi16(_mm_andnot_si128(_mm_srli_epi16(haux128_1, 4), mone), 4);
const __m128i q5l_0 = _mm_and_si128(_mm256_extractf128_si256(q5bits, 0), m4);
const __m128i q5l_1 = _mm_and_si128(_mm256_extractf128_si256(q5bits, 1), m4);
const __m128i q5l_2 = _mm_and_si128(_mm_srli_epi16(_mm256_extractf128_si256(q5bits, 0), 4), m4);
const __m128i q5l_3 = _mm_and_si128(_mm_srli_epi16(_mm256_extractf128_si256(q5bits, 1), 4), m4);
const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
const __m128i p16_0 = _mm_madd_epi16(scale_0, _mm_maddubs_epi16(q5l_0, _mm256_extractf128_si256(q8_0, 0)));
const __m128i p16_1 = _mm_madd_epi16(scale_1, _mm_maddubs_epi16(q5l_1, _mm256_extractf128_si256(q8_0, 1)));
const __m128i p16_2 = _mm_madd_epi16(scale_2, _mm_maddubs_epi16(q5l_2, _mm256_extractf128_si256(q8_1, 0)));
const __m128i p16_3 = _mm_madd_epi16(scale_3, _mm_maddubs_epi16(q5l_3, _mm256_extractf128_si256(q8_1, 1)));
const __m128i s16_0 = _mm_madd_epi16(scale_0, _mm_maddubs_epi16(q5h_0, _mm256_extractf128_si256(q8_0, 0)));
const __m128i s16_1 = _mm_madd_epi16(scale_1, _mm_maddubs_epi16(q5h_1, _mm256_extractf128_si256(q8_0, 1)));
const __m128i s16_2 = _mm_madd_epi16(scale_2, _mm_maddubs_epi16(q5h_2, _mm256_extractf128_si256(q8_1, 0)));
const __m128i s16_3 = _mm_madd_epi16(scale_3, _mm_maddubs_epi16(q5h_3, _mm256_extractf128_si256(q8_1, 1)));
const __m128i dot_0 = _mm_sub_epi32(_mm_add_epi32(p16_0, p16_2), _mm_add_epi32(s16_0, s16_2));
const __m128i dot_1 = _mm_sub_epi32(_mm_add_epi32(p16_1, p16_3), _mm_add_epi32(s16_1, s16_3));
ggml : workaround for missing _mm256_setr_m128i in GCC < 8 in k_quants.c (#2405)
2023-07-28 18:17:45 +00:00
acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(dot_1, dot_0))), acc);
k_quants : add AVX support to dot functions with QK_K as 64 (#2339) * add AVX to ggml_vec_dot_q2_K_q8_K() * add AVX to ggml_vec_dot_q3_K_q8_K() * add AVX to ggml_vec_dot_q4_K_q8_K() * add AVX to ggml_vec_dot_q5_K_q8_K() * add AVX to ggml_vec_dot_q6_K_q8_K() * refactor AVX code in ggml_vec_dot_q6_K_q8_K()
2023-07-25 12:13:41 +00:00
}
*s = hsum_float_8(acc);
ggml : add RISC-V Vector Support for K-Quants and improved the existing intrinsics (#3453) * Added RVV intrinsics support for Q8 quantize row and also improved the existing dot product function for risc-v. The RVV intrinsics is added for the following quantize row functions quantize_row_q8_0 quantize_row_q8_1 The following dot product functions have also been optimized by using LMUL = 1/2 instead of LMUL = 1 ggml_vec_dot_q4_0_q8_0 ggml_vec_dot_q4_1_q8_1 ggml_vec_dot_q5_0_q8_0 ggml_vec_dot_q5_1_q8_1 And vector initialization in Q5 by temporary array is also replaced by the vid intrinsics Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai> * Added RVV intrinsics support for k_quants This adds RISC-V Vector intrinsics support for the following K_quants functions for both QKK = 256 and QKK = 64 ggml_vec_dot_q2_K_q8_K ggml_vec_dot_q3_K_q8_K ggml_vec_dot_q4_K_q8_K ggml_vec_dot_q5_K_q8_K ggml_vec_dot_q6_K_q8_K Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai> --------- Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai>
2023-10-03 18:38:19 +00:00
#elif defined __riscv_v_intrinsic
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const float d = y[i].d * (float)x[i].d;
const int8_t * sc = x[i].scales;
const uint8_t * restrict q5 = x[i].qs;
const uint8_t * restrict qh = x[i].qh;
const int8_t * restrict q8 = y[i].qs;
vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
// load qh
vuint8mf4_t qh_x1 = __riscv_vle8_v_u8mf4(qh, 8);
vuint8mf2_t qh_x2 = __riscv_vlmul_ext_v_u8mf4_u8mf2(__riscv_vsrl_vx_u8mf4(qh_x1, 1, 8));
size_t vl = 16;
// combine both qh_1 and qh_2
vuint8mf2_t qh_x = __riscv_vslideup_vx_u8mf2(__riscv_vlmul_ext_v_u8mf4_u8mf2(qh_x1), qh_x2, vl/2, vl);
vuint8mf2_t qh_h0 = __riscv_vand_vx_u8mf2(__riscv_vnot_v_u8mf2(__riscv_vsll_vx_u8mf2(qh_x, 0x4, vl), vl), 16, vl);
vuint8mf2_t qh_h1 = __riscv_vand_vx_u8mf2(__riscv_vnot_v_u8mf2(__riscv_vsll_vx_u8mf2(qh_x, 0x2, vl), vl), 16, vl);
vuint8mf2_t qh_h2 = __riscv_vand_vx_u8mf2(__riscv_vnot_v_u8mf2(qh_x, vl), 16, vl);
vuint8mf2_t qh_h3 = __riscv_vand_vx_u8mf2(__riscv_vnot_v_u8mf2(__riscv_vsrl_vx_u8mf2(qh_x, 0x4, vl), vl), 16, vl);
vint8mf2_t qh_0 = __riscv_vreinterpret_v_u8mf2_i8mf2(qh_h0);
vint8mf2_t qh_1 = __riscv_vreinterpret_v_u8mf2_i8mf2(qh_h1);
vint8mf2_t qh_2 = __riscv_vreinterpret_v_u8mf2_i8mf2(qh_h2);
vint8mf2_t qh_3 = __riscv_vreinterpret_v_u8mf2_i8mf2(qh_h3);
// load q5
vuint8mf2_t q5_x1 = __riscv_vle8_v_u8mf2(q5, vl);
vuint8mf2_t q5_x2 = __riscv_vle8_v_u8mf2(q5+16, vl);
vint8mf2_t q5s_0 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(q5_x1, 0xF, vl));
vint8mf2_t q5s_1 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(q5_x2, 0xF, vl));
vint8mf2_t q5s_2 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vsrl_vx_u8mf2(q5_x1, 0x4, vl));
vint8mf2_t q5s_3 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vsrl_vx_u8mf2(q5_x2, 0x4, vl));
vint8mf2_t q5_0 = __riscv_vsub_vv_i8mf2(q5s_0, qh_0, vl);
vint8mf2_t q5_1 = __riscv_vsub_vv_i8mf2(q5s_1, qh_1, vl);
vint8mf2_t q5_2 = __riscv_vsub_vv_i8mf2(q5s_2, qh_2, vl);
vint8mf2_t q5_3 = __riscv_vsub_vv_i8mf2(q5s_3, qh_3, vl);
// load Q8 and multiply it with Q5
vint16m1_t p0 = __riscv_vwmul_vv_i16m1(q5_0, __riscv_vle8_v_i8mf2(q8, vl), vl);
vint16m1_t p1 = __riscv_vwmul_vv_i16m1(q5_1, __riscv_vle8_v_i8mf2(q8+16, vl), vl);
vint16m1_t p2 = __riscv_vwmul_vv_i16m1(q5_2, __riscv_vle8_v_i8mf2(q8+32, vl), vl);
vint16m1_t p3 = __riscv_vwmul_vv_i16m1(q5_3, __riscv_vle8_v_i8mf2(q8+48, vl), vl);
vint32m1_t vs_0 = __riscv_vwredsum_vs_i16m1_i32m1(p0, vzero, vl);
vint32m1_t vs_1 = __riscv_vwredsum_vs_i16m1_i32m1(p1, vzero, vl);
vint32m1_t vs_2 = __riscv_vwredsum_vs_i16m1_i32m1(p2, vzero, vl);
vint32m1_t vs_3 = __riscv_vwredsum_vs_i16m1_i32m1(p3, vzero, vl);
int32_t sumi1 = sc[0] * __riscv_vmv_x_s_i32m1_i32(vs_0);
int32_t sumi2 = sc[1] * __riscv_vmv_x_s_i32m1_i32(vs_1);
int32_t sumi3 = sc[2] * __riscv_vmv_x_s_i32m1_i32(vs_2);
int32_t sumi4 = sc[3] * __riscv_vmv_x_s_i32m1_i32(vs_3);
sumf += d * (sumi1 + sumi2 + sumi3 + sumi4);
}
*s = sumf;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
#else
Fix scalar version of Q5_K when QK_K = 64 (#2362) Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-07-24 09:55:02 +00:00
int8_t aux8[QK_K];
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
int16_t aux16[16];
float sums [8];
memset(sums, 0, 8*sizeof(float));
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * restrict q4 = x[i].qs;
const uint8_t * restrict hm = x[i].qh;
const int8_t * restrict q8 = y[i].qs;
Fix scalar version of Q5_K when QK_K = 64 (#2362) Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-07-24 09:55:02 +00:00
int8_t * restrict a = aux8;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
for (int l = 0; l < 32; ++l) {
a[l+ 0] = q4[l] & 0xF;
a[l+32] = q4[l] >> 4;
}
for (int is = 0; is < 8; ++is) {
uint8_t m = 1 << is;
for (int l = 0; l < 8; ++l) a[8*is + l] -= (hm[l] & m ? 0 : 16);
}
const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
const int8_t * restrict sc = x[i].scales;
for (int j = 0; j < QK_K/16; ++j) {
const float dl = d * sc[j];
for (int l = 0; l < 16; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) sums[l] += dl * (aux16[l] + aux16[8+l]);
q8 += 16; a += 16;
}
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
#endif
}
#endif
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
#if QK_K == 256
k-quants : allow to optionally disable at compile time (#1734) * k-quants : put behind optional compile flag LLAMA_K_QUANTS * build : enable k-quants by default
2023-06-07 07:59:52 +00:00
void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
assert(n % QK_K == 0);
k-quants : allow to optionally disable at compile time (#1734) * k-quants : put behind optional compile flag LLAMA_K_QUANTS * build : enable k-quants by default
2023-06-07 07:59:52 +00:00
const block_q6_K * restrict x = vx;
const block_q8_K * restrict y = vy;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
const int nb = n / QK_K;
#ifdef __ARM_NEON
float sum = 0;
const uint8x16_t m4b = vdupq_n_u8(0xF);
k-quants : fix build on armv7 (android only) (#2920) * k-quants : fix build on armv7 * ggml : cleanup unused arm32 specific impl * k-quants : avoid some unused vzero / mzero define * ggml-alloc : use 4g for MEASURE_MAX_SIZE in 32-bit arm
2023-09-02 12:23:45 +00:00
#if defined(__ARM_FEATURE_DOTPROD)
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
const int32x4_t vzero = vdupq_n_s32(0);
k-quants : fix build on armv7 (android only) (#2920) * k-quants : fix build on armv7 * ggml : cleanup unused arm32 specific impl * k-quants : avoid some unused vzero / mzero define * ggml-alloc : use 4g for MEASURE_MAX_SIZE in 32-bit arm
2023-09-02 12:23:45 +00:00
#endif
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
//const int8x16_t m32s = vdupq_n_s8(32);
const uint8x16_t mone = vdupq_n_u8(3);
int8x16x4_t q6bytes;
uint8x16x4_t q6h;
for (int i = 0; i < nb; ++i) {
const float d_all = ggml_fp16_to_fp32(x[i].d);
const uint8_t * restrict q6 = x[i].ql;
const uint8_t * restrict qh = x[i].qh;
const int8_t * restrict q8 = y[i].qs;
const int8_t * restrict scale = x[i].scales;
const int16x8x2_t q8sums = vld1q_s16_x2(y[i].bsums);
const int8x16_t scales = vld1q_s8(scale);
const int16x8x2_t q6scales = {vmovl_s8(vget_low_s8(scales)), vmovl_s8(vget_high_s8(scales))};
const int32x4_t prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums.val[0]), vget_low_s16 (q6scales.val[0])),
vmull_s16(vget_high_s16(q8sums.val[0]), vget_high_s16(q6scales.val[0]))),
vaddq_s32(vmull_s16(vget_low_s16 (q8sums.val[1]), vget_low_s16 (q6scales.val[1])),
vmull_s16(vget_high_s16(q8sums.val[1]), vget_high_s16(q6scales.val[1]))));
int32_t isum_mins = vaddvq_s32(prod);
int32_t isum = 0;
for (int j = 0; j < QK_K/128; ++j) {
uint8x16x2_t qhbits = vld1q_u8_x2(qh); qh += 32;
Revert "ggml : load data into int8x16x4_t using vld4q_s8 on arm64 (#1738)" This reverts commit 8432d4d9f716b25133e3ed671d91e21f6f3be867.
2023-06-08 17:48:14 +00:00
uint8x16x4_t q6bits = vld1q_u8_x4(q6); q6 += 64;
int8x16x4_t q8bytes = vld1q_s8_x4(q8); q8 += 64;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
q6h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits.val[0]), 4);
q6h.val[1] = vshlq_n_u8(vandq_u8(mone, qhbits.val[1]), 4);
uint8x16_t shifted = vshrq_n_u8(qhbits.val[0], 2);
q6h.val[2] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
shifted = vshrq_n_u8(qhbits.val[1], 2);
q6h.val[3] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
//q6bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[0], m4b), q6h.val[0])), m32s);
//q6bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[1], m4b), q6h.val[1])), m32s);
//q6bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[2], m4b), q6h.val[2])), m32s);
//q6bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[3], m4b), q6h.val[3])), m32s);
q6bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[0], m4b), q6h.val[0]));
q6bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[1], m4b), q6h.val[1]));
q6bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[2], m4b), q6h.val[2]));
q6bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[3], m4b), q6h.val[3]));
#if defined(__ARM_FEATURE_DOTPROD)
isum += vaddvq_s32(vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] +
vaddvq_s32(vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] +
vaddvq_s32(vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] +
vaddvq_s32(vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3];
scale += 4;
#else
int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[0]), vget_low_s8 (q8bytes.val[0])),
vmull_s8(vget_high_s8(q6bytes.val[0]), vget_high_s8(q8bytes.val[0])));
int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[1]), vget_low_s8 (q8bytes.val[1])),
vmull_s8(vget_high_s8(q6bytes.val[1]), vget_high_s8(q8bytes.val[1])));
isum += vaddvq_s16(p0) * scale[0] + vaddvq_s16(p1) * scale[1];
scale += 2;
int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[2]), vget_low_s8 (q8bytes.val[2])),
vmull_s8(vget_high_s8(q6bytes.val[2]), vget_high_s8(q8bytes.val[2])));
int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[3]), vget_low_s8 (q8bytes.val[3])),
vmull_s8(vget_high_s8(q6bytes.val[3]), vget_high_s8(q8bytes.val[3])));
isum += vaddvq_s16(p2) * scale[0] + vaddvq_s16(p3) * scale[1];
scale += 2;
#endif
Revert "ggml : load data into int8x16x4_t using vld4q_s8 on arm64 (#1738)" This reverts commit 8432d4d9f716b25133e3ed671d91e21f6f3be867.
2023-06-08 17:48:14 +00:00
q8bytes = vld1q_s8_x4(q8); q8 += 64;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
shifted = vshrq_n_u8(qhbits.val[0], 4);
q6h.val[0] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
shifted = vshrq_n_u8(qhbits.val[1], 4);
q6h.val[1] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
shifted = vshrq_n_u8(qhbits.val[0], 6);
q6h.val[2] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
shifted = vshrq_n_u8(qhbits.val[1], 6);
q6h.val[3] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
//q6bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[0], 4), q6h.val[0])), m32s);
//q6bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[1], 4), q6h.val[1])), m32s);
//q6bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[2], 4), q6h.val[2])), m32s);
//q6bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[3], 4), q6h.val[3])), m32s);
q6bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[0], 4), q6h.val[0]));
q6bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[1], 4), q6h.val[1]));
q6bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[2], 4), q6h.val[2]));
q6bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[3], 4), q6h.val[3]));
#if defined(__ARM_FEATURE_DOTPROD)
isum += vaddvq_s32(vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] +
vaddvq_s32(vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] +
vaddvq_s32(vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] +
vaddvq_s32(vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3];
scale += 4;
//for (int l = 0; l < 4; ++l) {
// const int32x4_t p = vdotq_s32(vzero, q6bytes.val[l], q8bytes.val[l]);
// isum += vaddvq_s32(p) * *scale++;
//}
#else
p0 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[0]), vget_low_s8 (q8bytes.val[0])),
vmull_s8(vget_high_s8(q6bytes.val[0]), vget_high_s8(q8bytes.val[0])));
p1 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[1]), vget_low_s8 (q8bytes.val[1])),
vmull_s8(vget_high_s8(q6bytes.val[1]), vget_high_s8(q8bytes.val[1])));
isum += vaddvq_s16(p0) * scale[0] + vaddvq_s16(p1) * scale[1];
scale += 2;
p2 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[2]), vget_low_s8 (q8bytes.val[2])),
vmull_s8(vget_high_s8(q6bytes.val[2]), vget_high_s8(q8bytes.val[2])));
p3 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[3]), vget_low_s8 (q8bytes.val[3])),
vmull_s8(vget_high_s8(q6bytes.val[3]), vget_high_s8(q8bytes.val[3])));
isum += vaddvq_s16(p2) * scale[0] + vaddvq_s16(p3) * scale[1];
scale += 2;
#endif
}
//sum += isum * d_all * y[i].d;
sum += d_all * y[i].d * (isum - 32 * isum_mins);
}
*s = sum;
#elif defined __AVX2__
const __m256i m4 = _mm256_set1_epi8(0xF);
const __m256i m2 = _mm256_set1_epi8(3);
const __m256i m32s = _mm256_set1_epi8(32);
__m256 acc = _mm256_setzero_ps();
for (int i = 0; i < nb; ++i) {
const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
const uint8_t * restrict q4 = x[i].ql;
const uint8_t * restrict qh = x[i].qh;
const int8_t * restrict q8 = y[i].qs;
const __m128i scales = _mm_loadu_si128((const __m128i*)x[i].scales);
__m256i sumi = _mm256_setzero_si256();
int is = 0;
for (int j = 0; j < QK_K/128; ++j) {
const __m128i scale_0 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 0));
const __m128i scale_1 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 1));
const __m128i scale_2 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 2));
const __m128i scale_3 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 3));
is += 4;
const __m256i q4bits1 = _mm256_loadu_si256((const __m256i*)q4); q4 += 32;
const __m256i q4bits2 = _mm256_loadu_si256((const __m256i*)q4); q4 += 32;
const __m256i q4bitsH = _mm256_loadu_si256((const __m256i*)qh); qh += 32;
const __m256i q4h_0 = _mm256_slli_epi16(_mm256_and_si256(q4bitsH, m2), 4);
const __m256i q4h_1 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bitsH, 2), m2), 4);
const __m256i q4h_2 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bitsH, 4), m2), 4);
const __m256i q4h_3 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bitsH, 6), m2), 4);
const __m256i q4_0 = _mm256_or_si256(_mm256_and_si256(q4bits1, m4), q4h_0);
const __m256i q4_1 = _mm256_or_si256(_mm256_and_si256(q4bits2, m4), q4h_1);
const __m256i q4_2 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits1, 4), m4), q4h_2);
const __m256i q4_3 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits2, 4), m4), q4h_3);
const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
const __m256i q8_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
const __m256i q8_3 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
__m256i q8s_0 = _mm256_maddubs_epi16(m32s, q8_0);
__m256i q8s_1 = _mm256_maddubs_epi16(m32s, q8_1);
__m256i q8s_2 = _mm256_maddubs_epi16(m32s, q8_2);
__m256i q8s_3 = _mm256_maddubs_epi16(m32s, q8_3);
__m256i p16_0 = _mm256_maddubs_epi16(q4_0, q8_0);
__m256i p16_1 = _mm256_maddubs_epi16(q4_1, q8_1);
__m256i p16_2 = _mm256_maddubs_epi16(q4_2, q8_2);
__m256i p16_3 = _mm256_maddubs_epi16(q4_3, q8_3);
p16_0 = _mm256_sub_epi16(p16_0, q8s_0);
p16_1 = _mm256_sub_epi16(p16_1, q8s_1);
p16_2 = _mm256_sub_epi16(p16_2, q8s_2);
p16_3 = _mm256_sub_epi16(p16_3, q8s_3);
p16_0 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_0), p16_0);
p16_1 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_1), p16_1);
p16_2 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_2), p16_2);
p16_3 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_3), p16_3);
sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_1));
sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_2, p16_3));
}
acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
}
*s = hsum_float_8(acc);
k-quants : add AVX support to dot functions (#1916) * k_quants : add AVX support * k_quants : apply review comments
2023-06-26 16:46:07 +00:00
#elif defined __AVX__
const __m128i m4 = _mm_set1_epi8(0xF);
const __m128i m3 = _mm_set1_epi8(3);
const __m128i m32s = _mm_set1_epi8(32);
const __m128i m2 = _mm_set1_epi8(2);
__m256 acc = _mm256_setzero_ps();
for (int i = 0; i < nb; ++i) {
const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
const uint8_t * restrict q4 = x[i].ql;
const uint8_t * restrict qh = x[i].qh;
const int8_t * restrict q8 = y[i].qs;
const __m128i scales = _mm_loadu_si128((const __m128i*)x[i].scales);
__m128i sumi_0 = _mm_setzero_si128();
__m128i sumi_1 = _mm_setzero_si128();
__m128i shuffle = _mm_set_epi64x(0x0101010101010101, 0x0000000000000000);
for (int j = 0; j < QK_K/128; ++j) {
const __m128i q4bitsH_0 = _mm_loadu_si128((const __m128i*)qh); qh += 16;
const __m128i q4bitsH_1 = _mm_loadu_si128((const __m128i*)qh); qh += 16;
const __m128i q4h_0 = _mm_slli_epi16(_mm_and_si128(q4bitsH_0, m3), 4);
const __m128i q4h_1 = _mm_slli_epi16(_mm_and_si128(q4bitsH_1, m3), 4);
const __m128i q4h_2 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_0, 2), m3), 4);
const __m128i q4h_3 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_1, 2), m3), 4);
const __m128i q4h_4 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_0, 4), m3), 4);
const __m128i q4h_5 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_1, 4), m3), 4);
const __m128i q4h_6 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_0, 6), m3), 4);
const __m128i q4h_7 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_1, 6), m3), 4);
const __m128i q4bits1_0 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
const __m128i q4bits1_1 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
const __m128i q4bits2_0 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
const __m128i q4bits2_1 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
const __m128i q4_0 = _mm_or_si128(_mm_and_si128(q4bits1_0, m4), q4h_0);
const __m128i q4_1 = _mm_or_si128(_mm_and_si128(q4bits1_1, m4), q4h_1);
const __m128i q4_2 = _mm_or_si128(_mm_and_si128(q4bits2_0, m4), q4h_2);
const __m128i q4_3 = _mm_or_si128(_mm_and_si128(q4bits2_1, m4), q4h_3);
const __m128i q4_4 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits1_0, 4), m4), q4h_4);
const __m128i q4_5 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits1_1, 4), m4), q4h_5);
const __m128i q4_6 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits2_0, 4), m4), q4h_6);
const __m128i q4_7 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits2_1, 4), m4), q4h_7);
const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
__m128i q8s_0 = _mm_maddubs_epi16(m32s, q8_0);
__m128i q8s_1 = _mm_maddubs_epi16(m32s, q8_1);
__m128i q8s_2 = _mm_maddubs_epi16(m32s, q8_2);
__m128i q8s_3 = _mm_maddubs_epi16(m32s, q8_3);
__m128i q8s_4 = _mm_maddubs_epi16(m32s, q8_4);
__m128i q8s_5 = _mm_maddubs_epi16(m32s, q8_5);
__m128i q8s_6 = _mm_maddubs_epi16(m32s, q8_6);
__m128i q8s_7 = _mm_maddubs_epi16(m32s, q8_7);
__m128i p16_0 = _mm_maddubs_epi16(q4_0, q8_0);
__m128i p16_1 = _mm_maddubs_epi16(q4_1, q8_1);
__m128i p16_2 = _mm_maddubs_epi16(q4_2, q8_2);
__m128i p16_3 = _mm_maddubs_epi16(q4_3, q8_3);
__m128i p16_4 = _mm_maddubs_epi16(q4_4, q8_4);
__m128i p16_5 = _mm_maddubs_epi16(q4_5, q8_5);
__m128i p16_6 = _mm_maddubs_epi16(q4_6, q8_6);
__m128i p16_7 = _mm_maddubs_epi16(q4_7, q8_7);
p16_0 = _mm_sub_epi16(p16_0, q8s_0);
p16_1 = _mm_sub_epi16(p16_1, q8s_1);
p16_2 = _mm_sub_epi16(p16_2, q8s_2);
p16_3 = _mm_sub_epi16(p16_3, q8s_3);
p16_4 = _mm_sub_epi16(p16_4, q8s_4);
p16_5 = _mm_sub_epi16(p16_5, q8s_5);
p16_6 = _mm_sub_epi16(p16_6, q8s_6);
p16_7 = _mm_sub_epi16(p16_7, q8s_7);
const __m128i scale_0 = _mm_shuffle_epi8(scales, shuffle);
shuffle = _mm_add_epi8(shuffle, m2);
const __m128i scale_1 = _mm_shuffle_epi8(scales, shuffle);
shuffle = _mm_add_epi8(shuffle, m2);
const __m128i scale_2 = _mm_shuffle_epi8(scales, shuffle);
shuffle = _mm_add_epi8(shuffle, m2);
const __m128i scale_3 = _mm_shuffle_epi8(scales, shuffle);
shuffle = _mm_add_epi8(shuffle, m2);
p16_0 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_0), p16_0);
p16_1 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_0, scale_0)), p16_1);
p16_2 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_1), p16_2);
p16_3 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_1, scale_1)), p16_3);
p16_4 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_2), p16_4);
p16_5 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_2, scale_2)), p16_5);
p16_6 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_3), p16_6);
p16_7 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_3, scale_3)), p16_7);
sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_4, p16_6));
sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_5, p16_7));
}
ggml : workaround for missing _mm256_setr_m128i in GCC < 8 in k_quants.c (#2405)
2023-07-28 18:17:45 +00:00
__m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
k-quants : add AVX support to dot functions (#1916) * k_quants : add AVX support * k_quants : apply review comments
2023-06-26 16:46:07 +00:00
acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi)), acc);
}
*s = hsum_float_8(acc);
ggml : add RISC-V Vector Support for K-Quants and improved the existing intrinsics (#3453) * Added RVV intrinsics support for Q8 quantize row and also improved the existing dot product function for risc-v. The RVV intrinsics is added for the following quantize row functions quantize_row_q8_0 quantize_row_q8_1 The following dot product functions have also been optimized by using LMUL = 1/2 instead of LMUL = 1 ggml_vec_dot_q4_0_q8_0 ggml_vec_dot_q4_1_q8_1 ggml_vec_dot_q5_0_q8_0 ggml_vec_dot_q5_1_q8_1 And vector initialization in Q5 by temporary array is also replaced by the vid intrinsics Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai> * Added RVV intrinsics support for k_quants This adds RISC-V Vector intrinsics support for the following K_quants functions for both QKK = 256 and QKK = 64 ggml_vec_dot_q2_K_q8_K ggml_vec_dot_q3_K_q8_K ggml_vec_dot_q4_K_q8_K ggml_vec_dot_q5_K_q8_K ggml_vec_dot_q6_K_q8_K Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai> --------- Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai>
2023-10-03 18:38:19 +00:00
#elif defined __riscv_v_intrinsic
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const float d = ggml_fp16_to_fp32(x[i].d) * y[i].d;
const uint8_t * restrict q6 = x[i].ql;
const uint8_t * restrict qh = x[i].qh;
const int8_t * restrict q8 = y[i].qs;
const int8_t * restrict scale = x[i].scales;
size_t vl;
vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
int sum_t = 0;
int is = 0;
for (int j = 0; j < QK_K/128; ++j) {
vl = 32;
// load qh
vuint8m1_t qh_x = __riscv_vle8_v_u8m1(qh, vl);
// load Q6
vuint8m1_t q6_0 = __riscv_vle8_v_u8m1(q6, vl);
vuint8m1_t q6_1 = __riscv_vle8_v_u8m1(q6+32, vl);
vuint8m1_t q6a_0 = __riscv_vand_vx_u8m1(q6_0, 0x0F, vl);
vuint8m1_t q6a_1 = __riscv_vand_vx_u8m1(q6_1, 0x0F, vl);
vuint8m1_t q6s_0 = __riscv_vsrl_vx_u8m1(q6_0, 0x04, vl);
vuint8m1_t q6s_1 = __riscv_vsrl_vx_u8m1(q6_1, 0x04, vl);
vuint8m1_t qh_0 = __riscv_vand_vx_u8m1(qh_x, 0x03, vl);
vuint8m1_t qh_1 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(qh_x, 0x2, vl), 0x03 , vl);
vuint8m1_t qh_2 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(qh_x, 0x4, vl), 0x03 , vl);
vuint8m1_t qh_3 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(qh_x, 0x6, vl), 0x03 , vl);
vuint8m1_t qhi_0 = __riscv_vor_vv_u8m1(q6a_0, __riscv_vsll_vx_u8m1(qh_0, 0x04, vl), vl);
vuint8m1_t qhi_1 = __riscv_vor_vv_u8m1(q6a_1, __riscv_vsll_vx_u8m1(qh_1, 0x04, vl), vl);
vuint8m1_t qhi_2 = __riscv_vor_vv_u8m1(q6s_0, __riscv_vsll_vx_u8m1(qh_2, 0x04, vl), vl);
vuint8m1_t qhi_3 = __riscv_vor_vv_u8m1(q6s_1, __riscv_vsll_vx_u8m1(qh_3, 0x04, vl), vl);
vint8m1_t a_0 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_0), 32, vl);
vint8m1_t a_1 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_1), 32, vl);
vint8m1_t a_2 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_2), 32, vl);
vint8m1_t a_3 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_3), 32, vl);
// load Q8 and take product
vint16m2_t va_q_0 = __riscv_vwmul_vv_i16m2(a_0, __riscv_vle8_v_i8m1(q8, vl), vl);
vint16m2_t va_q_1 = __riscv_vwmul_vv_i16m2(a_1, __riscv_vle8_v_i8m1(q8+32, vl), vl);
vint16m2_t va_q_2 = __riscv_vwmul_vv_i16m2(a_2, __riscv_vle8_v_i8m1(q8+64, vl), vl);
vint16m2_t va_q_3 = __riscv_vwmul_vv_i16m2(a_3, __riscv_vle8_v_i8m1(q8+96, vl), vl);
vl = 16;
vint32m2_t vaux_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_0, 0), scale[is+0], vl);
vint32m2_t vaux_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_0, 1), scale[is+1], vl);
vint32m2_t vaux_2 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_1, 0), scale[is+2], vl);
vint32m2_t vaux_3 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_1, 1), scale[is+3], vl);
vint32m2_t vaux_4 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_2, 0), scale[is+4], vl);
vint32m2_t vaux_5 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_2, 1), scale[is+5], vl);
vint32m2_t vaux_6 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_3, 0), scale[is+6], vl);
vint32m2_t vaux_7 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_3, 1), scale[is+7], vl);
vint32m1_t isum0 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_0, vaux_1, vl), vzero, vl);
vint32m1_t isum1 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_2, vaux_3, vl), isum0, vl);
vint32m1_t isum2 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_4, vaux_5, vl), isum1, vl);
vint32m1_t isum3 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_6, vaux_7, vl), isum2, vl);
sum_t += __riscv_vmv_x_s_i32m1_i32(isum3);
q6 += 64; qh += 32; q8 += 128; is=8;
}
sumf += d * sum_t;
}
*s = sumf;
ggml : add SOTA 2,3,4,5,6 bit k-quantizations (#1684) * Starting to add k-quantization to ggml I think it is better to have quantization separate from ggml. For now just adding the k-quants there, but it would be better to also factor out the existing ggml quantizations. * Adding Q3_K and Q8_K (de)-quantization * Q3_K now working on CUDA and AVX2/scalar CUDA is not ideal - ~50% slower than Q4_0 for single token prediction, about the same in batch mode (perplexity). CPU single token is ~55 ms (on Ryzen 7950X). * Some improvement for Q3_K on CUDA It is now ~22.5 ms/token on my GPU, so ~30% slower than Q4_0. * Some more CUDA optimizations for Q3_K Single token is now 20.5 ms/token (~20% slower than Q4_0). Perplexity is on par with Q4_0. * Adding Q4_K - scalar, AVX2, CUDA Performance is the same or perhaps very slightly better than Q4_0 on the CPU. On the GPU, single token prediction is ~10% better than Q4_0, batch mode (perplexity is about the same). * Adding Q6_K - scalar, AVX2, CUDA Performance is ~40% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 6-bit model is ~44% larger than the 4-bit. On the GPU, single token prediction is ~6% lower than Q4_0, batch mode (perplexity) is even closer (but still slower). * Adding Q5_K - scalar, AVX2, CUDA Performance is ~20% lower compared to Q4_K on the CPU. This is to be expected, considering that we are memory bound on the CPU and the 5-bit model is ~22% larger than the 4-bit. On the GPU, single token prediction is about the same as Q4_0 for both, single token and batch prediction. * Per convention, all QX_K quantizations use Q5_K for output.weight * Adding quantization mixes * Quantization mixes: didn't quite get what I wanted in the last commit * Q4_K dot product for ARM_NEON * Q6_K dot product for ARM_NEON * Q5_K dot product for ARM_NEON * Adding Q3_K dot for ARM_NEON It is 22% slower than Q4_K, despite the smaller model size. On x86_64, where we are memory bound, the Q3_K model is quite a bit faster than Q4_K. * A very slightly faster ARM_NEON Q3_K dot * Adding Q2_K - just CUDA for now Token prediction is pretty good - about 15.5 ms on a RTX 4080. Perplexity is about the same as Q4_K. * Adding scalar and AVX2 Q2_K dot * Adding ARM_NEON Q2_K dot About the same performance as Q4_K. * A slightly faster ARM_NEON Q2_K dot Single token prediction is now ~36 ms on M2 Max. The code is much simpler too. * Fixed bug in Q2_K CUDA dot product kernel Stranegly enough, for the few prompts I tried with the 7B model the responses looked perfectly reasonable. Only realized something is not quite right when I tried the larger models and started getting nonse back. In any case, Q2_K single token evaluation time on an RTX 4080 in a Ryzen7950X box iusing CUDA and model fully loaded on the GPU are ~15.5 ms for 7B, ~25.4 ms for 13B, and ~55.8 ms for 30B. The max number of layers that fit in VRAM for The 65B is 32. With that, we get ~330 ms per token, which is not that much faster than just running on the CPU (~470 ms per token). * Don't print zeros/NaNs when no count histogram has been collected * A 10% faster CUDA vector dot kernel for Q3_K Q3_K is now running at ~18.5 ms / token on CUDA, so the gap to Q4_0 is only 10%. It seems memory acccess pattern is more important for performance than the amount of computation the kernel does. * A slightly daster Q4_K AVX2 dot product For perplexity, where we are less memory bound, time per pass drops by ~5%. Barely measurable difference for single token prediction. * A slightly faster ARM_NEON A4_K dot product * Minor * Fix quantization error test We cannot possibly be expecting rmse < 0.002 for 2- and 3-bit quantization variants. * Fix docker build I have been sloppy with vector reinterpret casts on ARM_NEON. It seems clang is very forgiving in that regard. * Added forgotten ggml.o dependence on k_quants.h to the Makefile * Had unintentionally committed the Makefile with -Ofast enabled * ggml : rename k_quants -> ggml-quants-k, use lowercase in code --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-06-05 19:56:18 +00:00
#else
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * restrict q4 = x[i].ql;
const uint8_t * restrict qh = x[i].qh;
const int8_t * restrict q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * restrict a = aux8;
for (int j = 0; j < QK_K; j += 128) {
for (int l = 0; l < 32; ++l) {
a[l + 0] = (int8_t)((q4[l + 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
a[l + 64] = (int8_t)((q4[l + 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32;
a[l + 96] = (int8_t)((q4[l + 32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32;
}
a += 128;
q4 += 64;
qh += 32;
}
a = aux8;
int is = 0;
for (int j = 0; j < QK_K/16; ++j) {
int scale = x[i].scales[is++];
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
}
const float d = ggml_fp16_to_fp32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
#endif
}
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
#else
void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
assert(n % QK_K == 0);
const block_q6_K * restrict x = vx;
const block_q8_K * restrict y = vy;
const int nb = n / QK_K;
#ifdef __ARM_NEON
float sum = 0;
const uint8x16_t m4b = vdupq_n_u8(0xF);
const int8x16_t m32s = vdupq_n_s8(32);
k-quants : fix build on armv7 (android only) (#2920) * k-quants : fix build on armv7 * ggml : cleanup unused arm32 specific impl * k-quants : avoid some unused vzero / mzero define * ggml-alloc : use 4g for MEASURE_MAX_SIZE in 32-bit arm
2023-09-02 12:23:45 +00:00
#if defined(__ARM_FEATURE_DOTPROD)
const int32x4_t vzero = vdupq_n_s32(0);
#endif
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
const uint8x16_t mone = vdupq_n_u8(3);
int8x16x4_t q6bytes;
uint8x16x4_t q6h;
for (int i = 0; i < nb; ++i) {
const float d_all = (float)x[i].d;
const uint8_t * restrict q6 = x[i].ql;
const uint8_t * restrict qh = x[i].qh;
const int8_t * restrict q8 = y[i].qs;
const int8_t * restrict scale = x[i].scales;
int32_t isum = 0;
uint8x16_t qhbits = vld1q_u8(qh);
uint8x16x2_t q6bits = vld1q_u8_x2(q6);
int8x16x4_t q8bytes = vld1q_s8_x4(q8);
q6h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits), 4);
uint8x16_t shifted = vshrq_n_u8(qhbits, 2);
q6h.val[1] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
shifted = vshrq_n_u8(qhbits, 4);
q6h.val[2] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
shifted = vshrq_n_u8(qhbits, 6);
q6h.val[3] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
q6bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[0], m4b), q6h.val[0])), m32s);
q6bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[1], m4b), q6h.val[1])), m32s);
q6bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[0], 4), q6h.val[2])), m32s);
q6bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[1], 4), q6h.val[3])), m32s);
#if defined(__ARM_FEATURE_DOTPROD)
isum += vaddvq_s32(vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] +
vaddvq_s32(vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] +
vaddvq_s32(vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] +
vaddvq_s32(vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3];
#else
int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[0]), vget_low_s8 (q8bytes.val[0])),
vmull_s8(vget_high_s8(q6bytes.val[0]), vget_high_s8(q8bytes.val[0])));
int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[1]), vget_low_s8 (q8bytes.val[1])),
vmull_s8(vget_high_s8(q6bytes.val[1]), vget_high_s8(q8bytes.val[1])));
isum += vaddvq_s16(p0) * scale[0] + vaddvq_s16(p1) * scale[1];
int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[2]), vget_low_s8 (q8bytes.val[2])),
vmull_s8(vget_high_s8(q6bytes.val[2]), vget_high_s8(q8bytes.val[2])));
int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q6bytes.val[3]), vget_low_s8 (q8bytes.val[3])),
vmull_s8(vget_high_s8(q6bytes.val[3]), vget_high_s8(q8bytes.val[3])));
isum += vaddvq_s16(p2) * scale[2] + vaddvq_s16(p3) * scale[3];
#endif
sum += isum * d_all * y[i].d;
}
*s = sum;
#elif defined __AVX2__
const __m256i m4 = _mm256_set1_epi8(0xF);
const __m256i m2 = _mm256_set1_epi8(3);
const __m256i m32s = _mm256_set1_epi8(32);
__m256 acc = _mm256_setzero_ps();
for (int i = 0; i < nb; ++i) {
const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
const uint8_t * restrict q4 = x[i].ql;
const uint8_t * restrict qh = x[i].qh;
const int8_t * restrict q8 = y[i].qs;
const __m64 scales_1 = _mm_set1_pi8(x[i].scales[0]);
const __m64 scales_2 = _mm_set1_pi8(x[i].scales[1]);
const __m64 scales_3 = _mm_set1_pi8(x[i].scales[2]);
const __m64 scales_4 = _mm_set1_pi8(x[i].scales[3]);
__m256i sumi = _mm256_setzero_si256();
const __m128i scale_0 = _mm_set_epi64(scales_2, scales_1);
const __m128i scale_1 = _mm_set_epi64(scales_4, scales_3);
const __m256i q4bits1 = _mm256_loadu_si256((const __m256i*)q4);
const __m128i q4bitsH = _mm_loadu_si128((const __m128i*)qh);
ggml : workaround for missing _mm256_setr_m128i in GCC < 8 in k_quants.c (#2405)
2023-07-28 18:17:45 +00:00
const __m256i q4h_0 = _mm256_slli_epi16(_mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(q4bitsH, 2), q4bitsH), m2), 4);
const __m256i q4h_1 = _mm256_slli_epi16(_mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(q4bitsH, 6), _mm_srli_epi16(q4bitsH, 4)), m2), 4);
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
const __m256i q4_0 = _mm256_or_si256(_mm256_and_si256(q4bits1, m4), q4h_0);
const __m256i q4_1 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits1, 4), m4), q4h_1);
const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
__m256i q8s_0 = _mm256_maddubs_epi16(m32s, q8_0);
__m256i q8s_1 = _mm256_maddubs_epi16(m32s, q8_1);
__m256i p16_0 = _mm256_maddubs_epi16(q4_0, q8_0);
__m256i p16_1 = _mm256_maddubs_epi16(q4_1, q8_1);
p16_0 = _mm256_sub_epi16(p16_0, q8s_0);
p16_1 = _mm256_sub_epi16(p16_1, q8s_1);
p16_0 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_0), p16_0);
p16_1 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_1), p16_1);
sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_1));
acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
}
*s = hsum_float_8(acc);
k_quants : add AVX support to dot functions with QK_K as 64 (#2339) * add AVX to ggml_vec_dot_q2_K_q8_K() * add AVX to ggml_vec_dot_q3_K_q8_K() * add AVX to ggml_vec_dot_q4_K_q8_K() * add AVX to ggml_vec_dot_q5_K_q8_K() * add AVX to ggml_vec_dot_q6_K_q8_K() * refactor AVX code in ggml_vec_dot_q6_K_q8_K()
2023-07-25 12:13:41 +00:00
#elif defined __AVX__
const __m128i m4 = _mm_set1_epi8(0xF);
const __m128i m2 = _mm_set1_epi8(3);
const __m128i m32s = _mm_set1_epi8(32);
__m256 acc = _mm256_setzero_ps();
for (int i = 0; i < nb; ++i) {
const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
const uint8_t * restrict q4 = x[i].ql;
const uint8_t * restrict qh = x[i].qh;
const int8_t * restrict q8 = y[i].qs;
const __m64 scales_1 = _mm_set1_pi8(x[i].scales[0]);
const __m64 scales_2 = _mm_set1_pi8(x[i].scales[1]);
const __m64 scales_3 = _mm_set1_pi8(x[i].scales[2]);
const __m64 scales_4 = _mm_set1_pi8(x[i].scales[3]);
__m128i sumi_0 = _mm_setzero_si128();
__m128i sumi_1 = _mm_setzero_si128();
const __m128i scale_0 = _mm_set_epi64(scales_2, scales_1);
const __m128i scale_1 = _mm_set_epi64(scales_4, scales_3);
const __m256i q4bits1 = _mm256_loadu_si256((const __m256i*)q4);
const __m128i q4bitsH = _mm_loadu_si128((const __m128i*)qh);
const __m128i q4h_0 = _mm_slli_epi16(_mm_and_si128(q4bitsH, m2), 4);
const __m128i q4h_1 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH, 2), m2), 4);
const __m128i q4h_2 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH, 4), m2), 4);
const __m128i q4h_3 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH, 6), m2), 4);
const __m128i q4_0 = _mm_or_si128(_mm_and_si128(_mm256_extractf128_si256(q4bits1, 0), m4), q4h_0);
const __m128i q4_1 = _mm_or_si128(_mm_and_si128(_mm256_extractf128_si256(q4bits1, 1), m4), q4h_1);
const __m128i q4_2 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(_mm256_extractf128_si256(q4bits1, 0), 4), m4), q4h_2);
const __m128i q4_3 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(_mm256_extractf128_si256(q4bits1, 1), 4), m4), q4h_3);
const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
__m128i q8s_0 = _mm_maddubs_epi16(m32s, _mm256_extractf128_si256(q8_0, 0));
__m128i q8s_1 = _mm_maddubs_epi16(m32s, _mm256_extractf128_si256(q8_0, 1));
__m128i q8s_2 = _mm_maddubs_epi16(m32s, _mm256_extractf128_si256(q8_1, 0));
__m128i q8s_3 = _mm_maddubs_epi16(m32s, _mm256_extractf128_si256(q8_1, 1));
__m128i p16_0 = _mm_maddubs_epi16(q4_0, _mm256_extractf128_si256(q8_0, 0));
__m128i p16_1 = _mm_maddubs_epi16(q4_1, _mm256_extractf128_si256(q8_0, 1));
__m128i p16_2 = _mm_maddubs_epi16(q4_2, _mm256_extractf128_si256(q8_1, 0));
__m128i p16_3 = _mm_maddubs_epi16(q4_3, _mm256_extractf128_si256(q8_1, 1));
p16_0 = _mm_sub_epi16(p16_0, q8s_0);
p16_1 = _mm_sub_epi16(p16_1, q8s_1);
p16_2 = _mm_sub_epi16(p16_2, q8s_2);
p16_3 = _mm_sub_epi16(p16_3, q8s_3);
p16_0 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_0), p16_0);
p16_1 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_0, scale_0)), p16_1);
p16_2 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_1), p16_2);
p16_3 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_1, scale_1)), p16_3);
sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
ggml : workaround for missing _mm256_setr_m128i in GCC < 8 in k_quants.c (#2405)
2023-07-28 18:17:45 +00:00
acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(MM256_SET_M128I(sumi_1, sumi_0))), acc);
k_quants : add AVX support to dot functions with QK_K as 64 (#2339) * add AVX to ggml_vec_dot_q2_K_q8_K() * add AVX to ggml_vec_dot_q3_K_q8_K() * add AVX to ggml_vec_dot_q4_K_q8_K() * add AVX to ggml_vec_dot_q5_K_q8_K() * add AVX to ggml_vec_dot_q6_K_q8_K() * refactor AVX code in ggml_vec_dot_q6_K_q8_K()
2023-07-25 12:13:41 +00:00
}
*s = hsum_float_8(acc);
ggml : add RISC-V Vector Support for K-Quants and improved the existing intrinsics (#3453) * Added RVV intrinsics support for Q8 quantize row and also improved the existing dot product function for risc-v. The RVV intrinsics is added for the following quantize row functions quantize_row_q8_0 quantize_row_q8_1 The following dot product functions have also been optimized by using LMUL = 1/2 instead of LMUL = 1 ggml_vec_dot_q4_0_q8_0 ggml_vec_dot_q4_1_q8_1 ggml_vec_dot_q5_0_q8_0 ggml_vec_dot_q5_1_q8_1 And vector initialization in Q5 by temporary array is also replaced by the vid intrinsics Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai> * Added RVV intrinsics support for k_quants This adds RISC-V Vector intrinsics support for the following K_quants functions for both QKK = 256 and QKK = 64 ggml_vec_dot_q2_K_q8_K ggml_vec_dot_q3_K_q8_K ggml_vec_dot_q4_K_q8_K ggml_vec_dot_q5_K_q8_K ggml_vec_dot_q6_K_q8_K Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai> --------- Signed-off-by: Ahmad Tameem <ahmad.tameem@10xengineers.ai>
2023-10-03 18:38:19 +00:00
#elif defined __riscv_v_intrinsic
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const float d_all = (float)x[i].d;
const uint8_t * restrict q6 = x[i].ql;
const uint8_t * restrict qh = x[i].qh;
const int8_t * restrict q8 = y[i].qs;
const int8_t * restrict scale = x[i].scales;
int32_t isum = 0;
size_t vl = 16;
vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
// load Q6
vuint8mf2_t q6_0 = __riscv_vle8_v_u8mf2(q6, vl);
vuint8mf2_t q6_1 = __riscv_vle8_v_u8mf2(q6+16, vl);
// load qh
vuint8mf2_t qh_x = __riscv_vle8_v_u8mf2(qh, vl);
vuint8mf2_t qh0 = __riscv_vsll_vx_u8mf2(__riscv_vand_vx_u8mf2(qh_x, 0x3, vl), 0x4, vl);
qh_x = __riscv_vsrl_vx_u8mf2(qh_x, 0x2, vl);
vuint8mf2_t qh1 = __riscv_vsll_vx_u8mf2(__riscv_vand_vx_u8mf2(qh_x, 0x3, vl), 0x4, vl);
qh_x = __riscv_vsrl_vx_u8mf2(qh_x, 0x2, vl);
vuint8mf2_t qh2 = __riscv_vsll_vx_u8mf2(__riscv_vand_vx_u8mf2(qh_x, 0x3, vl), 0x4, vl);
qh_x = __riscv_vsrl_vx_u8mf2(qh_x, 0x2, vl);
vuint8mf2_t qh3 = __riscv_vsll_vx_u8mf2(__riscv_vand_vx_u8mf2(qh_x, 0x3, vl), 0x4, vl);
vuint8mf2_t q6h_0 = __riscv_vor_vv_u8mf2(__riscv_vand_vx_u8mf2(q6_0, 0xF, vl), qh0, vl);
vuint8mf2_t q6h_1 = __riscv_vor_vv_u8mf2(__riscv_vand_vx_u8mf2(q6_1, 0xF, vl), qh1, vl);
vuint8mf2_t q6h_2 = __riscv_vor_vv_u8mf2(__riscv_vsrl_vx_u8mf2(q6_0, 0x4, vl), qh2, vl);
vuint8mf2_t q6h_3 = __riscv_vor_vv_u8mf2(__riscv_vsrl_vx_u8mf2(q6_1, 0x4, vl), qh3, vl);
vint8mf2_t q6v_0 = __riscv_vsub_vx_i8mf2(__riscv_vreinterpret_v_u8mf2_i8mf2(q6h_0), 32, vl);
vint8mf2_t q6v_1 = __riscv_vsub_vx_i8mf2(__riscv_vreinterpret_v_u8mf2_i8mf2(q6h_1), 32, vl);
vint8mf2_t q6v_2 = __riscv_vsub_vx_i8mf2(__riscv_vreinterpret_v_u8mf2_i8mf2(q6h_2), 32, vl);
vint8mf2_t q6v_3 = __riscv_vsub_vx_i8mf2(__riscv_vreinterpret_v_u8mf2_i8mf2(q6h_3), 32, vl);
// load Q8 and take product
vint16m1_t p0 = __riscv_vwmul_vv_i16m1(q6v_0, __riscv_vle8_v_i8mf2(q8, vl), vl);
vint16m1_t p1 = __riscv_vwmul_vv_i16m1(q6v_1, __riscv_vle8_v_i8mf2(q8+16, vl), vl);
vint16m1_t p2 = __riscv_vwmul_vv_i16m1(q6v_2, __riscv_vle8_v_i8mf2(q8+32, vl), vl);
vint16m1_t p3 = __riscv_vwmul_vv_i16m1(q6v_3, __riscv_vle8_v_i8mf2(q8+48, vl), vl);
vint32m1_t vs_0 = __riscv_vwredsum_vs_i16m1_i32m1(p0, vzero, vl);
vint32m1_t vs_1 = __riscv_vwredsum_vs_i16m1_i32m1(p1, vzero, vl);
vint32m1_t vs_2 = __riscv_vwredsum_vs_i16m1_i32m1(p2, vzero, vl);
vint32m1_t vs_3 = __riscv_vwredsum_vs_i16m1_i32m1(p3, vzero, vl);
isum += __riscv_vmv_x_s_i32m1_i32(vs_0) * scale[0];
isum += __riscv_vmv_x_s_i32m1_i32(vs_1) * scale[1];
isum += __riscv_vmv_x_s_i32m1_i32(vs_2) * scale[2];
isum += __riscv_vmv_x_s_i32m1_i32(vs_3) * scale[3];
sumf += isum * d_all * y[i].d;
}
*s = sumf;
k-quants : support for super-block size of 64 (#2001) * k_quants: WIP super-blocks with 64 weights * k_quants: WIP super-blocks with 64 weights Q6_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q4_K scalar and AVX2 works * k_quants: WIP super-blocks with 64 weights Q2_K scalar and AVX2 works. Q2_K is way too slow (it is actually slower than the scalar implementation) * k_quants: WIP super-blocks with 64 weights Q3_K scalar and AVX2 works. * k_quants: WIP super-blocks with 64 weights Q5_K scalar and AVX2 works, and with that all k_quants are done on AVX2 and scalar * k_quants: WIP super-blocks with 64 weights Q6_K working on CUDA. Cannot make it run quite as gast as with super-blocks with 256 weigths: 8% slower on 4080, 20% slower on the 1660 (but there we fit 1 less layer on the GPU because pf the larger model size), so some fraction of these 20% is due to that, * k_quants: WIP super-blocks with 64 weights Q4_K working on CUDA. ~10% slower on GTX-1660, 16% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q2_K working on CUDA. ~3% slower on GTX-1660, 10% slower on 4080. * k_quants: WIP super-blocks with 64 weights Q3_K working on CUDA. * k_quants: WIP super-blocks with 64 weights Q5_K working on CUDA, and with this CUDA is done. * k_quants: WIP super-blocks with 64 weights Q6_K working on ARM_NEON * k_quants: WIP super-blocks with 64 weights Q4_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q2_K working on ARM_NEON, but quite a bit slower than 256 weights * k_quants: WIP super-blocks with 64 weights Q3_K working on ARM_NEON, but quite a bit slower than 256 weights. * k_quants: WIP super-blocks with 64 weights Q5_K working on ARM_NEON, but quite a bit slower than 256 weights. With that, we have full support for ARM_NEON, although performance is not quite there. * k_quants: WIP super-blocks with 64 weights Slightly more efficient Q3_K and Q5_K * k_quants: WIP super-blocks with 64 weights Another small improvement for Q3_K and Q5_K on ARM_NEON * k_quants: WIP super-blocks with 64 weights Yet another speedup for Q5_K on ARM_NEON. We are now within 10% of the QK_K = 256 version. * k_quants: WIP super-blocks with 64 weights * We are able to pass preprocessor macros to the Metal compiler * Q6_K works and is actually slightly more efficient than the QK_K = 256 version (25.2 ms vs 25.8 ms) * k_quants: WIP super-blocks with 64 weights Q4_K works on Metal and is actually slightly faster than QK_K = 256 (21.95 ms vs 24.0 ms). * k_quants: WIP super-blocks with 64 weights Q2_K works on Metal and is very slightly faster than QK_K = 256 (23.8 ms vs 24.2 ms). * k_quants: WIP super-blocks with 64 weights Q3_K works on Metal and is slightly faster than QK_K = 256 (26.6 ms vs 28.3 ms). * k_quants: WIP super-blocks with 64 weights Q5_K works on Metal and is slightly faster than QK_K = 256 (23.7 ms vs 26.3 ms). * k_quants: call them _K, not _k, also on Metal * k_quants: correctly define QK_K in llama.cpp * Fixed bug in q4_K quantization added with the 64-block addition * Simplify via lambda * k_quants: swicth Q3_K to 4-bit scales when QK_K = 64 Otherwise there isn't much benefit from this quantization type. There is some very slight loss in accuracy, but we reduce size by ~7%. E.g., for OpenLLaMA-3B, Q3_K_S perplexity is 8.6131 with 8-bit scales and 8.6352 with 4-bit, while file size decreases from 1.53G to 1.44G. * k_quants: switch Q4_K to 4-bit scales when QK_K = 64 Here the loss in accuracy is greater than for Q3_K, but the Q4_K points still move further to the left on the perplexity vs size curve. * k_quants: forgot to add the Metal changes in last commit * k_quants: change Q5_K to be type 0 when QK_K = 64 Still needs AVX2 implementation * k_quants: AVX2 implementation for new 64-weight Q5_K * k_quants: 10% faster ARM_NEON Q5_K dot product * k_quants: fixed issue caused by merging with master --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-06-26 16:43:07 +00:00
#else
int8_t aux8[QK_K];
int16_t aux16[8];
float sums [8];
int32_t aux32[8];
memset(sums, 0, 8*sizeof(float));
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * restrict q4 = x[i].ql;
const uint8_t * restrict qh = x[i].qh;
const int8_t * restrict q8 = y[i].qs;
memset(aux32, 0, 8*sizeof(int32_t));
int8_t * restrict a = aux8;
for (int l = 0; l < 16; ++l) {
a[l+ 0] = (int8_t)((q4[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
a[l+16] = (int8_t)((q4[l+16] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
a[l+32] = (int8_t)((q4[l+ 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32;
a[l+48] = (int8_t)((q4[l+16] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32;
}
int is = 0;
for (int j = 0; j < QK_K/16; ++j) {
int scale = x[i].scales[is++];
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
q8 += 8; a += 8;
}
const float d = ggml_fp16_to_fp32(x[i].d) * y[i].d;
for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
}
for (int l = 0; l < 8; ++l) sumf += sums[l];
*s = sumf;
#endif
}
#endif
Reference in New Issue Copy Permalink