2024-02-22 21:21:39 +00:00
|
|
|
#include "ggml-cuda.h"
|
|
|
|
#include "ggml.h"
|
|
|
|
#include "ggml-backend-impl.h"
|
|
|
|
|
2023-09-28 19:42:38 +00:00
|
|
|
#include <algorithm>
|
2023-12-13 12:04:25 +00:00
|
|
|
#include <assert.h>
|
|
|
|
#include <atomic>
|
|
|
|
#include <cinttypes>
|
2023-05-01 16:11:07 +00:00
|
|
|
#include <cstddef>
|
|
|
|
#include <cstdint>
|
2023-12-07 20:26:54 +00:00
|
|
|
#include <float.h>
|
2023-06-14 17:47:19 +00:00
|
|
|
#include <limits>
|
2023-04-20 01:14:14 +00:00
|
|
|
#include <stdint.h>
|
2023-04-21 19:59:17 +00:00
|
|
|
#include <stdio.h>
|
2024-01-12 19:07:38 +00:00
|
|
|
#include <string>
|
2023-12-13 12:04:25 +00:00
|
|
|
#include <vector>
|
2024-01-12 19:07:38 +00:00
|
|
|
#include <map>
|
|
|
|
#include <array>
|
2023-04-20 01:14:14 +00:00
|
|
|
|
2024-01-23 12:31:56 +00:00
|
|
|
// stringize macro for converting __CUDA_ARCH_LIST__ (list of integers) to string
|
|
|
|
#define STRINGIZE_IMPL(...) #__VA_ARGS__
|
|
|
|
#define STRINGIZE(...) STRINGIZE_IMPL(__VA_ARGS__)
|
|
|
|
|
2023-08-25 09:09:42 +00:00
|
|
|
#if defined(GGML_USE_HIPBLAS)
|
|
|
|
#include <hip/hip_runtime.h>
|
|
|
|
#include <hipblas/hipblas.h>
|
|
|
|
#include <hip/hip_fp16.h>
|
|
|
|
#ifdef __HIP_PLATFORM_AMD__
|
|
|
|
// for rocblas_initialize()
|
|
|
|
#include "rocblas/rocblas.h"
|
2023-09-13 09:20:24 +00:00
|
|
|
#endif // __HIP_PLATFORM_AMD__
|
2023-09-28 10:08:28 +00:00
|
|
|
#define CUBLAS_COMPUTE_16F HIPBLAS_R_16F
|
2023-08-25 09:09:42 +00:00
|
|
|
#define CUBLAS_COMPUTE_32F HIPBLAS_R_32F
|
|
|
|
#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
|
|
|
|
#define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
|
2023-09-28 10:08:28 +00:00
|
|
|
#define CUBLAS_GEMM_DEFAULT_TENSOR_OP HIPBLAS_GEMM_DEFAULT
|
2023-08-25 09:09:42 +00:00
|
|
|
#define CUBLAS_OP_N HIPBLAS_OP_N
|
|
|
|
#define CUBLAS_OP_T HIPBLAS_OP_T
|
|
|
|
#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
|
|
|
|
#define CUBLAS_TF32_TENSOR_OP_MATH 0
|
|
|
|
#define CUDA_R_16F HIPBLAS_R_16F
|
|
|
|
#define CUDA_R_32F HIPBLAS_R_32F
|
|
|
|
#define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
|
2023-12-18 21:33:45 +00:00
|
|
|
#define cublasComputeType_t hipblasDatatype_t //deprecated, new hipblasComputeType_t not in 5.6
|
2023-08-25 09:09:42 +00:00
|
|
|
#define cublasCreate hipblasCreate
|
|
|
|
#define cublasGemmEx hipblasGemmEx
|
2023-10-24 13:48:37 +00:00
|
|
|
#define cublasGemmBatchedEx hipblasGemmBatchedEx
|
|
|
|
#define cublasGemmStridedBatchedEx hipblasGemmStridedBatchedEx
|
2023-08-25 09:09:42 +00:00
|
|
|
#define cublasHandle_t hipblasHandle_t
|
|
|
|
#define cublasSetMathMode(handle, mode) CUBLAS_STATUS_SUCCESS
|
|
|
|
#define cublasSetStream hipblasSetStream
|
|
|
|
#define cublasSgemm hipblasSgemm
|
|
|
|
#define cublasStatus_t hipblasStatus_t
|
2023-12-18 21:33:45 +00:00
|
|
|
#define cudaDataType_t hipblasDatatype_t //deprecated, new hipblasDatatype not in 5.6
|
2023-09-17 14:37:53 +00:00
|
|
|
#define cudaDeviceCanAccessPeer hipDeviceCanAccessPeer
|
|
|
|
#define cudaDeviceDisablePeerAccess hipDeviceDisablePeerAccess
|
|
|
|
#define cudaDeviceEnablePeerAccess hipDeviceEnablePeerAccess
|
2023-08-25 09:09:42 +00:00
|
|
|
#define cudaDeviceProp hipDeviceProp_t
|
|
|
|
#define cudaDeviceSynchronize hipDeviceSynchronize
|
|
|
|
#define cudaError_t hipError_t
|
2024-02-19 22:40:26 +00:00
|
|
|
#define cudaErrorPeerAccessAlreadyEnabled hipErrorPeerAccessAlreadyEnabled
|
|
|
|
#define cudaErrorPeerAccessNotEnabled hipErrorPeerAccessNotEnabled
|
2023-08-25 09:09:42 +00:00
|
|
|
#define cudaEventCreateWithFlags hipEventCreateWithFlags
|
|
|
|
#define cudaEventDisableTiming hipEventDisableTiming
|
|
|
|
#define cudaEventRecord hipEventRecord
|
|
|
|
#define cudaEvent_t hipEvent_t
|
|
|
|
#define cudaEventDestroy hipEventDestroy
|
|
|
|
#define cudaFree hipFree
|
|
|
|
#define cudaFreeHost hipHostFree
|
|
|
|
#define cudaGetDevice hipGetDevice
|
|
|
|
#define cudaGetDeviceCount hipGetDeviceCount
|
|
|
|
#define cudaGetDeviceProperties hipGetDeviceProperties
|
|
|
|
#define cudaGetErrorString hipGetErrorString
|
|
|
|
#define cudaGetLastError hipGetLastError
|
2023-12-21 19:45:32 +00:00
|
|
|
#ifdef GGML_HIP_UMA
|
|
|
|
#define cudaMalloc hipMallocManaged
|
|
|
|
#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size)
|
|
|
|
#else
|
2023-08-25 09:09:42 +00:00
|
|
|
#define cudaMalloc hipMalloc
|
|
|
|
#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault)
|
2023-12-21 19:45:32 +00:00
|
|
|
#endif
|
2023-08-25 09:09:42 +00:00
|
|
|
#define cudaMemcpy hipMemcpy
|
|
|
|
#define cudaMemcpyAsync hipMemcpyAsync
|
2023-12-26 20:23:59 +00:00
|
|
|
#define cudaMemcpyPeerAsync hipMemcpyPeerAsync
|
|
|
|
#define cudaMemcpy2DAsync hipMemcpy2DAsync
|
2023-08-25 09:09:42 +00:00
|
|
|
#define cudaMemcpyDeviceToDevice hipMemcpyDeviceToDevice
|
|
|
|
#define cudaMemcpyDeviceToHost hipMemcpyDeviceToHost
|
|
|
|
#define cudaMemcpyHostToDevice hipMemcpyHostToDevice
|
|
|
|
#define cudaMemcpyKind hipMemcpyKind
|
|
|
|
#define cudaMemset hipMemset
|
2023-10-08 17:19:14 +00:00
|
|
|
#define cudaMemsetAsync hipMemsetAsync
|
2024-01-12 19:07:38 +00:00
|
|
|
#define cudaMemGetInfo hipMemGetInfo
|
2023-08-25 09:09:42 +00:00
|
|
|
#define cudaOccupancyMaxPotentialBlockSize hipOccupancyMaxPotentialBlockSize
|
|
|
|
#define cudaSetDevice hipSetDevice
|
|
|
|
#define cudaStreamCreateWithFlags hipStreamCreateWithFlags
|
2023-12-07 20:26:54 +00:00
|
|
|
#define cudaStreamFireAndForget hipStreamFireAndForget
|
2023-08-25 09:09:42 +00:00
|
|
|
#define cudaStreamNonBlocking hipStreamNonBlocking
|
|
|
|
#define cudaStreamSynchronize hipStreamSynchronize
|
2023-09-16 14:55:43 +00:00
|
|
|
#define cudaStreamWaitEvent(stream, event, flags) hipStreamWaitEvent(stream, event, flags)
|
2023-08-25 09:09:42 +00:00
|
|
|
#define cudaStream_t hipStream_t
|
|
|
|
#define cudaSuccess hipSuccess
|
2023-12-21 19:13:25 +00:00
|
|
|
#define __trap abort
|
2023-12-24 13:34:22 +00:00
|
|
|
#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
|
|
|
|
#define CUBLAS_STATUS_NOT_INITIALIZED HIPBLAS_STATUS_NOT_INITIALIZED
|
|
|
|
#define CUBLAS_STATUS_ALLOC_FAILED HIPBLAS_STATUS_ALLOC_FAILED
|
|
|
|
#define CUBLAS_STATUS_INVALID_VALUE HIPBLAS_STATUS_INVALID_VALUE
|
|
|
|
#define CUBLAS_STATUS_ARCH_MISMATCH HIPBLAS_STATUS_ARCH_MISMATCH
|
|
|
|
#define CUBLAS_STATUS_MAPPING_ERROR HIPBLAS_STATUS_MAPPING_ERROR
|
|
|
|
#define CUBLAS_STATUS_EXECUTION_FAILED HIPBLAS_STATUS_EXECUTION_FAILED
|
|
|
|
#define CUBLAS_STATUS_INTERNAL_ERROR HIPBLAS_STATUS_INTERNAL_ERROR
|
|
|
|
#define CUBLAS_STATUS_NOT_SUPPORTED HIPBLAS_STATUS_NOT_SUPPORTED
|
2023-08-25 09:09:42 +00:00
|
|
|
#else
|
2023-05-01 16:11:07 +00:00
|
|
|
#include <cuda_runtime.h>
|
2023-12-24 13:34:22 +00:00
|
|
|
#include <cuda.h>
|
2023-05-01 16:11:07 +00:00
|
|
|
#include <cublas_v2.h>
|
|
|
|
#include <cuda_fp16.h>
|
2023-12-24 13:34:22 +00:00
|
|
|
|
|
|
|
#if CUDART_VERSION < 11020
|
2023-12-26 10:38:36 +00:00
|
|
|
#define CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED CU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED
|
2023-12-22 15:11:12 +00:00
|
|
|
#define CUBLAS_TF32_TENSOR_OP_MATH CUBLAS_TENSOR_OP_MATH
|
|
|
|
#define CUBLAS_COMPUTE_16F CUDA_R_16F
|
|
|
|
#define CUBLAS_COMPUTE_32F CUDA_R_32F
|
|
|
|
#define cublasComputeType_t cudaDataType_t
|
2023-12-24 13:34:22 +00:00
|
|
|
#endif // CUDART_VERSION < 11020
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
#endif // defined(GGML_USE_HIPBLAS)
|
2023-05-01 16:11:07 +00:00
|
|
|
|
2024-01-12 11:30:41 +00:00
|
|
|
#define CUDART_HMAX 11070 // CUDA 11.7, min. ver. for which __hmax and __hmax2 are known to work (may be higher than needed)
|
|
|
|
|
2024-01-09 07:58:55 +00:00
|
|
|
#define CC_PASCAL 600
|
2023-09-13 09:20:24 +00:00
|
|
|
#define MIN_CC_DP4A 610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products
|
2023-09-30 16:12:57 +00:00
|
|
|
#define CC_VOLTA 700
|
2023-09-13 09:20:24 +00:00
|
|
|
#define CC_OFFSET_AMD 1000000
|
2023-12-30 12:52:01 +00:00
|
|
|
#define CC_RDNA1 (CC_OFFSET_AMD + 1010)
|
2023-09-30 16:12:57 +00:00
|
|
|
#define CC_RDNA2 (CC_OFFSET_AMD + 1030)
|
2023-12-30 12:52:01 +00:00
|
|
|
#define CC_RDNA3 (CC_OFFSET_AMD + 1100)
|
2023-08-25 09:09:42 +00:00
|
|
|
|
2023-11-15 12:58:13 +00:00
|
|
|
#define GGML_CUDA_MAX_NODES 8192
|
|
|
|
|
2023-12-30 12:52:01 +00:00
|
|
|
// define this if you want to always fallback to MMQ kernels and not use cuBLAS for matrix multiplication
|
|
|
|
// on modern hardware, using cuBLAS is recommended as it utilizes F16 tensor cores which are very performant
|
|
|
|
// for large computational tasks. the drawback is that this requires some extra amount of VRAM:
|
|
|
|
// - 7B quantum model: +100-200 MB
|
|
|
|
// - 13B quantum model: +200-400 MB
|
|
|
|
//
|
|
|
|
//#define GGML_CUDA_FORCE_MMQ
|
|
|
|
|
|
|
|
// TODO: improve this to be correct for more hardware
|
|
|
|
// for example, currently fails for GeForce GTX 1660 which is TURING arch (> VOLTA) but does not have tensor cores
|
|
|
|
#if !defined(GGML_CUDA_FORCE_MMQ)
|
|
|
|
#define CUDA_USE_TENSOR_CORES
|
|
|
|
#endif
|
|
|
|
|
2024-02-11 18:08:39 +00:00
|
|
|
#define MMVQ_MAX_BATCH_SIZE 8 // max batch size to use MMVQ kernels
|
|
|
|
#define MMQ_MAX_BATCH_SIZE 32 // max batch size to use MMQ kernels when tensor cores are available
|
2023-12-30 12:52:01 +00:00
|
|
|
|
2023-08-25 09:09:42 +00:00
|
|
|
#if defined(GGML_USE_HIPBLAS)
|
|
|
|
#define __CUDA_ARCH__ 1300
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \
|
|
|
|
defined(__gfx1150__) || defined(__gfx1151__)
|
|
|
|
#define RDNA3
|
|
|
|
#endif
|
|
|
|
|
|
|
|
#if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || defined(__gfx1033__) || \
|
|
|
|
defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) || defined(__gfx1037__)
|
|
|
|
#define RDNA2
|
|
|
|
#endif
|
|
|
|
|
2023-09-01 21:33:19 +00:00
|
|
|
#ifndef __has_builtin
|
|
|
|
#define __has_builtin(x) 0
|
|
|
|
#endif
|
|
|
|
|
2023-08-25 09:09:42 +00:00
|
|
|
typedef int8_t int8x4_t __attribute__((ext_vector_type(4)));
|
2024-02-24 14:23:52 +00:00
|
|
|
typedef uint8_t uint8x4_t __attribute__((ext_vector_type(4)));
|
2023-08-25 09:09:42 +00:00
|
|
|
static __device__ __forceinline__ int __vsubss4(const int a, const int b) {
|
|
|
|
const int8x4_t va = reinterpret_cast<const int8x4_t&>(a);
|
|
|
|
const int8x4_t vb = reinterpret_cast<const int8x4_t&>(b);
|
2023-09-01 21:33:19 +00:00
|
|
|
#if __has_builtin(__builtin_elementwise_sub_sat)
|
2023-08-25 09:09:42 +00:00
|
|
|
const int8x4_t c = __builtin_elementwise_sub_sat(va, vb);
|
2023-12-26 20:23:59 +00:00
|
|
|
return reinterpret_cast<const int &>(c);
|
2023-09-01 21:33:19 +00:00
|
|
|
#else
|
|
|
|
int8x4_t c;
|
|
|
|
int16_t tmp;
|
|
|
|
#pragma unroll
|
|
|
|
for (int i = 0; i < 4; i++) {
|
|
|
|
tmp = va[i] - vb[i];
|
|
|
|
if(tmp > std::numeric_limits<int8_t>::max()) tmp = std::numeric_limits<int8_t>::max();
|
|
|
|
if(tmp < std::numeric_limits<int8_t>::min()) tmp = std::numeric_limits<int8_t>::min();
|
|
|
|
c[i] = tmp;
|
|
|
|
}
|
2023-12-26 20:23:59 +00:00
|
|
|
return reinterpret_cast<int &>(c);
|
2023-09-01 21:33:19 +00:00
|
|
|
#endif // __has_builtin(__builtin_elementwise_sub_sat)
|
2023-08-25 09:09:42 +00:00
|
|
|
}
|
|
|
|
|
2024-01-30 13:14:12 +00:00
|
|
|
static __device__ __forceinline__ int __vsub4(const int a, const int b) {
|
|
|
|
return __vsubss4(a, b);
|
|
|
|
}
|
|
|
|
|
2024-02-24 14:23:52 +00:00
|
|
|
static __device__ __forceinline__ unsigned int __vcmpeq4(unsigned int a, unsigned int b) {
|
|
|
|
const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
|
|
|
|
const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
|
|
|
|
unsigned int c;
|
|
|
|
uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
|
|
|
|
#pragma unroll
|
|
|
|
for (int i = 0; i < 4; ++i) {
|
|
|
|
vc[i] = va[i] == vb[i] ? 0xff : 0x00;
|
|
|
|
}
|
|
|
|
return c;
|
|
|
|
}
|
|
|
|
|
2023-08-25 09:09:42 +00:00
|
|
|
static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) {
|
|
|
|
#if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx1030__)
|
|
|
|
c = __builtin_amdgcn_sdot4(a, b, c, false);
|
2024-01-07 06:52:42 +00:00
|
|
|
#elif defined(RDNA3)
|
2023-08-25 09:09:42 +00:00
|
|
|
c = __builtin_amdgcn_sudot4( true, a, true, b, c, false);
|
|
|
|
#elif defined(__gfx1010__) || defined(__gfx900__)
|
|
|
|
int tmp1;
|
|
|
|
int tmp2;
|
|
|
|
asm("\n \
|
|
|
|
v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0 \n \
|
|
|
|
v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1 \n \
|
|
|
|
v_add3_u32 %0, %1, %2, %0 \n \
|
|
|
|
v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2 \n \
|
|
|
|
v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3 \n \
|
|
|
|
v_add3_u32 %0, %1, %2, %0 \n \
|
|
|
|
"
|
|
|
|
: "+v"(c), "=&v"(tmp1), "=&v"(tmp2)
|
|
|
|
: "v"(a), "v"(b)
|
|
|
|
);
|
|
|
|
#else
|
|
|
|
const int8x4_t va = reinterpret_cast<const int8x4_t&>(a);
|
|
|
|
const int8x4_t vb = reinterpret_cast<const int8x4_t&>(b);
|
|
|
|
c += va[0] * vb[0] + va[1] * vb[1] + va[2] * vb[2] + va[3] * vb[3];
|
|
|
|
#endif
|
|
|
|
return c;
|
|
|
|
}
|
2023-09-13 09:20:24 +00:00
|
|
|
#endif // defined(GGML_USE_HIPBLAS)
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-06-17 15:46:15 +00:00
|
|
|
#if defined(_MSC_VER)
|
|
|
|
#pragma warning(disable: 4244 4267) // possible loss of data
|
|
|
|
#endif
|
|
|
|
|
2023-05-01 16:11:07 +00:00
|
|
|
static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size");
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
[[noreturn]]
|
|
|
|
static void ggml_cuda_error(const char * stmt, const char * func, const char * file, const int line, const char * msg) {
|
|
|
|
int id = -1; // in case cudaGetDevice fails
|
|
|
|
cudaGetDevice(&id);
|
|
|
|
|
|
|
|
fprintf(stderr, "CUDA error: %s\n", msg);
|
|
|
|
fprintf(stderr, " current device: %d, in function %s at %s:%d\n", id, func, file, line);
|
|
|
|
fprintf(stderr, " %s\n", stmt);
|
|
|
|
// abort with GGML_ASSERT to get a stack trace
|
|
|
|
GGML_ASSERT(!"CUDA error");
|
|
|
|
}
|
|
|
|
|
|
|
|
#define CUDA_CHECK_GEN(err, success, error_fn) \
|
|
|
|
do { \
|
|
|
|
auto err_ = (err); \
|
|
|
|
if (err_ != (success)) { \
|
|
|
|
ggml_cuda_error(#err, __func__, __FILE__, __LINE__, error_fn(err_)); \
|
|
|
|
} \
|
|
|
|
} while (0)
|
|
|
|
|
|
|
|
#define CUDA_CHECK(err) CUDA_CHECK_GEN(err, cudaSuccess, cudaGetErrorString)
|
|
|
|
|
2023-06-15 19:49:08 +00:00
|
|
|
#if CUDART_VERSION >= 12000
|
2023-12-24 13:34:22 +00:00
|
|
|
static const char * cublas_get_error_str(const cublasStatus_t err) {
|
|
|
|
return cublasGetStatusString(err);
|
|
|
|
}
|
2023-06-06 19:33:23 +00:00
|
|
|
#else
|
2023-12-24 13:34:22 +00:00
|
|
|
static const char * cublas_get_error_str(const cublasStatus_t err) {
|
|
|
|
switch (err) {
|
|
|
|
case CUBLAS_STATUS_SUCCESS: return "CUBLAS_STATUS_SUCCESS";
|
|
|
|
case CUBLAS_STATUS_NOT_INITIALIZED: return "CUBLAS_STATUS_NOT_INITIALIZED";
|
|
|
|
case CUBLAS_STATUS_ALLOC_FAILED: return "CUBLAS_STATUS_ALLOC_FAILED";
|
|
|
|
case CUBLAS_STATUS_INVALID_VALUE: return "CUBLAS_STATUS_INVALID_VALUE";
|
|
|
|
case CUBLAS_STATUS_ARCH_MISMATCH: return "CUBLAS_STATUS_ARCH_MISMATCH";
|
|
|
|
case CUBLAS_STATUS_MAPPING_ERROR: return "CUBLAS_STATUS_MAPPING_ERROR";
|
|
|
|
case CUBLAS_STATUS_EXECUTION_FAILED: return "CUBLAS_STATUS_EXECUTION_FAILED";
|
|
|
|
case CUBLAS_STATUS_INTERNAL_ERROR: return "CUBLAS_STATUS_INTERNAL_ERROR";
|
|
|
|
case CUBLAS_STATUS_NOT_SUPPORTED: return "CUBLAS_STATUS_NOT_SUPPORTED";
|
|
|
|
default: return "unknown error";
|
|
|
|
}
|
|
|
|
}
|
|
|
|
#endif // CUDART_VERSION >= 12000
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
#define CUBLAS_CHECK(err) CUDA_CHECK_GEN(err, CUBLAS_STATUS_SUCCESS, cublas_get_error_str)
|
2023-12-24 13:34:22 +00:00
|
|
|
|
|
|
|
#if !defined(GGML_USE_HIPBLAS)
|
|
|
|
static const char * cu_get_error_str(CUresult err) {
|
|
|
|
const char * err_str;
|
|
|
|
cuGetErrorString(err, &err_str);
|
|
|
|
return err_str;
|
|
|
|
}
|
2023-12-26 20:23:59 +00:00
|
|
|
#define CU_CHECK(err) CUDA_CHECK_GEN(err, CUDA_SUCCESS, cu_get_error_str)
|
2023-12-24 13:34:22 +00:00
|
|
|
#endif
|
2023-05-01 16:11:07 +00:00
|
|
|
|
2023-09-16 14:55:43 +00:00
|
|
|
#if CUDART_VERSION >= 11100
|
|
|
|
#define GGML_CUDA_ASSUME(x) __builtin_assume(x)
|
|
|
|
#else
|
|
|
|
#define GGML_CUDA_ASSUME(x)
|
|
|
|
#endif // CUDART_VERSION >= 11100
|
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
#ifdef GGML_CUDA_F16
|
2023-06-19 08:23:56 +00:00
|
|
|
typedef half dfloat; // dequantize float
|
|
|
|
typedef half2 dfloat2;
|
|
|
|
#else
|
|
|
|
typedef float dfloat; // dequantize float
|
|
|
|
typedef float2 dfloat2;
|
2023-07-29 21:04:44 +00:00
|
|
|
#endif //GGML_CUDA_F16
|
|
|
|
|
|
|
|
static __device__ __forceinline__ int get_int_from_int8(const int8_t * x8, const int & i32) {
|
2023-11-18 15:11:18 +00:00
|
|
|
const uint16_t * x16 = (const uint16_t *) (x8 + sizeof(int) * i32); // assume at least 2 byte alignment
|
2023-07-29 21:04:44 +00:00
|
|
|
|
|
|
|
int x32 = 0;
|
|
|
|
x32 |= x16[0] << 0;
|
|
|
|
x32 |= x16[1] << 16;
|
|
|
|
|
|
|
|
return x32;
|
|
|
|
}
|
|
|
|
|
|
|
|
static __device__ __forceinline__ int get_int_from_uint8(const uint8_t * x8, const int & i32) {
|
2023-11-18 15:11:18 +00:00
|
|
|
const uint16_t * x16 = (const uint16_t *) (x8 + sizeof(int) * i32); // assume at least 2 byte alignment
|
2023-07-29 21:04:44 +00:00
|
|
|
|
|
|
|
int x32 = 0;
|
|
|
|
x32 |= x16[0] << 0;
|
|
|
|
x32 |= x16[1] << 16;
|
|
|
|
|
|
|
|
return x32;
|
|
|
|
}
|
|
|
|
|
|
|
|
static __device__ __forceinline__ int get_int_from_int8_aligned(const int8_t * x8, const int & i32) {
|
2023-11-18 15:11:18 +00:00
|
|
|
return *((const int *) (x8 + sizeof(int) * i32)); // assume at least 4 byte alignment
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static __device__ __forceinline__ int get_int_from_uint8_aligned(const uint8_t * x8, const int & i32) {
|
2023-11-18 15:11:18 +00:00
|
|
|
return *((const int *) (x8 + sizeof(int) * i32)); // assume at least 4 byte alignment
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
2023-06-19 08:23:56 +00:00
|
|
|
|
2023-09-28 10:08:28 +00:00
|
|
|
template<typename T>
|
|
|
|
using to_t_cuda_t = void (*)(const void * __restrict__ x, T * __restrict__ y, int k, cudaStream_t stream);
|
|
|
|
typedef to_t_cuda_t<float> to_fp32_cuda_t;
|
|
|
|
typedef to_t_cuda_t<half> to_fp16_cuda_t;
|
|
|
|
|
2023-06-19 08:23:56 +00:00
|
|
|
typedef void (*dequantize_kernel_t)(const void * vx, const int ib, const int iqs, dfloat2 & v);
|
2023-07-07 22:25:15 +00:00
|
|
|
typedef void (*dot_kernel_k_t)(const void * __restrict__ vx, const int ib, const int iqs, const float * __restrict__ y, float & v);
|
2023-06-14 17:47:19 +00:00
|
|
|
typedef void (*cpy_kernel_t)(const char * cx, char * cdst);
|
2023-06-06 19:33:23 +00:00
|
|
|
typedef void (*ggml_cuda_func_t)(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
|
2023-09-11 17:55:51 +00:00
|
|
|
typedef void (*ggml_cuda_op_mul_mat_t)(
|
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
|
|
|
|
const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
|
2023-12-26 20:23:59 +00:00
|
|
|
const int64_t src1_padded_row_size, cudaStream_t stream);
|
2023-09-11 17:55:51 +00:00
|
|
|
typedef void (*ggml_cuda_op_flatten_t)(
|
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream);
|
2023-05-13 13:38:36 +00:00
|
|
|
|
|
|
|
// QK = number of values after dequantization
|
|
|
|
// QR = QK / number of values before dequantization
|
2023-07-05 12:19:42 +00:00
|
|
|
// QI = number of 32 bit integers before dequantization
|
2023-04-20 01:14:14 +00:00
|
|
|
|
|
|
|
#define QK4_0 32
|
2023-05-13 13:38:36 +00:00
|
|
|
#define QR4_0 2
|
2023-07-14 17:44:08 +00:00
|
|
|
#define QI4_0 (QK4_0 / (4 * QR4_0))
|
2023-04-20 01:14:14 +00:00
|
|
|
typedef struct {
|
2023-05-19 19:17:18 +00:00
|
|
|
half d; // delta
|
2023-04-20 01:14:14 +00:00
|
|
|
uint8_t qs[QK4_0 / 2]; // nibbles / quants
|
|
|
|
} block_q4_0;
|
2023-05-19 19:17:18 +00:00
|
|
|
static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
|
2023-04-20 01:14:14 +00:00
|
|
|
|
|
|
|
#define QK4_1 32
|
2023-05-13 13:38:36 +00:00
|
|
|
#define QR4_1 2
|
2023-07-14 17:44:08 +00:00
|
|
|
#define QI4_1 (QK4_1 / (4 * QR4_1))
|
2023-04-20 01:14:14 +00:00
|
|
|
typedef struct {
|
2023-07-29 21:04:44 +00:00
|
|
|
half2 dm; // dm.x = delta, dm.y = min
|
2023-04-20 01:14:14 +00:00
|
|
|
uint8_t qs[QK4_1 / 2]; // nibbles / quants
|
|
|
|
} block_q4_1;
|
2023-05-19 19:17:18 +00:00
|
|
|
static_assert(sizeof(block_q4_1) == sizeof(ggml_fp16_t) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");
|
2023-04-20 01:14:14 +00:00
|
|
|
|
2023-04-26 20:14:13 +00:00
|
|
|
#define QK5_0 32
|
2023-05-13 13:38:36 +00:00
|
|
|
#define QR5_0 2
|
2023-07-14 17:44:08 +00:00
|
|
|
#define QI5_0 (QK5_0 / (4 * QR5_0))
|
2023-04-26 20:14:13 +00:00
|
|
|
typedef struct {
|
2023-05-01 16:11:07 +00:00
|
|
|
half d; // delta
|
2023-04-26 20:14:13 +00:00
|
|
|
uint8_t qh[4]; // 5-th bit of quants
|
|
|
|
uint8_t qs[QK5_0 / 2]; // nibbles / quants
|
|
|
|
} block_q5_0;
|
|
|
|
static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
|
|
|
|
|
|
|
|
#define QK5_1 32
|
2023-05-13 13:38:36 +00:00
|
|
|
#define QR5_1 2
|
2023-07-14 17:44:08 +00:00
|
|
|
#define QI5_1 (QK5_1 / (4 * QR5_1))
|
2023-04-26 20:14:13 +00:00
|
|
|
typedef struct {
|
2023-07-29 21:04:44 +00:00
|
|
|
half2 dm; // dm.x = delta, dm.y = min
|
2023-05-01 16:11:07 +00:00
|
|
|
uint8_t qh[4]; // 5-th bit of quants
|
2023-04-26 20:14:13 +00:00
|
|
|
uint8_t qs[QK5_1 / 2]; // nibbles / quants
|
|
|
|
} block_q5_1;
|
|
|
|
static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
|
|
|
|
|
2023-04-25 20:40:51 +00:00
|
|
|
#define QK8_0 32
|
2023-05-13 13:38:36 +00:00
|
|
|
#define QR8_0 1
|
2023-07-14 17:44:08 +00:00
|
|
|
#define QI8_0 (QK8_0 / (4 * QR8_0))
|
2023-04-25 20:40:51 +00:00
|
|
|
typedef struct {
|
2023-05-19 19:17:18 +00:00
|
|
|
half d; // delta
|
2023-04-25 20:40:51 +00:00
|
|
|
int8_t qs[QK8_0]; // quants
|
|
|
|
} block_q8_0;
|
2023-05-19 19:17:18 +00:00
|
|
|
static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
|
2023-04-25 20:40:51 +00:00
|
|
|
|
2023-07-05 12:19:42 +00:00
|
|
|
#define QK8_1 32
|
|
|
|
#define QR8_1 1
|
2023-07-14 17:44:08 +00:00
|
|
|
#define QI8_1 (QK8_1 / (4 * QR8_1))
|
2023-07-05 12:19:42 +00:00
|
|
|
typedef struct {
|
2023-07-29 21:04:44 +00:00
|
|
|
half2 ds; // ds.x = delta, ds.y = sum
|
2023-07-05 12:19:42 +00:00
|
|
|
int8_t qs[QK8_0]; // quants
|
|
|
|
} block_q8_1;
|
|
|
|
static_assert(sizeof(block_q8_1) == 2*sizeof(ggml_fp16_t) + QK8_0, "wrong q8_1 block size/padding");
|
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
typedef float (*vec_dot_q_cuda_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs);
|
|
|
|
typedef void (*allocate_tiles_cuda_t)(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc);
|
|
|
|
typedef void (*load_tiles_cuda_t)(
|
|
|
|
const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
|
2023-08-02 14:48:10 +00:00
|
|
|
int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row);
|
2023-07-29 21:04:44 +00:00
|
|
|
typedef float (*vec_dot_q_mul_mat_cuda_t)(
|
|
|
|
const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
|
|
|
|
const int * __restrict__ y_qs, const half2 * __restrict__ y_ms, const int & i, const int & j, const int & k);
|
2023-07-05 12:19:42 +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
|
|
|
|
|
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
|
|
|
#ifdef GGML_QKK_64
|
|
|
|
#define QK_K 64
|
|
|
|
#define K_SCALE_SIZE 4
|
|
|
|
#else
|
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
|
|
|
#define QK_K 256
|
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
|
|
|
#define K_SCALE_SIZE 12
|
|
|
|
#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
|
|
|
|
2023-07-14 17:44:08 +00:00
|
|
|
#define QR2_K 4
|
|
|
|
#define QI2_K (QK_K / (4*QR2_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
|
|
|
typedef struct {
|
|
|
|
uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
|
|
|
|
uint8_t qs[QK_K/4]; // quants
|
2023-07-29 21:04:44 +00:00
|
|
|
half2 dm; // super-block scale for quantized scales/mins
|
2023-06-07 07:59:52 +00:00
|
|
|
} block_q2_K;
|
|
|
|
static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
|
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
|
|
|
|
2023-07-14 17:44:08 +00:00
|
|
|
#define QR3_K 4
|
|
|
|
#define QI3_K (QK_K / (4*QR3_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
|
|
|
typedef struct {
|
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
|
|
|
uint8_t hmask[QK_K/8]; // quants - high bit
|
|
|
|
uint8_t qs[QK_K/4]; // quants - low 2 bits
|
|
|
|
#ifdef GGML_QKK_64
|
|
|
|
uint8_t scales[2]; // scales, quantized with 8 bits
|
|
|
|
#else
|
|
|
|
uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
|
|
|
|
#endif
|
|
|
|
half d; // super-block scale
|
2023-06-07 07:59:52 +00:00
|
|
|
} block_q3_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
|
|
|
//static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + K_SCALE_SIZE, "wrong q3_K block size/padding");
|
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
|
|
|
|
2023-07-14 17:44:08 +00:00
|
|
|
#define QR4_K 2
|
|
|
|
#define QI4_K (QK_K / (4*QR4_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
|
|
|
#ifdef GGML_QKK_64
|
|
|
|
typedef struct {
|
2023-08-27 12:19:59 +00:00
|
|
|
half dm[2]; // super-block scales/mins
|
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
|
|
|
uint8_t scales[2]; // 4-bit block scales/mins
|
|
|
|
uint8_t qs[QK_K/2]; // 4--bit quants
|
|
|
|
} block_q4_K;
|
2023-08-27 12:19:59 +00:00
|
|
|
static_assert(sizeof(block_q4_K) == sizeof(half2) + QK_K/2 + 2, "wrong q4_K block size/padding");
|
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 : 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
|
|
|
typedef struct {
|
2023-07-29 21:04:44 +00:00
|
|
|
half2 dm; // super-block scale for quantized scales/mins
|
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
|
|
|
uint8_t scales[3*QK_K/64]; // scales, quantized with 6 bits
|
|
|
|
uint8_t qs[QK_K/2]; // 4--bit quants
|
2023-06-07 07:59:52 +00:00
|
|
|
} block_q4_K;
|
|
|
|
static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + 3*QK_K/64 + QK_K/2, "wrong q4_K block size/padding");
|
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
|
|
|
|
2023-07-14 17:44:08 +00:00
|
|
|
#define QR5_K 2
|
|
|
|
#define QI5_K (QK_K / (4*QR5_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
|
|
|
#ifdef GGML_QKK_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
|
|
|
typedef struct {
|
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
|
|
|
half d; // super-block scale
|
|
|
|
int8_t scales[QK_K/16]; // block scales
|
|
|
|
uint8_t qh[QK_K/8]; // quants, high bit
|
|
|
|
uint8_t qs[QK_K/2]; // quants, low 4 bits
|
|
|
|
} block_q5_K;
|
|
|
|
static_assert(sizeof(block_q5_K) == sizeof(ggml_fp16_t) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
|
|
|
|
#else
|
|
|
|
typedef struct {
|
2023-07-29 21:04:44 +00:00
|
|
|
half2 dm; // super-block scale for quantized scales/mins
|
|
|
|
uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
|
|
|
|
uint8_t qh[QK_K/8]; // quants, high bit
|
|
|
|
uint8_t qs[QK_K/2]; // quants, low 4 bits
|
2023-06-07 07:59:52 +00:00
|
|
|
} block_q5_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
|
|
|
static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
|
|
|
|
#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
|
|
|
|
2023-07-14 17:44:08 +00:00
|
|
|
#define QR6_K 2
|
|
|
|
#define QI6_K (QK_K / (4*QR6_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
|
|
|
typedef struct {
|
|
|
|
uint8_t ql[QK_K/2]; // quants, lower 4 bits
|
|
|
|
uint8_t qh[QK_K/4]; // quants, upper 2 bits
|
|
|
|
int8_t scales[QK_K/16]; // scales
|
|
|
|
half d; // delta
|
2023-06-07 07:59:52 +00:00
|
|
|
} block_q6_K;
|
|
|
|
static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_K block size/padding");
|
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
|
|
|
|
2024-01-08 15:02:32 +00:00
|
|
|
#define QR2_XXS 8
|
|
|
|
#define QI2_XXS (QK_K / (4*QR2_XXS))
|
|
|
|
typedef struct {
|
|
|
|
half d;
|
|
|
|
uint16_t qs[QK_K/8];
|
|
|
|
} block_iq2_xxs;
|
|
|
|
static_assert(sizeof(block_iq2_xxs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t), "wrong iq2_xxs block size/padding");
|
|
|
|
|
2024-01-11 19:39:39 +00:00
|
|
|
#define QR2_XS 8
|
|
|
|
#define QI2_XS (QK_K / (4*QR2_XS))
|
|
|
|
typedef struct {
|
|
|
|
half d;
|
|
|
|
uint16_t qs[QK_K/8];
|
|
|
|
uint8_t scales[QK_K/32];
|
|
|
|
} block_iq2_xs;
|
|
|
|
static_assert(sizeof(block_iq2_xs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding");
|
|
|
|
|
2024-02-26 16:28:38 +00:00
|
|
|
// 2.5625 bpw quants
|
|
|
|
#define QR2_S 8
|
|
|
|
#define QI2_S (QK_K / (4*QR2_S))
|
|
|
|
typedef struct {
|
|
|
|
half d;
|
|
|
|
uint8_t qs[QK_K/4];
|
|
|
|
uint8_t qh[QK_K/32];
|
|
|
|
uint8_t scales[QK_K/32];
|
|
|
|
} block_iq2_s;
|
|
|
|
static_assert(sizeof(block_iq2_s) == sizeof(ggml_fp16_t) + QK_K/4 + QK_K/16, "wrong iq2_s block size/padding");
|
|
|
|
|
2024-01-30 13:14:12 +00:00
|
|
|
#define QR3_XXS 8
|
|
|
|
#define QI3_XXS (QK_K / (4*QR3_XXS))
|
|
|
|
typedef struct {
|
|
|
|
half d;
|
|
|
|
uint8_t qs[3*(QK_K/8)];
|
|
|
|
} block_iq3_xxs;
|
|
|
|
static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_fp16_t) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
|
|
|
|
|
2024-02-24 14:23:52 +00:00
|
|
|
#define QR3_XS 8
|
|
|
|
#define QI3_XS (QK_K / (4*QR3_XS))
|
2024-02-28 08:37:02 +00:00
|
|
|
#if QK_K == 64
|
|
|
|
#define IQ3S_N_SCALE 2
|
|
|
|
#else
|
|
|
|
#define IQ3S_N_SCALE QK_K/64
|
|
|
|
#endif
|
2024-02-24 14:23:52 +00:00
|
|
|
typedef struct {
|
|
|
|
half d;
|
|
|
|
uint8_t qs[QK_K/4];
|
|
|
|
uint8_t qh[QK_K/32];
|
|
|
|
uint8_t signs[QK_K/8];
|
2024-02-28 08:37:02 +00:00
|
|
|
uint8_t scales[IQ3S_N_SCALE];
|
2024-02-24 14:23:52 +00:00
|
|
|
} block_iq3_s;
|
2024-02-28 08:37:02 +00:00
|
|
|
static_assert(sizeof(block_iq3_s) == sizeof(ggml_fp16_t) + 13*(QK_K/32) + IQ3S_N_SCALE, "wrong iq3_s block size/padding");
|
2024-02-24 14:23:52 +00:00
|
|
|
|
2024-02-18 16:16:55 +00:00
|
|
|
#define QR1_S 8
|
|
|
|
#define QI1_S (QK_K / (4*QR1_S))
|
|
|
|
typedef struct {
|
|
|
|
half d;
|
|
|
|
uint8_t qs[QK_K/8];
|
|
|
|
uint8_t scales[QK_K/16];
|
|
|
|
} block_iq1_s;
|
|
|
|
static_assert(sizeof(block_iq1_s) == sizeof(ggml_fp16_t) + QK_K/8 + QK_K/16, "wrong iq1_s block size/padding");
|
|
|
|
|
2024-02-21 09:39:52 +00:00
|
|
|
#define QK4_NL 32
|
|
|
|
#define QR4_NL 2
|
|
|
|
#define QI4_NL (QK4_NL / (4*QR4_NL))
|
|
|
|
typedef struct {
|
|
|
|
half d;
|
|
|
|
uint8_t qs[QK4_NL/2];
|
|
|
|
} block_iq4_nl;
|
|
|
|
static_assert(sizeof(block_iq4_nl) == sizeof(ggml_fp16_t) + QK4_NL/2, "wrong iq4_nl block size/padding");
|
|
|
|
|
2024-02-28 08:37:02 +00:00
|
|
|
#if QK_K == 64
|
|
|
|
#define block_iq4_xs block_iq4_nl
|
|
|
|
#define QR4_XS QR4_NL
|
|
|
|
#define QI4_XS QI4_NL
|
|
|
|
#else
|
2024-02-27 14:34:24 +00:00
|
|
|
// QR4_XS = 8 is very slightly faster than QR4_XS = 4
|
|
|
|
#define QR4_XS 8
|
|
|
|
#define QI4_XS (QK_K / (4*QR4_XS))
|
|
|
|
typedef struct {
|
|
|
|
half d;
|
|
|
|
uint16_t scales_h;
|
|
|
|
uint8_t scales_l[QK_K/64];
|
|
|
|
uint8_t qs[QK_K/2];
|
|
|
|
} block_iq4_xs;
|
|
|
|
static_assert(sizeof(block_iq4_xs) == sizeof(ggml_fp16_t) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding");
|
2024-02-28 08:37:02 +00:00
|
|
|
#endif
|
2024-02-27 14:34:24 +00:00
|
|
|
|
2023-05-25 21:07:29 +00:00
|
|
|
#define WARP_SIZE 32
|
2023-07-22 19:27:34 +00:00
|
|
|
#define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
|
2023-05-25 21:07:29 +00:00
|
|
|
|
2023-07-12 17:26:18 +00:00
|
|
|
#define CUDA_GELU_BLOCK_SIZE 256
|
2023-06-06 19:33:23 +00:00
|
|
|
#define CUDA_SILU_BLOCK_SIZE 256
|
2023-12-13 19:54:54 +00:00
|
|
|
#define CUDA_TANH_BLOCK_SIZE 256
|
2023-11-13 08:58:15 +00:00
|
|
|
#define CUDA_RELU_BLOCK_SIZE 256
|
2024-01-31 13:10:15 +00:00
|
|
|
#define CUDA_HARDSIGMOID_BLOCK_SIZE 256
|
|
|
|
#define CUDA_HARDSWISH_BLOCK_SIZE 256
|
2023-11-13 08:58:15 +00:00
|
|
|
#define CUDA_SQR_BLOCK_SIZE 256
|
2023-06-14 17:47:19 +00:00
|
|
|
#define CUDA_CPY_BLOCK_SIZE 32
|
|
|
|
#define CUDA_SCALE_BLOCK_SIZE 256
|
2023-10-10 07:50:23 +00:00
|
|
|
#define CUDA_CLAMP_BLOCK_SIZE 256
|
2023-06-06 19:33:23 +00:00
|
|
|
#define CUDA_ROPE_BLOCK_SIZE 256
|
2023-12-01 08:51:24 +00:00
|
|
|
#define CUDA_SOFT_MAX_BLOCK_SIZE 1024
|
2023-08-22 11:22:08 +00:00
|
|
|
#define CUDA_ALIBI_BLOCK_SIZE 32
|
2023-06-14 17:47:19 +00:00
|
|
|
#define CUDA_DIAG_MASK_INF_BLOCK_SIZE 32
|
2023-07-05 12:19:42 +00:00
|
|
|
#define CUDA_QUANTIZE_BLOCK_SIZE 256
|
2023-05-14 18:53:23 +00:00
|
|
|
#define CUDA_DEQUANTIZE_BLOCK_SIZE 256
|
2023-10-08 17:19:14 +00:00
|
|
|
#define CUDA_GET_ROWS_BLOCK_SIZE 256
|
2023-12-13 19:54:54 +00:00
|
|
|
#define CUDA_UPSCALE_BLOCK_SIZE 256
|
|
|
|
#define CUDA_CONCAT_BLOCK_SIZE 256
|
|
|
|
#define CUDA_PAD_BLOCK_SIZE 256
|
|
|
|
#define CUDA_ACC_BLOCK_SIZE 256
|
|
|
|
#define CUDA_IM2COL_BLOCK_SIZE 256
|
2024-01-31 13:10:15 +00:00
|
|
|
#define CUDA_POOL2D_BLOCK_SIZE 256
|
2023-05-25 21:07:29 +00:00
|
|
|
|
2024-01-12 19:38:54 +00:00
|
|
|
#define CUDA_Q8_0_NE_ALIGN 2048
|
|
|
|
|
2023-05-25 21:07:29 +00:00
|
|
|
// dmmv = dequantize_mul_mat_vec
|
|
|
|
#ifndef GGML_CUDA_DMMV_X
|
|
|
|
#define GGML_CUDA_DMMV_X 32
|
|
|
|
#endif
|
2023-07-05 12:19:42 +00:00
|
|
|
#ifndef GGML_CUDA_MMV_Y
|
|
|
|
#define GGML_CUDA_MMV_Y 1
|
2023-05-25 21:07:29 +00:00
|
|
|
#endif
|
2023-05-13 13:38:36 +00:00
|
|
|
|
2023-06-16 17:08:44 +00:00
|
|
|
#ifndef K_QUANTS_PER_ITERATION
|
|
|
|
#define K_QUANTS_PER_ITERATION 2
|
|
|
|
#else
|
|
|
|
static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2");
|
|
|
|
#endif
|
|
|
|
|
2023-09-17 14:37:53 +00:00
|
|
|
#ifndef GGML_CUDA_PEER_MAX_BATCH_SIZE
|
|
|
|
#define GGML_CUDA_PEER_MAX_BATCH_SIZE 128
|
|
|
|
#endif // GGML_CUDA_PEER_MAX_BATCH_SIZE
|
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
#define MUL_MAT_SRC1_COL_STRIDE 128
|
|
|
|
|
|
|
|
#define MAX_STREAMS 8
|
2023-11-18 15:11:18 +00:00
|
|
|
static cudaStream_t g_cudaStreams[GGML_CUDA_MAX_DEVICES][MAX_STREAMS] = { { nullptr } };
|
2023-09-11 17:55:51 +00:00
|
|
|
|
2023-07-01 19:49:44 +00:00
|
|
|
struct ggml_tensor_extra_gpu {
|
|
|
|
void * data_device[GGML_CUDA_MAX_DEVICES]; // 1 pointer for each device for split tensors
|
2023-09-11 17:55:51 +00:00
|
|
|
cudaEvent_t events[GGML_CUDA_MAX_DEVICES][MAX_STREAMS]; // events for synchronizing multiple GPUs
|
2023-07-01 19:49:44 +00:00
|
|
|
};
|
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
// this is faster on Windows
|
|
|
|
// probably because the Windows CUDA libraries forget to make this check before invoking the drivers
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_set_device(const int device) {
|
2023-09-11 17:55:51 +00:00
|
|
|
int current_device;
|
|
|
|
CUDA_CHECK(cudaGetDevice(¤t_device));
|
|
|
|
|
|
|
|
if (device == current_device) {
|
2023-12-26 20:23:59 +00:00
|
|
|
return;
|
2023-09-11 17:55:51 +00:00
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
CUDA_CHECK(cudaSetDevice(device));
|
2023-09-11 17:55:51 +00:00
|
|
|
}
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
static int g_device_count = -1;
|
|
|
|
static int g_main_device = 0;
|
2024-01-12 19:07:38 +00:00
|
|
|
static std::array<float, GGML_CUDA_MAX_DEVICES> g_default_tensor_split = {};
|
2023-08-09 07:42:34 +00:00
|
|
|
|
2023-12-24 13:34:22 +00:00
|
|
|
struct cuda_device_capabilities {
|
|
|
|
int cc; // compute capability
|
2024-01-09 07:58:55 +00:00
|
|
|
size_t smpb; // max. shared memory per block
|
2023-12-24 13:34:22 +00:00
|
|
|
bool vmm; // virtual memory support
|
|
|
|
size_t vmm_granularity; // granularity of virtual memory
|
|
|
|
};
|
|
|
|
|
2024-01-09 07:58:55 +00:00
|
|
|
static cuda_device_capabilities g_device_caps[GGML_CUDA_MAX_DEVICES] = { {0, 0, false, 0} };
|
2023-12-24 13:34:22 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
static cublasHandle_t g_cublas_handles[GGML_CUDA_MAX_DEVICES] = {nullptr};
|
|
|
|
|
2023-12-21 17:02:30 +00:00
|
|
|
[[noreturn]]
|
2024-01-23 12:31:56 +00:00
|
|
|
static __device__ void no_device_code(
|
|
|
|
const char * file_name, const int line, const char * function_name, const int arch, const char * arch_list) {
|
|
|
|
|
|
|
|
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
printf("%s:%d: ERROR: HIP kernel %s has no device code compatible with HIP arch %d.\n",
|
|
|
|
file_name, line, function_name, arch);
|
|
|
|
(void) arch_list;
|
|
|
|
#else
|
|
|
|
printf("%s:%d: ERROR: CUDA kernel %s has no device code compatible with CUDA arch %d. ggml-cuda.cu was compiled for: %s\n",
|
|
|
|
file_name, line, function_name, arch, arch_list);
|
|
|
|
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
2023-12-21 17:02:30 +00:00
|
|
|
__trap();
|
|
|
|
|
2024-01-23 12:31:56 +00:00
|
|
|
(void) no_device_code; // suppress unused function warning
|
2023-12-21 17:02:30 +00:00
|
|
|
}
|
|
|
|
|
2024-01-23 12:31:56 +00:00
|
|
|
#ifdef __CUDA_ARCH__
|
|
|
|
#define NO_DEVICE_CODE no_device_code(__FILE__, __LINE__, __FUNCTION__, __CUDA_ARCH__, STRINGIZE(__CUDA_ARCH_LIST__))
|
|
|
|
#else
|
|
|
|
#define NO_DEVICE_CODE GGML_ASSERT(false && "NO_DEVICE_CODE not valid in host code.")
|
|
|
|
#endif // __CUDA_ARCH__
|
|
|
|
|
2023-12-01 08:51:24 +00:00
|
|
|
static __device__ __forceinline__ float warp_reduce_sum(float x) {
|
|
|
|
#pragma unroll
|
|
|
|
for (int mask = 16; mask > 0; mask >>= 1) {
|
|
|
|
x += __shfl_xor_sync(0xffffffff, x, mask, 32);
|
|
|
|
}
|
|
|
|
return x;
|
|
|
|
}
|
|
|
|
|
|
|
|
static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) {
|
|
|
|
#pragma unroll
|
|
|
|
for (int mask = 16; mask > 0; mask >>= 1) {
|
|
|
|
a.x += __shfl_xor_sync(0xffffffff, a.x, mask, 32);
|
|
|
|
a.y += __shfl_xor_sync(0xffffffff, a.y, mask, 32);
|
|
|
|
}
|
|
|
|
return a;
|
|
|
|
}
|
|
|
|
|
2024-02-27 13:22:45 +00:00
|
|
|
#ifdef GGML_CUDA_F16
|
|
|
|
static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
|
|
|
|
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
|
|
|
|
#pragma unroll
|
|
|
|
for (int mask = 16; mask > 0; mask >>= 1) {
|
|
|
|
a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, mask, 32));
|
|
|
|
}
|
|
|
|
return a;
|
|
|
|
#else
|
|
|
|
(void) a;
|
|
|
|
NO_DEVICE_CODE;
|
|
|
|
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
|
|
|
|
}
|
|
|
|
#endif // GGML_CUDA_F16
|
2024-01-09 07:58:55 +00:00
|
|
|
|
2023-12-01 08:51:24 +00:00
|
|
|
static __device__ __forceinline__ float warp_reduce_max(float x) {
|
|
|
|
#pragma unroll
|
|
|
|
for (int mask = 16; mask > 0; mask >>= 1) {
|
|
|
|
x = fmaxf(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
|
|
|
|
}
|
|
|
|
return x;
|
|
|
|
}
|
|
|
|
|
2024-02-19 12:45:41 +00:00
|
|
|
//static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
|
|
|
|
//#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX
|
|
|
|
//#pragma unroll
|
|
|
|
// for (int mask = 16; mask > 0; mask >>= 1) {
|
|
|
|
// x = __hmax2(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
|
|
|
|
// }
|
|
|
|
// return x;
|
|
|
|
//#else
|
|
|
|
// (void) x;
|
|
|
|
// NO_DEVICE_CODE;
|
|
|
|
//#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX
|
|
|
|
//}
|
2024-01-09 07:58:55 +00:00
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
static __device__ __forceinline__ float op_repeat(const float a, const float b) {
|
|
|
|
return b;
|
2023-12-26 20:23:59 +00:00
|
|
|
GGML_UNUSED(a);
|
2023-12-07 20:26:54 +00:00
|
|
|
}
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
static __device__ __forceinline__ float op_add(const float a, const float b) {
|
|
|
|
return a + b;
|
2023-06-06 19:33:23 +00:00
|
|
|
}
|
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
static __device__ __forceinline__ float op_mul(const float a, const float b) {
|
|
|
|
return a * b;
|
|
|
|
}
|
2023-06-28 16:35:54 +00:00
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
static __device__ __forceinline__ float op_div(const float a, const float b) {
|
|
|
|
return a / b;
|
2023-06-28 16:35:54 +00:00
|
|
|
}
|
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
|
|
|
|
static __global__ void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst_t * dst,
|
|
|
|
int ne0, int ne1, int ne2, int ne3,
|
|
|
|
int ne10, int ne11, int ne12, int ne13,
|
|
|
|
/*int s0, */ int s1, int s2, int s3,
|
|
|
|
/*int s10,*/ int s11, int s12, int s13) {
|
|
|
|
const int i0s = blockDim.x*blockIdx.x + threadIdx.x;
|
|
|
|
const int i1 = (blockDim.y*blockIdx.y + threadIdx.y);
|
|
|
|
const int i2 = (blockDim.z*blockIdx.z + threadIdx.z) / ne3;
|
|
|
|
const int i3 = (blockDim.z*blockIdx.z + threadIdx.z) % ne3;
|
2023-11-01 11:49:04 +00:00
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
if (i0s >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
|
2023-11-01 11:49:04 +00:00
|
|
|
return;
|
|
|
|
}
|
2023-12-07 20:26:54 +00:00
|
|
|
|
|
|
|
const int i11 = i1 % ne11;
|
|
|
|
const int i12 = i2 % ne12;
|
|
|
|
const int i13 = i3 % ne13;
|
|
|
|
|
|
|
|
const size_t i_src0 = i3*s3 + i2*s2 + i1*s1;
|
|
|
|
const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
|
|
|
|
const size_t i_dst = i_src0;
|
|
|
|
|
|
|
|
const src0_t * src0_row = src0 + i_src0;
|
|
|
|
const src1_t * src1_row = src1 + i_src1;
|
|
|
|
dst_t * dst_row = dst + i_dst;
|
|
|
|
|
|
|
|
for (int i0 = i0s; i0 < ne0; i0 += blockDim.x*gridDim.x) {
|
|
|
|
const int i10 = i0 % ne10;
|
|
|
|
dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
|
|
|
|
}
|
2023-11-01 11:49:04 +00:00
|
|
|
}
|
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
|
|
|
|
static __global__ void k_bin_bcast_unravel(const src0_t * src0, const src1_t * src1, dst_t * dst,
|
|
|
|
int ne0, int ne1, int ne2, int ne3,
|
|
|
|
int ne10, int ne11, int ne12, int ne13,
|
|
|
|
/*int s0, */ int s1, int s2, int s3,
|
|
|
|
/*int s10,*/ int s11, int s12, int s13) {
|
|
|
|
|
cuda : loading models directly into VRAM, norm calculation on GPU, broadcasting for ggml_mul (#1483)
* Broadcasting for ggml_mul
* CUDA kernel for ggml_mul, norms in VRAM
* GPU weights not in RAM, direct loading with cuFile
* fixup! GPU weights not in RAM, direct loading with cuFile
* fixup! GPU weights not in RAM, direct loading with cuFile
* define default model path once, sync path with readme (#1366)
* ~7% faster Q5_1 AVX2 code (#1477)
* convert.py: Support models which are stored in a single pytorch_model.bin (#1469)
* Support models in a single pytorch_model.bin
* Remove spurious line with typo
* benchmark-matmul: Print the average of the test results (#1490)
* Remove unused n_parts parameter (#1509)
* Fixes #1511 lambda issue for w64devkit (mingw) (#1513)
* Fix for w64devkit and mingw
* make kv_f16 the default for api users (#1517)
* minor : fix compile warnings
* readme : adds WizardLM to the list of supported models (#1485)
* main : make reverse prompt option act as a stop token in non-interactive mode (#1032)
* Make reverse prompt option act as a stop token in non-interactive scenarios
* Making requested review changes
* Update gpt_params_parse and fix a merge error
* Revert "Update gpt_params_parse and fix a merge error"
This reverts commit 2bb2ff1748513591ad45b175a75ed1d8089d84c8.
* Update gpt_params_parse and fix a merge error take 2
* examples : add persistent chat (#1495)
* examples : add persistent chat
* examples : fix whitespace
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* tests : add missing header
* ggml : use F16 instead of F32 in Q4_0, Q4_1, Q8_0 (#1508)
* ggml : use F16 instead of F32 in Q4_0, Q4_1 and Q8_0
* llama : bump LLAMA_FILE_VERSION to 3
* cuda : update Q4 and Q8 dequantize kernels
* ggml : fix AVX dot products
* readme : update performance table + hot topics
* ggml : fix scalar implementation of Q4_1 dot
* llama : fix compile warnings in llama_set_state_data()
* llama : fix name shadowing and C4146 (#1526)
* Fix name shadowing and C4146
* Fix if macros not using defined when required
* Update llama-util.h
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
* Update llama-util.h
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
* Code style
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Fix for mingw (#1462)
* llama : add llama_init_backend() API (close #1527)
* feature : add blis and other BLAS implementation support (#1502)
* feature: add blis support
* feature: allow all BLA_VENDOR to be assigned in cmake arguments. align with whisper.cpp pr 927
* fix: version detection for BLA_SIZEOF_INTEGER, recover min version of cmake
* Fix typo in INTEGER
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Revert "feature : add blis and other BLAS implementation support (#1502)"
This reverts commit 07e9ace0f9da424d82e75df969642522880feb92.
* GPU weights not in RAM, direct loading with cuFile
* llama : code style fixes + progress print fix
* ggml : ggml_mul better broadcast support
* cmake : workarounds for cufile when CMake version < 3.25
* gg rebase fixup
* Loop in llama.cpp, fixed progress callback
* Attempt clang-tidy fix
* llama : fix vram size computation
* Add forgotten fclose()
---------
Co-authored-by: András Salamon <ott2@users.noreply.github.com>
Co-authored-by: Ilya Kurdyukov <59548320+ilyakurdyukov@users.noreply.github.com>
Co-authored-by: Tom Jobbins <784313+TheBloke@users.noreply.github.com>
Co-authored-by: rankaiyx <rankaiyx@rankaiyx.com>
Co-authored-by: Stephan Walter <stephan@walter.name>
Co-authored-by: DannyDaemonic <DannyDaemonic@gmail.com>
Co-authored-by: Erik Scholz <Green-Sky@users.noreply.github.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: David Kennedy <dakennedyd@gmail.com>
Co-authored-by: Jason McCartney <jmac@theroot.org>
Co-authored-by: Evan Jones <evan.q.jones@gmail.com>
Co-authored-by: Maxime <672982+maximegmd@users.noreply.github.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Zenix <zenixls2@gmail.com>
2023-05-20 12:19:28 +00:00
|
|
|
const int i = blockDim.x*blockIdx.x + threadIdx.x;
|
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
const int i3 = i/(ne2*ne1*ne0);
|
|
|
|
const int i2 = (i/(ne1*ne0)) % ne2;
|
|
|
|
const int i1 = (i/ne0) % ne1;
|
|
|
|
const int i0 = i % ne0;
|
|
|
|
|
|
|
|
if (i0 >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
|
cuda : loading models directly into VRAM, norm calculation on GPU, broadcasting for ggml_mul (#1483)
* Broadcasting for ggml_mul
* CUDA kernel for ggml_mul, norms in VRAM
* GPU weights not in RAM, direct loading with cuFile
* fixup! GPU weights not in RAM, direct loading with cuFile
* fixup! GPU weights not in RAM, direct loading with cuFile
* define default model path once, sync path with readme (#1366)
* ~7% faster Q5_1 AVX2 code (#1477)
* convert.py: Support models which are stored in a single pytorch_model.bin (#1469)
* Support models in a single pytorch_model.bin
* Remove spurious line with typo
* benchmark-matmul: Print the average of the test results (#1490)
* Remove unused n_parts parameter (#1509)
* Fixes #1511 lambda issue for w64devkit (mingw) (#1513)
* Fix for w64devkit and mingw
* make kv_f16 the default for api users (#1517)
* minor : fix compile warnings
* readme : adds WizardLM to the list of supported models (#1485)
* main : make reverse prompt option act as a stop token in non-interactive mode (#1032)
* Make reverse prompt option act as a stop token in non-interactive scenarios
* Making requested review changes
* Update gpt_params_parse and fix a merge error
* Revert "Update gpt_params_parse and fix a merge error"
This reverts commit 2bb2ff1748513591ad45b175a75ed1d8089d84c8.
* Update gpt_params_parse and fix a merge error take 2
* examples : add persistent chat (#1495)
* examples : add persistent chat
* examples : fix whitespace
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* tests : add missing header
* ggml : use F16 instead of F32 in Q4_0, Q4_1, Q8_0 (#1508)
* ggml : use F16 instead of F32 in Q4_0, Q4_1 and Q8_0
* llama : bump LLAMA_FILE_VERSION to 3
* cuda : update Q4 and Q8 dequantize kernels
* ggml : fix AVX dot products
* readme : update performance table + hot topics
* ggml : fix scalar implementation of Q4_1 dot
* llama : fix compile warnings in llama_set_state_data()
* llama : fix name shadowing and C4146 (#1526)
* Fix name shadowing and C4146
* Fix if macros not using defined when required
* Update llama-util.h
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
* Update llama-util.h
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
* Code style
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Fix for mingw (#1462)
* llama : add llama_init_backend() API (close #1527)
* feature : add blis and other BLAS implementation support (#1502)
* feature: add blis support
* feature: allow all BLA_VENDOR to be assigned in cmake arguments. align with whisper.cpp pr 927
* fix: version detection for BLA_SIZEOF_INTEGER, recover min version of cmake
* Fix typo in INTEGER
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Revert "feature : add blis and other BLAS implementation support (#1502)"
This reverts commit 07e9ace0f9da424d82e75df969642522880feb92.
* GPU weights not in RAM, direct loading with cuFile
* llama : code style fixes + progress print fix
* ggml : ggml_mul better broadcast support
* cmake : workarounds for cufile when CMake version < 3.25
* gg rebase fixup
* Loop in llama.cpp, fixed progress callback
* Attempt clang-tidy fix
* llama : fix vram size computation
* Add forgotten fclose()
---------
Co-authored-by: András Salamon <ott2@users.noreply.github.com>
Co-authored-by: Ilya Kurdyukov <59548320+ilyakurdyukov@users.noreply.github.com>
Co-authored-by: Tom Jobbins <784313+TheBloke@users.noreply.github.com>
Co-authored-by: rankaiyx <rankaiyx@rankaiyx.com>
Co-authored-by: Stephan Walter <stephan@walter.name>
Co-authored-by: DannyDaemonic <DannyDaemonic@gmail.com>
Co-authored-by: Erik Scholz <Green-Sky@users.noreply.github.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: David Kennedy <dakennedyd@gmail.com>
Co-authored-by: Jason McCartney <jmac@theroot.org>
Co-authored-by: Evan Jones <evan.q.jones@gmail.com>
Co-authored-by: Maxime <672982+maximegmd@users.noreply.github.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Zenix <zenixls2@gmail.com>
2023-05-20 12:19:28 +00:00
|
|
|
return;
|
|
|
|
}
|
2023-12-07 20:26:54 +00:00
|
|
|
|
|
|
|
const int i11 = i1 % ne11;
|
|
|
|
const int i12 = i2 % ne12;
|
|
|
|
const int i13 = i3 % ne13;
|
|
|
|
|
|
|
|
const size_t i_src0 = i3*s3 + i2*s2 + i1*s1;
|
|
|
|
const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
|
|
|
|
const size_t i_dst = i_src0;
|
|
|
|
|
|
|
|
const src0_t * src0_row = src0 + i_src0;
|
|
|
|
const src1_t * src1_row = src1 + i_src1;
|
|
|
|
dst_t * dst_row = dst + i_dst;
|
|
|
|
|
|
|
|
const int i10 = i0 % ne10;
|
|
|
|
dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
|
cuda : loading models directly into VRAM, norm calculation on GPU, broadcasting for ggml_mul (#1483)
* Broadcasting for ggml_mul
* CUDA kernel for ggml_mul, norms in VRAM
* GPU weights not in RAM, direct loading with cuFile
* fixup! GPU weights not in RAM, direct loading with cuFile
* fixup! GPU weights not in RAM, direct loading with cuFile
* define default model path once, sync path with readme (#1366)
* ~7% faster Q5_1 AVX2 code (#1477)
* convert.py: Support models which are stored in a single pytorch_model.bin (#1469)
* Support models in a single pytorch_model.bin
* Remove spurious line with typo
* benchmark-matmul: Print the average of the test results (#1490)
* Remove unused n_parts parameter (#1509)
* Fixes #1511 lambda issue for w64devkit (mingw) (#1513)
* Fix for w64devkit and mingw
* make kv_f16 the default for api users (#1517)
* minor : fix compile warnings
* readme : adds WizardLM to the list of supported models (#1485)
* main : make reverse prompt option act as a stop token in non-interactive mode (#1032)
* Make reverse prompt option act as a stop token in non-interactive scenarios
* Making requested review changes
* Update gpt_params_parse and fix a merge error
* Revert "Update gpt_params_parse and fix a merge error"
This reverts commit 2bb2ff1748513591ad45b175a75ed1d8089d84c8.
* Update gpt_params_parse and fix a merge error take 2
* examples : add persistent chat (#1495)
* examples : add persistent chat
* examples : fix whitespace
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* tests : add missing header
* ggml : use F16 instead of F32 in Q4_0, Q4_1, Q8_0 (#1508)
* ggml : use F16 instead of F32 in Q4_0, Q4_1 and Q8_0
* llama : bump LLAMA_FILE_VERSION to 3
* cuda : update Q4 and Q8 dequantize kernels
* ggml : fix AVX dot products
* readme : update performance table + hot topics
* ggml : fix scalar implementation of Q4_1 dot
* llama : fix compile warnings in llama_set_state_data()
* llama : fix name shadowing and C4146 (#1526)
* Fix name shadowing and C4146
* Fix if macros not using defined when required
* Update llama-util.h
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
* Update llama-util.h
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
* Code style
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Fix for mingw (#1462)
* llama : add llama_init_backend() API (close #1527)
* feature : add blis and other BLAS implementation support (#1502)
* feature: add blis support
* feature: allow all BLA_VENDOR to be assigned in cmake arguments. align with whisper.cpp pr 927
* fix: version detection for BLA_SIZEOF_INTEGER, recover min version of cmake
* Fix typo in INTEGER
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Revert "feature : add blis and other BLAS implementation support (#1502)"
This reverts commit 07e9ace0f9da424d82e75df969642522880feb92.
* GPU weights not in RAM, direct loading with cuFile
* llama : code style fixes + progress print fix
* ggml : ggml_mul better broadcast support
* cmake : workarounds for cufile when CMake version < 3.25
* gg rebase fixup
* Loop in llama.cpp, fixed progress callback
* Attempt clang-tidy fix
* llama : fix vram size computation
* Add forgotten fclose()
---------
Co-authored-by: András Salamon <ott2@users.noreply.github.com>
Co-authored-by: Ilya Kurdyukov <59548320+ilyakurdyukov@users.noreply.github.com>
Co-authored-by: Tom Jobbins <784313+TheBloke@users.noreply.github.com>
Co-authored-by: rankaiyx <rankaiyx@rankaiyx.com>
Co-authored-by: Stephan Walter <stephan@walter.name>
Co-authored-by: DannyDaemonic <DannyDaemonic@gmail.com>
Co-authored-by: Erik Scholz <Green-Sky@users.noreply.github.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: David Kennedy <dakennedyd@gmail.com>
Co-authored-by: Jason McCartney <jmac@theroot.org>
Co-authored-by: Evan Jones <evan.q.jones@gmail.com>
Co-authored-by: Maxime <672982+maximegmd@users.noreply.github.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Zenix <zenixls2@gmail.com>
2023-05-20 12:19:28 +00:00
|
|
|
}
|
|
|
|
|
2023-12-13 19:54:54 +00:00
|
|
|
static __global__ void acc_f32(const float * x, const float * y, float * dst, const int ne,
|
|
|
|
const int ne10, const int ne11, const int ne12,
|
|
|
|
const int nb1, const int nb2, int offset) {
|
|
|
|
const int i = blockDim.x * blockIdx.x + threadIdx.x;
|
|
|
|
if (i >= ne) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
int src1_idx = i - offset;
|
|
|
|
int oz = src1_idx / nb2;
|
|
|
|
int oy = (src1_idx - (oz * nb2)) / nb1;
|
|
|
|
int ox = src1_idx % nb1;
|
|
|
|
if (src1_idx >= 0 && ox < ne10 && oy < ne11 && oz < ne12) {
|
|
|
|
dst[i] = x[i] + y[ox + oy * ne10 + oz * ne10 * ne11];
|
|
|
|
} else {
|
|
|
|
dst[i] = x[i];
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-07-12 17:26:18 +00:00
|
|
|
static __global__ void gelu_f32(const float * x, float * dst, const int k) {
|
2023-07-13 13:58:09 +00:00
|
|
|
const float GELU_COEF_A = 0.044715f;
|
|
|
|
const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
|
2023-07-12 17:26:18 +00:00
|
|
|
const int i = blockDim.x*blockIdx.x + threadIdx.x;
|
|
|
|
|
|
|
|
if (i >= k) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
float xi = x[i];
|
|
|
|
dst[i] = 0.5f*xi*(1.0f + tanhf(SQRT_2_OVER_PI*xi*(1.0f + GELU_COEF_A*xi*xi)));
|
|
|
|
}
|
|
|
|
|
2023-06-06 19:33:23 +00:00
|
|
|
static __global__ void silu_f32(const float * x, float * dst, const int k) {
|
|
|
|
const int i = blockDim.x*blockIdx.x + threadIdx.x;
|
|
|
|
|
|
|
|
if (i >= k) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
dst[i] = x[i] / (1.0f + expf(-x[i]));
|
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static __global__ void gelu_quick_f32(const float * x, float * dst, int k) {
|
2023-12-13 19:54:54 +00:00
|
|
|
const float GELU_QUICK_COEF = -1.702f;
|
|
|
|
const int i = blockDim.x*blockIdx.x + threadIdx.x;
|
|
|
|
if (i >= k) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
dst[i] = x[i] * (1.0f / (1.0f + expf(GELU_QUICK_COEF * x[i])));
|
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static __global__ void tanh_f32(const float * x, float * dst, int k) {
|
2023-12-13 19:54:54 +00:00
|
|
|
const int i = blockDim.x*blockIdx.x + threadIdx.x;
|
|
|
|
if (i >= k) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
dst[i] = tanhf(x[i]);
|
|
|
|
}
|
|
|
|
|
2023-11-13 08:58:15 +00:00
|
|
|
static __global__ void relu_f32(const float * x, float * dst, const int k) {
|
|
|
|
const int i = blockDim.x*blockIdx.x + threadIdx.x;
|
|
|
|
|
|
|
|
if (i >= k) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
dst[i] = fmaxf(x[i], 0);
|
|
|
|
}
|
|
|
|
|
2024-01-31 13:10:15 +00:00
|
|
|
static __global__ void hardsigmoid_f32(const float * x, float * dst, const int k) {
|
|
|
|
const int i = blockDim.x*blockIdx.x + threadIdx.x;
|
|
|
|
|
|
|
|
if (i >= k) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
dst[i] = fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f));
|
|
|
|
}
|
|
|
|
|
|
|
|
static __global__ void hardswish_f32(const float * x, float * dst, const int k) {
|
|
|
|
const int i = blockDim.x*blockIdx.x + threadIdx.x;
|
|
|
|
|
|
|
|
if (i >= k) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
dst[i] = x[i] * fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f));
|
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static __global__ void leaky_relu_f32(const float * x, float * dst, const int k, const float negative_slope) {
|
2023-12-13 19:54:54 +00:00
|
|
|
const int i = blockDim.x*blockIdx.x + threadIdx.x;
|
|
|
|
if (i >= k) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
dst[i] = fmaxf(x[i], 0) + fminf(x[i], 0.0f) * negative_slope;
|
|
|
|
}
|
|
|
|
|
2023-11-13 08:58:15 +00:00
|
|
|
static __global__ void sqr_f32(const float * x, float * dst, const int k) {
|
|
|
|
const int i = blockDim.x*blockIdx.x + threadIdx.x;
|
|
|
|
|
|
|
|
if (i >= k) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
dst[i] = x[i] * x[i];
|
|
|
|
}
|
|
|
|
|
2023-09-04 06:53:30 +00:00
|
|
|
template <int block_size>
|
2023-12-07 20:26:54 +00:00
|
|
|
static __global__ void norm_f32(const float * x, float * dst, const int ncols, const float eps) {
|
2023-07-11 19:53:34 +00:00
|
|
|
const int row = blockIdx.x*blockDim.y + threadIdx.y;
|
|
|
|
const int tid = threadIdx.x;
|
|
|
|
|
2023-09-04 06:53:30 +00:00
|
|
|
float2 mean_var = make_float2(0.f, 0.f);
|
2023-07-11 19:53:34 +00:00
|
|
|
|
2023-09-04 06:53:30 +00:00
|
|
|
for (int col = tid; col < ncols; col += block_size) {
|
2023-07-11 19:53:34 +00:00
|
|
|
const float xi = x[row*ncols + col];
|
2023-09-04 06:53:30 +00:00
|
|
|
mean_var.x += xi;
|
|
|
|
mean_var.y += xi * xi;
|
2023-07-11 19:53:34 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// sum up partial sums
|
2023-09-04 06:53:30 +00:00
|
|
|
mean_var = warp_reduce_sum(mean_var);
|
|
|
|
if (block_size > WARP_SIZE) {
|
|
|
|
__shared__ float2 s_sum[32];
|
|
|
|
int warp_id = threadIdx.x / WARP_SIZE;
|
|
|
|
int lane_id = threadIdx.x % WARP_SIZE;
|
|
|
|
if (lane_id == 0) {
|
|
|
|
s_sum[warp_id] = mean_var;
|
|
|
|
}
|
|
|
|
__syncthreads();
|
|
|
|
mean_var = s_sum[lane_id];
|
|
|
|
mean_var = warp_reduce_sum(mean_var);
|
2023-07-11 19:53:34 +00:00
|
|
|
}
|
|
|
|
|
2023-09-04 06:53:30 +00:00
|
|
|
const float mean = mean_var.x / ncols;
|
|
|
|
const float var = mean_var.y / ncols - mean * mean;
|
|
|
|
const float inv_std = rsqrtf(var + eps);
|
2023-07-11 19:53:34 +00:00
|
|
|
|
2023-09-04 06:53:30 +00:00
|
|
|
for (int col = tid; col < ncols; col += block_size) {
|
|
|
|
dst[row*ncols + col] = (x[row*ncols + col] - mean) * inv_std;
|
2023-07-11 19:53:34 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static __global__ void concat_f32(const float * x,const float * y, float * dst, const int ne0, const int ne02) {
|
2023-12-13 19:54:54 +00:00
|
|
|
int nidx = threadIdx.x + blockIdx.x * blockDim.x;
|
|
|
|
if (nidx >= ne0) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
// operation
|
|
|
|
int offset_dst =
|
|
|
|
nidx +
|
|
|
|
blockIdx.y * ne0 +
|
|
|
|
blockIdx.z * ne0 * gridDim.y;
|
|
|
|
if (blockIdx.z < ne02) { // src0
|
|
|
|
int offset_src =
|
|
|
|
nidx +
|
|
|
|
blockIdx.y * ne0 +
|
|
|
|
blockIdx.z * ne0 * gridDim.y;
|
|
|
|
dst[offset_dst] = x[offset_src];
|
|
|
|
} else {
|
|
|
|
int offset_src =
|
|
|
|
nidx +
|
|
|
|
blockIdx.y * ne0 +
|
|
|
|
(blockIdx.z - ne02) * ne0 * gridDim.y;
|
|
|
|
dst[offset_dst] = y[offset_src];
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static __global__ void upscale_f32(const float * x, float * dst, const int ne00, const int nb02, const int scale_factor) {
|
2023-12-13 19:54:54 +00:00
|
|
|
int ne0 = ne00 * scale_factor;
|
|
|
|
int nidx = threadIdx.x + blockIdx.x * blockDim.x;
|
|
|
|
if (nidx >= ne0) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
// operation
|
|
|
|
int i00 = nidx / scale_factor;
|
|
|
|
int i01 = blockIdx.y / scale_factor;
|
|
|
|
int offset_src =
|
|
|
|
i00 +
|
|
|
|
i01 * ne00 +
|
|
|
|
blockIdx.z * nb02;
|
|
|
|
int offset_dst =
|
|
|
|
nidx +
|
|
|
|
blockIdx.y * ne0 +
|
|
|
|
blockIdx.z * ne0 * gridDim.y;
|
|
|
|
dst[offset_dst] = x[offset_src];
|
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static __global__ void pad_f32(const float * x, float * dst, const int ne0, const int ne00, const int ne01, const int ne02) {
|
2023-12-13 19:54:54 +00:00
|
|
|
int nidx = threadIdx.x + blockIdx.x * blockDim.x;
|
|
|
|
if (nidx >= ne0) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
// operation
|
|
|
|
int offset_dst =
|
|
|
|
nidx +
|
|
|
|
blockIdx.y * ne0 +
|
|
|
|
blockIdx.z * ne0 * gridDim.y;
|
|
|
|
if (nidx < ne00 && blockIdx.y < ne01 && blockIdx.z < ne02) {
|
|
|
|
int offset_src =
|
|
|
|
nidx +
|
|
|
|
blockIdx.y * ne00 +
|
|
|
|
blockIdx.z * ne00 * ne01;
|
|
|
|
dst[offset_dst] = x[offset_src];
|
|
|
|
} else {
|
|
|
|
dst[offset_dst] = 0.0f;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
template <int block_size>
|
|
|
|
static __global__ void group_norm_f32(const float * x, float * dst, const int group_size, const int ne_elements, const float eps) {
|
|
|
|
int start = blockIdx.x * group_size;
|
|
|
|
int end = start + group_size;
|
|
|
|
|
|
|
|
start += threadIdx.x;
|
|
|
|
|
|
|
|
if (end >= ne_elements) {
|
|
|
|
end = ne_elements;
|
|
|
|
}
|
|
|
|
|
|
|
|
float tmp = 0.0f; // partial sum for thread in warp
|
|
|
|
|
|
|
|
for (int j = start; j < end; j += block_size) {
|
|
|
|
tmp += x[j];
|
|
|
|
}
|
|
|
|
|
|
|
|
tmp = warp_reduce_sum(tmp);
|
|
|
|
if (block_size > WARP_SIZE) {
|
|
|
|
__shared__ float s_sum[32];
|
|
|
|
int warp_id = threadIdx.x / WARP_SIZE;
|
|
|
|
int lane_id = threadIdx.x % WARP_SIZE;
|
|
|
|
if (lane_id == 0) {
|
|
|
|
s_sum[warp_id] = tmp;
|
|
|
|
}
|
|
|
|
__syncthreads();
|
|
|
|
tmp = s_sum[lane_id];
|
|
|
|
tmp = warp_reduce_sum(tmp);
|
|
|
|
}
|
|
|
|
|
|
|
|
float mean = tmp / group_size;
|
|
|
|
tmp = 0.0f;
|
|
|
|
|
|
|
|
for (int j = start; j < end; j += block_size) {
|
|
|
|
float xi = x[j] - mean;
|
|
|
|
dst[j] = xi;
|
|
|
|
tmp += xi * xi;
|
|
|
|
}
|
|
|
|
|
|
|
|
tmp = warp_reduce_sum(tmp);
|
|
|
|
if (block_size > WARP_SIZE) {
|
|
|
|
__shared__ float s_sum[32];
|
|
|
|
int warp_id = threadIdx.x / WARP_SIZE;
|
|
|
|
int lane_id = threadIdx.x % WARP_SIZE;
|
|
|
|
if (lane_id == 0) {
|
|
|
|
s_sum[warp_id] = tmp;
|
|
|
|
}
|
|
|
|
__syncthreads();
|
|
|
|
tmp = s_sum[lane_id];
|
|
|
|
tmp = warp_reduce_sum(tmp);
|
|
|
|
}
|
|
|
|
|
|
|
|
float variance = tmp / group_size;
|
|
|
|
float scale = rsqrtf(variance + eps);
|
|
|
|
for (int j = start; j < end; j += block_size) {
|
|
|
|
dst[j] *= scale;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-09-04 06:53:30 +00:00
|
|
|
template <int block_size>
|
2023-07-24 15:57:12 +00:00
|
|
|
static __global__ void rms_norm_f32(const float * x, float * dst, const int ncols, const float eps) {
|
2023-06-06 19:33:23 +00:00
|
|
|
const int row = blockIdx.x*blockDim.y + threadIdx.y;
|
|
|
|
const int tid = threadIdx.x;
|
|
|
|
|
|
|
|
float tmp = 0.0f; // partial sum for thread in warp
|
|
|
|
|
2023-09-04 06:53:30 +00:00
|
|
|
for (int col = tid; col < ncols; col += block_size) {
|
2023-06-06 19:33:23 +00:00
|
|
|
const float xi = x[row*ncols + col];
|
|
|
|
tmp += xi * xi;
|
|
|
|
}
|
|
|
|
|
|
|
|
// sum up partial sums
|
2023-09-04 06:53:30 +00:00
|
|
|
tmp = warp_reduce_sum(tmp);
|
|
|
|
if (block_size > WARP_SIZE) {
|
|
|
|
__shared__ float s_sum[32];
|
|
|
|
int warp_id = threadIdx.x / WARP_SIZE;
|
|
|
|
int lane_id = threadIdx.x % WARP_SIZE;
|
|
|
|
if (lane_id == 0) {
|
|
|
|
s_sum[warp_id] = tmp;
|
|
|
|
}
|
|
|
|
__syncthreads();
|
|
|
|
tmp = s_sum[lane_id];
|
|
|
|
tmp = warp_reduce_sum(tmp);
|
2023-06-06 19:33:23 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
const float mean = tmp / ncols;
|
2023-07-11 19:53:34 +00:00
|
|
|
const float scale = rsqrtf(mean + eps);
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2023-09-04 06:53:30 +00:00
|
|
|
for (int col = tid; col < ncols; col += block_size) {
|
2023-06-06 19:33:23 +00:00
|
|
|
dst[row*ncols + col] = scale * x[row*ncols + col];
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-06-19 08:23:56 +00:00
|
|
|
static __device__ __forceinline__ void dequantize_q4_0(const void * vx, const int ib, const int iqs, dfloat2 & v){
|
2023-05-13 13:38:36 +00:00
|
|
|
const block_q4_0 * x = (const block_q4_0 *) vx;
|
|
|
|
|
2023-06-19 08:23:56 +00:00
|
|
|
const dfloat d = x[ib].d;
|
2023-05-13 13:38:36 +00:00
|
|
|
|
2023-06-19 08:23:56 +00:00
|
|
|
const int vui = x[ib].qs[iqs];
|
2023-05-13 13:38:36 +00:00
|
|
|
|
2023-06-19 08:23:56 +00:00
|
|
|
v.x = vui & 0xF;
|
|
|
|
v.y = vui >> 4;
|
2023-05-13 13:38:36 +00:00
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
#ifdef GGML_CUDA_F16
|
2023-06-19 08:23:56 +00:00
|
|
|
v = __hsub2(v, {8.0f, 8.0f});
|
|
|
|
v = __hmul2(v, {d, d});
|
|
|
|
#else
|
|
|
|
v.x = (v.x - 8.0f) * d;
|
|
|
|
v.y = (v.y - 8.0f) * d;
|
2023-07-29 21:04:44 +00:00
|
|
|
#endif // GGML_CUDA_F16
|
2023-05-13 13:38:36 +00:00
|
|
|
}
|
|
|
|
|
2023-06-19 08:23:56 +00:00
|
|
|
static __device__ __forceinline__ void dequantize_q4_1(const void * vx, const int ib, const int iqs, dfloat2 & v){
|
2023-05-13 13:38:36 +00:00
|
|
|
const block_q4_1 * x = (const block_q4_1 *) vx;
|
|
|
|
|
2023-08-25 09:09:42 +00:00
|
|
|
const dfloat d = __low2half(x[ib].dm);
|
|
|
|
const dfloat m = __high2half(x[ib].dm);
|
2023-05-13 13:38:36 +00:00
|
|
|
|
2023-06-19 08:23:56 +00:00
|
|
|
const int vui = x[ib].qs[iqs];
|
2023-05-13 13:38:36 +00:00
|
|
|
|
2023-06-19 08:23:56 +00:00
|
|
|
v.x = vui & 0xF;
|
|
|
|
v.y = vui >> 4;
|
2023-05-13 13:38:36 +00:00
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
#ifdef GGML_CUDA_F16
|
2023-06-19 08:23:56 +00:00
|
|
|
v = __hmul2(v, {d, d});
|
|
|
|
v = __hadd2(v, {m, m});
|
|
|
|
#else
|
|
|
|
v.x = (v.x * d) + m;
|
|
|
|
v.y = (v.y * d) + m;
|
2023-07-29 21:04:44 +00:00
|
|
|
#endif // GGML_CUDA_F16
|
2023-05-13 13:38:36 +00:00
|
|
|
}
|
|
|
|
|
2023-06-19 08:23:56 +00:00
|
|
|
static __device__ __forceinline__ void dequantize_q5_0(const void * vx, const int ib, const int iqs, dfloat2 & v){
|
2023-05-13 13:38:36 +00:00
|
|
|
const block_q5_0 * x = (const block_q5_0 *) vx;
|
|
|
|
|
2023-06-19 08:23:56 +00:00
|
|
|
const dfloat d = x[ib].d;
|
2023-05-13 13:38:36 +00:00
|
|
|
|
|
|
|
uint32_t qh;
|
|
|
|
memcpy(&qh, x[ib].qh, sizeof(qh));
|
|
|
|
|
2023-06-19 08:23:56 +00:00
|
|
|
const int xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10;
|
|
|
|
const int xh_1 = ((qh >> (iqs + 12)) ) & 0x10;
|
2023-05-13 13:38:36 +00:00
|
|
|
|
2023-06-19 08:23:56 +00:00
|
|
|
v.x = ((x[ib].qs[iqs] & 0xf) | xh_0);
|
|
|
|
v.y = ((x[ib].qs[iqs] >> 4) | xh_1);
|
2023-05-13 13:38:36 +00:00
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
#ifdef GGML_CUDA_F16
|
2023-06-19 08:23:56 +00:00
|
|
|
v = __hsub2(v, {16.0f, 16.0f});
|
|
|
|
v = __hmul2(v, {d, d});
|
|
|
|
#else
|
|
|
|
v.x = (v.x - 16.0f) * d;
|
|
|
|
v.y = (v.y - 16.0f) * d;
|
2023-07-29 21:04:44 +00:00
|
|
|
#endif // GGML_CUDA_F16
|
2023-05-13 13:38:36 +00:00
|
|
|
}
|
|
|
|
|
2023-06-19 08:23:56 +00:00
|
|
|
static __device__ __forceinline__ void dequantize_q5_1(const void * vx, const int ib, const int iqs, dfloat2 & v){
|
2023-05-13 13:38:36 +00:00
|
|
|
const block_q5_1 * x = (const block_q5_1 *) vx;
|
|
|
|
|
2023-08-25 09:09:42 +00:00
|
|
|
const dfloat d = __low2half(x[ib].dm);
|
|
|
|
const dfloat m = __high2half(x[ib].dm);
|
2023-05-13 13:38:36 +00:00
|
|
|
|
|
|
|
uint32_t qh;
|
|
|
|
memcpy(&qh, x[ib].qh, sizeof(qh));
|
|
|
|
|
2023-06-19 08:23:56 +00:00
|
|
|
const int xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10;
|
|
|
|
const int xh_1 = ((qh >> (iqs + 12)) ) & 0x10;
|
2023-05-13 13:38:36 +00:00
|
|
|
|
2023-06-19 08:23:56 +00:00
|
|
|
v.x = ((x[ib].qs[iqs] & 0xf) | xh_0);
|
|
|
|
v.y = ((x[ib].qs[iqs] >> 4) | xh_1);
|
2023-05-13 13:38:36 +00:00
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
#ifdef GGML_CUDA_F16
|
2023-06-19 08:23:56 +00:00
|
|
|
v = __hmul2(v, {d, d});
|
|
|
|
v = __hadd2(v, {m, m});
|
|
|
|
#else
|
|
|
|
v.x = (v.x * d) + m;
|
|
|
|
v.y = (v.y * d) + m;
|
2023-07-29 21:04:44 +00:00
|
|
|
#endif // GGML_CUDA_F16
|
2023-05-13 13:38:36 +00:00
|
|
|
}
|
|
|
|
|
2023-06-19 08:23:56 +00:00
|
|
|
static __device__ __forceinline__ void dequantize_q8_0(const void * vx, const int ib, const int iqs, dfloat2 & v){
|
2023-05-13 13:38:36 +00:00
|
|
|
const block_q8_0 * x = (const block_q8_0 *) vx;
|
|
|
|
|
2023-06-19 08:23:56 +00:00
|
|
|
const dfloat d = x[ib].d;
|
2023-05-13 13:38:36 +00:00
|
|
|
|
2023-06-19 08:23:56 +00:00
|
|
|
v.x = x[ib].qs[iqs + 0];
|
|
|
|
v.y = x[ib].qs[iqs + 1];
|
2023-05-13 13:38:36 +00:00
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
#ifdef GGML_CUDA_F16
|
2023-06-19 08:23:56 +00:00
|
|
|
v = __hmul2(v, {d, d});
|
|
|
|
#else
|
|
|
|
v.x *= d;
|
|
|
|
v.y *= d;
|
2023-07-29 21:04:44 +00:00
|
|
|
#endif // GGML_CUDA_F16
|
2023-05-13 13:38:36 +00:00
|
|
|
}
|
|
|
|
|
2024-01-15 05:48:06 +00:00
|
|
|
template<typename dst_t>
|
|
|
|
static __global__ void dequantize_block_q4_0(const void * __restrict__ vx, dst_t * __restrict__ yy, int nb32) {
|
|
|
|
|
|
|
|
const int i = blockIdx.x;
|
|
|
|
|
|
|
|
// assume 32 threads
|
|
|
|
const int tid = threadIdx.x;
|
|
|
|
const int il = tid/8;
|
|
|
|
const int ir = tid%8;
|
|
|
|
const int ib = 8*i + ir;
|
|
|
|
if (ib >= nb32) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
dst_t * y = yy + 256*i + 32*ir + 4*il;
|
|
|
|
|
|
|
|
const block_q4_0 * x = (const block_q4_0 *)vx + ib;
|
|
|
|
const float d = __half2float(x->d);
|
|
|
|
const float dm = -8*d;
|
|
|
|
|
|
|
|
const uint8_t * q = x->qs + 4*il;
|
|
|
|
|
|
|
|
for (int l = 0; l < 4; ++l) {
|
|
|
|
y[l+ 0] = d * (q[l] & 0xF) + dm;
|
|
|
|
y[l+16] = d * (q[l] >> 4) + dm;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
template<typename dst_t>
|
|
|
|
static __global__ void dequantize_block_q4_1(const void * __restrict__ vx, dst_t * __restrict__ yy, int nb32) {
|
|
|
|
|
|
|
|
const int i = blockIdx.x;
|
|
|
|
|
|
|
|
// assume 32 threads
|
|
|
|
const int tid = threadIdx.x;
|
|
|
|
const int il = tid/8;
|
|
|
|
const int ir = tid%8;
|
|
|
|
const int ib = 8*i + ir;
|
|
|
|
if (ib >= nb32) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
dst_t * y = yy + 256*i + 32*ir + 4*il;
|
|
|
|
|
|
|
|
const block_q4_1 * x = (const block_q4_1 *)vx + ib;
|
|
|
|
const float2 d = __half22float2(x->dm);
|
|
|
|
|
|
|
|
const uint8_t * q = x->qs + 4*il;
|
|
|
|
|
|
|
|
for (int l = 0; l < 4; ++l) {
|
|
|
|
y[l+ 0] = d.x * (q[l] & 0xF) + d.y;
|
|
|
|
y[l+16] = d.x * (q[l] >> 4) + d.y;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
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
|
|
|
|
|
2023-09-30 16:12:57 +00:00
|
|
|
template<typename dst_t>
|
|
|
|
static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
|
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 i = blockIdx.x;
|
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_q2_K * x = (const block_q2_K *) vx;
|
|
|
|
|
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 tid = threadIdx.x;
|
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 : 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 n = tid/32;
|
|
|
|
const int l = tid - 32*n;
|
|
|
|
const int is = 8*n + l/16;
|
|
|
|
|
|
|
|
const uint8_t q = x[i].qs[32*n + l];
|
2023-09-30 16:12:57 +00:00
|
|
|
dst_t * y = yy + i*QK_K + 128*n;
|
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
|
|
|
|
2023-08-25 09:09:42 +00:00
|
|
|
float dall = __low2half(x[i].dm);
|
|
|
|
float dmin = __high2half(x[i].dm);
|
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
|
|
|
y[l+ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
|
|
|
|
y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4);
|
|
|
|
y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4);
|
|
|
|
y[l+96] = dall * (x[i].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is+6] >> 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
|
|
|
#else
|
|
|
|
const int is = tid/16; // 0 or 1
|
|
|
|
const int il = tid%16; // 0...15
|
|
|
|
const uint8_t q = x[i].qs[il] >> (2*is);
|
2023-09-30 16:12:57 +00:00
|
|
|
dst_t * y = yy + i*QK_K + 16*is + il;
|
2023-08-25 09:09:42 +00:00
|
|
|
float dall = __low2half(x[i].dm);
|
|
|
|
float dmin = __high2half(x[i].dm);
|
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
|
|
|
y[ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
|
|
|
|
y[32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+2] >> 4);
|
|
|
|
#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
|
|
|
|
|
|
|
}
|
|
|
|
|
2023-09-30 16:12:57 +00:00
|
|
|
template<typename dst_t>
|
|
|
|
static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
|
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 int i = blockIdx.x;
|
2023-06-07 07:59:52 +00:00
|
|
|
const block_q3_K * x = (const block_q3_K *) vx;
|
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
|
|
|
|
const int r = threadIdx.x/4;
|
|
|
|
const int tid = r/2;
|
|
|
|
const int is0 = r%2;
|
|
|
|
const int l0 = 16*is0 + 4*(threadIdx.x%4);
|
|
|
|
const int n = tid / 4;
|
|
|
|
const int j = tid - 4*n;
|
|
|
|
|
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
|
|
|
uint8_t m = 1 << (4*n + j);
|
|
|
|
int is = 8*n + 2*j + is0;
|
|
|
|
int shift = 2*j;
|
|
|
|
|
|
|
|
int8_t us = is < 4 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+8] >> 0) & 3) << 4) :
|
|
|
|
is < 8 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+4] >> 2) & 3) << 4) :
|
|
|
|
is < 12 ? (x[i].scales[is-8] >> 4) | (((x[i].scales[is+0] >> 4) & 3) << 4) :
|
|
|
|
(x[i].scales[is-8] >> 4) | (((x[i].scales[is-4] >> 6) & 3) << 4);
|
|
|
|
float d_all = x[i].d;
|
|
|
|
float dl = d_all * (us - 32);
|
|
|
|
|
2023-09-30 16:12:57 +00:00
|
|
|
dst_t * y = yy + i*QK_K + 128*n + 32*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
|
|
|
const uint8_t * q = x[i].qs + 32*n;
|
|
|
|
const uint8_t * hm = x[i].hmask;
|
|
|
|
|
|
|
|
for (int l = l0; l < l0+4; ++l) y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 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
|
|
|
#else
|
|
|
|
const int tid = threadIdx.x;
|
|
|
|
const int is = tid/16; // 0 or 1
|
|
|
|
const int il = tid%16; // 0...15
|
|
|
|
const int im = il/8; // 0...1
|
|
|
|
const int in = il%8; // 0...7
|
|
|
|
|
2023-09-30 16:12:57 +00:00
|
|
|
dst_t * y = yy + i*QK_K + 16*is + il;
|
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 q = x[i].qs[il] >> (2*is);
|
|
|
|
const uint8_t h = x[i].hmask[in] >> (2*is + im);
|
|
|
|
const float d = (float)x[i].d;
|
|
|
|
|
|
|
|
if (is == 0) {
|
|
|
|
y[ 0] = d * ((x[i].scales[0] & 0xF) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4));
|
|
|
|
y[32] = d * ((x[i].scales[1] & 0xF) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4));
|
|
|
|
} else {
|
|
|
|
y[ 0] = d * ((x[i].scales[0] >> 4) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4));
|
|
|
|
y[32] = d * ((x[i].scales[1] >> 4) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4));
|
|
|
|
}
|
|
|
|
#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
|
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
|
|
|
static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & 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);
|
|
|
|
}
|
|
|
|
}
|
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
|
|
|
|
2023-09-30 16:12:57 +00:00
|
|
|
template<typename dst_t>
|
|
|
|
static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
|
2023-06-07 07:59:52 +00:00
|
|
|
const block_q4_K * x = (const block_q4_K *) vx;
|
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 i = blockIdx.x;
|
|
|
|
|
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 : 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
|
|
|
// assume 32 threads
|
|
|
|
const int tid = threadIdx.x;
|
|
|
|
const int il = tid/8;
|
|
|
|
const int ir = tid%8;
|
|
|
|
const int is = 2*il;
|
|
|
|
const int n = 4;
|
|
|
|
|
2023-09-30 16:12:57 +00:00
|
|
|
dst_t * y = yy + i*QK_K + 64*il + n*ir;
|
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
|
|
|
|
2023-08-25 09:09:42 +00:00
|
|
|
const float dall = __low2half(x[i].dm);
|
|
|
|
const float dmin = __high2half(x[i].dm);
|
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 uint8_t * q = x[i].qs + 32*il + n*ir;
|
|
|
|
|
|
|
|
uint8_t sc, m;
|
|
|
|
get_scale_min_k4(is + 0, x[i].scales, sc, m);
|
|
|
|
const float d1 = dall * sc; const float m1 = dmin * m;
|
|
|
|
get_scale_min_k4(is + 1, x[i].scales, sc, m);
|
|
|
|
const float d2 = dall * sc; const float m2 = dmin * m;
|
|
|
|
for (int l = 0; l < n; ++l) {
|
|
|
|
y[l + 0] = d1 * (q[l] & 0xF) - m1;
|
|
|
|
y[l +32] = d2 * (q[l] >> 4) - m2;
|
|
|
|
}
|
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
|
|
|
|
const int tid = threadIdx.x;
|
|
|
|
const uint8_t * q = x[i].qs;
|
2023-09-30 16:12:57 +00:00
|
|
|
dst_t * y = yy + i*QK_K;
|
2023-08-27 12:19:59 +00:00
|
|
|
const float d = (float)x[i].dm[0];
|
|
|
|
const float m = (float)x[i].dm[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
|
|
|
y[tid+ 0] = d * (x[i].scales[0] & 0xF) * (q[tid] & 0xF) - m * (x[i].scales[0] >> 4);
|
|
|
|
y[tid+32] = d * (x[i].scales[1] & 0xF) * (q[tid] >> 4) - m * (x[i].scales[1] >> 4);
|
|
|
|
#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
|
|
|
}
|
|
|
|
|
2023-09-30 16:12:57 +00:00
|
|
|
template<typename dst_t>
|
|
|
|
static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
|
2023-06-07 07:59:52 +00:00
|
|
|
const block_q5_K * x = (const block_q5_K *) vx;
|
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 i = blockIdx.x;
|
|
|
|
|
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 : 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
|
|
|
// assume 64 threads - this is very slightly better than the one below
|
|
|
|
const int tid = threadIdx.x;
|
|
|
|
const int il = tid/16; // il is in 0...3
|
|
|
|
const int ir = tid%16; // ir is in 0...15
|
|
|
|
const int is = 2*il; // is is in 0...6
|
|
|
|
|
2023-09-30 16:12:57 +00:00
|
|
|
dst_t * y = yy + i*QK_K + 64*il + 2*ir;
|
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
|
|
|
|
2023-08-25 09:09:42 +00:00
|
|
|
const float dall = __low2half(x[i].dm);
|
|
|
|
const float dmin = __high2half(x[i].dm);
|
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 uint8_t * ql = x[i].qs + 32*il + 2*ir;
|
|
|
|
const uint8_t * qh = x[i].qh + 2*ir;
|
|
|
|
|
|
|
|
uint8_t sc, m;
|
|
|
|
get_scale_min_k4(is + 0, x[i].scales, sc, m);
|
|
|
|
const float d1 = dall * sc; const float m1 = dmin * m;
|
|
|
|
get_scale_min_k4(is + 1, x[i].scales, sc, m);
|
|
|
|
const float d2 = dall * sc; const float m2 = dmin * m;
|
|
|
|
|
|
|
|
uint8_t hm = 1 << (2*il);
|
|
|
|
y[ 0] = d1 * ((ql[ 0] & 0xF) + (qh[ 0] & hm ? 16 : 0)) - m1;
|
|
|
|
y[ 1] = d1 * ((ql[ 1] & 0xF) + (qh[ 1] & hm ? 16 : 0)) - m1;
|
|
|
|
hm <<= 1;
|
|
|
|
y[32] = d2 * ((ql[ 0] >> 4) + (qh[ 0] & hm ? 16 : 0)) - m2;
|
|
|
|
y[33] = d2 * ((ql[ 1] >> 4) + (qh[ 1] & hm ? 16 : 0)) - m2;
|
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
|
|
|
|
const int tid = threadIdx.x;
|
|
|
|
const uint8_t q = x[i].qs[tid];
|
|
|
|
const int im = tid/8; // 0...3
|
|
|
|
const int in = tid%8; // 0...7
|
|
|
|
const int is = tid/16; // 0 or 1
|
|
|
|
const uint8_t h = x[i].qh[in] >> im;
|
|
|
|
const float d = x[i].d;
|
2023-09-30 16:12:57 +00:00
|
|
|
dst_t * y = yy + i*QK_K + tid;
|
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
|
|
|
y[ 0] = d * x[i].scales[is+0] * ((q & 0xF) - ((h >> 0) & 1 ? 0 : 16));
|
|
|
|
y[32] = d * x[i].scales[is+2] * ((q >> 4) - ((h >> 4) & 1 ? 0 : 16));
|
|
|
|
#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
|
|
|
}
|
|
|
|
|
2023-09-30 16:12:57 +00:00
|
|
|
template<typename dst_t>
|
|
|
|
static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
|
2023-06-07 07:59:52 +00:00
|
|
|
const block_q6_K * x = (const block_q6_K *) vx;
|
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 i = blockIdx.x;
|
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 : 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
|
|
|
|
|
|
|
// assume 64 threads - this is very slightly better than the one below
|
|
|
|
const int tid = threadIdx.x;
|
|
|
|
const int ip = tid/32; // ip is 0 or 1
|
|
|
|
const int il = tid - 32*ip; // 0...32
|
|
|
|
const int is = 8*ip + il/16;
|
|
|
|
|
2023-09-30 16:12:57 +00:00
|
|
|
dst_t * y = yy + i*QK_K + 128*ip + il;
|
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 float d = x[i].d;
|
|
|
|
|
|
|
|
const uint8_t * ql = x[i].ql + 64*ip + il;
|
|
|
|
const uint8_t qh = x[i].qh[32*ip + il];
|
|
|
|
const int8_t * sc = x[i].scales + is;
|
|
|
|
|
|
|
|
y[ 0] = d * sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
|
|
|
|
y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32);
|
|
|
|
y[64] = d * sc[4] * ((int8_t)((ql[ 0] >> 4) | (((qh >> 4) & 3) << 4)) - 32);
|
|
|
|
y[96] = d * sc[6] * ((int8_t)((ql[32] >> 4) | (((qh >> 6) & 3) << 4)) - 32);
|
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
|
|
|
|
|
|
|
|
// assume 32 threads
|
|
|
|
const int tid = threadIdx.x;
|
|
|
|
const int ip = tid/16; // 0 or 1
|
|
|
|
const int il = tid - 16*ip; // 0...15
|
|
|
|
|
2023-09-30 16:12:57 +00:00
|
|
|
dst_t * y = yy + i*QK_K + 16*ip + il;
|
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 = x[i].d;
|
|
|
|
|
|
|
|
const uint8_t ql = x[i].ql[16*ip + il];
|
|
|
|
const uint8_t qh = x[i].qh[il] >> (2*ip);
|
|
|
|
const int8_t * sc = x[i].scales;
|
|
|
|
|
|
|
|
y[ 0] = d * sc[ip+0] * ((int8_t)((ql & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
|
|
|
|
y[32] = d * sc[ip+2] * ((int8_t)((ql >> 4) | (((qh >> 4) & 3) << 4)) - 32);
|
|
|
|
#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
|
|
|
}
|
|
|
|
|
2024-01-11 19:39:39 +00:00
|
|
|
static const __device__ uint64_t iq2xxs_grid[256] = {
|
2024-01-08 15:02:32 +00:00
|
|
|
0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
|
|
|
|
0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x08080808082b0808,
|
|
|
|
0x08080808082b082b, 0x08080808082b2b08, 0x08080808082b2b2b, 0x0808080819080819,
|
|
|
|
0x0808080819081908, 0x0808080819190808, 0x0808080819192b08, 0x08080808192b0819,
|
|
|
|
0x08080808192b1908, 0x080808082b080808, 0x080808082b08082b, 0x080808082b082b2b,
|
|
|
|
0x080808082b2b082b, 0x0808081908080819, 0x0808081908081908, 0x0808081908190808,
|
|
|
|
0x0808081908191919, 0x0808081919080808, 0x080808192b081908, 0x080808192b192b08,
|
|
|
|
0x0808082b08080808, 0x0808082b0808082b, 0x0808082b082b082b, 0x0808082b2b08082b,
|
|
|
|
0x0808190808080819, 0x0808190808081908, 0x0808190808190808, 0x08081908082b0819,
|
|
|
|
0x08081908082b1908, 0x0808190819080808, 0x080819081908082b, 0x0808190819082b08,
|
|
|
|
0x08081908192b0808, 0x080819082b080819, 0x080819082b081908, 0x080819082b190808,
|
|
|
|
0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b, 0x0808191908082b08,
|
|
|
|
0x08081919082b0808, 0x080819191908192b, 0x08081919192b2b19, 0x080819192b080808,
|
|
|
|
0x080819192b190819, 0x0808192b08082b19, 0x0808192b08190808, 0x0808192b19080808,
|
|
|
|
0x0808192b2b081908, 0x0808192b2b2b1908, 0x08082b0808080808, 0x08082b0808081919,
|
|
|
|
0x08082b0808082b08, 0x08082b0808191908, 0x08082b08082b2b08, 0x08082b0819080819,
|
|
|
|
0x08082b0819081908, 0x08082b0819190808, 0x08082b081919082b, 0x08082b082b082b08,
|
|
|
|
0x08082b1908081908, 0x08082b1919080808, 0x08082b2b0808082b, 0x08082b2b08191908,
|
|
|
|
0x0819080808080819, 0x0819080808081908, 0x0819080808190808, 0x08190808082b0819,
|
|
|
|
0x0819080819080808, 0x08190808192b0808, 0x081908082b081908, 0x081908082b190808,
|
|
|
|
0x081908082b191919, 0x0819081908080808, 0x0819081908082b08, 0x08190819082b0808,
|
|
|
|
0x0819081919190808, 0x0819081919192b2b, 0x081908192b080808, 0x0819082b082b1908,
|
|
|
|
0x0819082b19081919, 0x0819190808080808, 0x0819190808082b08, 0x08191908082b0808,
|
|
|
|
0x08191908082b1919, 0x0819190819082b19, 0x081919082b080808, 0x0819191908192b08,
|
|
|
|
0x08191919192b082b, 0x0819192b08080808, 0x0819192b0819192b, 0x08192b0808080819,
|
|
|
|
0x08192b0808081908, 0x08192b0808190808, 0x08192b0819080808, 0x08192b082b080819,
|
|
|
|
0x08192b1908080808, 0x08192b1908081919, 0x08192b192b2b0808, 0x08192b2b19190819,
|
|
|
|
0x082b080808080808, 0x082b08080808082b, 0x082b080808082b2b, 0x082b080819081908,
|
|
|
|
0x082b0808192b0819, 0x082b08082b080808, 0x082b08082b08082b, 0x082b0819082b2b19,
|
|
|
|
0x082b081919082b08, 0x082b082b08080808, 0x082b082b0808082b, 0x082b190808080819,
|
|
|
|
0x082b190808081908, 0x082b190808190808, 0x082b190819080808, 0x082b19081919192b,
|
|
|
|
0x082b191908080808, 0x082b191919080819, 0x082b1919192b1908, 0x082b192b2b190808,
|
|
|
|
0x082b2b0808082b08, 0x082b2b08082b0808, 0x082b2b082b191908, 0x082b2b2b19081908,
|
|
|
|
0x1908080808080819, 0x1908080808081908, 0x1908080808190808, 0x1908080808192b08,
|
|
|
|
0x19080808082b0819, 0x19080808082b1908, 0x1908080819080808, 0x1908080819082b08,
|
|
|
|
0x190808081919192b, 0x19080808192b0808, 0x190808082b080819, 0x190808082b081908,
|
|
|
|
0x190808082b190808, 0x1908081908080808, 0x19080819082b0808, 0x19080819192b0819,
|
|
|
|
0x190808192b080808, 0x190808192b081919, 0x1908082b08080819, 0x1908082b08190808,
|
|
|
|
0x1908082b19082b08, 0x1908082b1919192b, 0x1908082b192b2b08, 0x1908190808080808,
|
|
|
|
0x1908190808082b08, 0x19081908082b0808, 0x190819082b080808, 0x190819082b192b19,
|
|
|
|
0x190819190819082b, 0x19081919082b1908, 0x1908192b08080808, 0x19082b0808080819,
|
|
|
|
0x19082b0808081908, 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919,
|
|
|
|
0x19082b1908080808, 0x19082b1919192b08, 0x19082b19192b0819, 0x19082b192b08082b,
|
|
|
|
0x19082b2b19081919, 0x19082b2b2b190808, 0x1919080808080808, 0x1919080808082b08,
|
|
|
|
0x1919080808190819, 0x1919080808192b19, 0x19190808082b0808, 0x191908082b080808,
|
|
|
|
0x191908082b082b08, 0x1919081908081908, 0x191908191908082b, 0x191908192b2b1908,
|
|
|
|
0x1919082b2b190819, 0x191919082b190808, 0x191919082b19082b, 0x1919191908082b2b,
|
|
|
|
0x1919192b08080819, 0x1919192b19191908, 0x19192b0808080808, 0x19192b0808190819,
|
|
|
|
0x19192b0808192b19, 0x19192b08192b1908, 0x19192b1919080808, 0x19192b2b08082b08,
|
|
|
|
0x192b080808081908, 0x192b080808190808, 0x192b080819080808, 0x192b0808192b2b08,
|
|
|
|
0x192b081908080808, 0x192b081919191919, 0x192b082b08192b08, 0x192b082b192b0808,
|
|
|
|
0x192b190808080808, 0x192b190808081919, 0x192b191908190808, 0x192b19190819082b,
|
|
|
|
0x192b19192b081908, 0x192b2b081908082b, 0x2b08080808080808, 0x2b0808080808082b,
|
|
|
|
0x2b08080808082b2b, 0x2b08080819080819, 0x2b0808082b08082b, 0x2b08081908081908,
|
|
|
|
0x2b08081908192b08, 0x2b08081919080808, 0x2b08082b08190819, 0x2b08190808080819,
|
|
|
|
0x2b08190808081908, 0x2b08190808190808, 0x2b08190808191919, 0x2b08190819080808,
|
|
|
|
0x2b081908192b0808, 0x2b08191908080808, 0x2b0819191908192b, 0x2b0819192b191908,
|
|
|
|
0x2b08192b08082b19, 0x2b08192b19080808, 0x2b08192b192b0808, 0x2b082b080808082b,
|
|
|
|
0x2b082b1908081908, 0x2b082b2b08190819, 0x2b19080808081908, 0x2b19080808190808,
|
|
|
|
0x2b190808082b1908, 0x2b19080819080808, 0x2b1908082b2b0819, 0x2b1908190819192b,
|
|
|
|
0x2b1908192b080808, 0x2b19082b19081919, 0x2b19190808080808, 0x2b191908082b082b,
|
|
|
|
0x2b19190819081908, 0x2b19191919190819, 0x2b192b082b080819, 0x2b192b19082b0808,
|
|
|
|
0x2b2b08080808082b, 0x2b2b080819190808, 0x2b2b08082b081919, 0x2b2b081908082b19,
|
|
|
|
0x2b2b082b08080808, 0x2b2b190808192b08, 0x2b2b2b0819190808, 0x2b2b2b1908081908,
|
|
|
|
};
|
|
|
|
|
2024-01-11 19:39:39 +00:00
|
|
|
static const __device__ uint64_t iq2xs_grid[512] = {
|
|
|
|
0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
|
|
|
|
0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b,
|
|
|
|
0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919,
|
|
|
|
0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b,
|
|
|
|
0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919,
|
|
|
|
0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x080808082b080808,
|
|
|
|
0x080808082b08082b, 0x080808082b081919, 0x080808082b082b08, 0x080808082b190819,
|
|
|
|
0x080808082b191908, 0x080808082b192b19, 0x080808082b2b0808, 0x0808081908080819,
|
|
|
|
0x0808081908081908, 0x080808190808192b, 0x0808081908082b19, 0x0808081908190808,
|
|
|
|
0x080808190819082b, 0x0808081908191919, 0x0808081908192b08, 0x0808081908192b2b,
|
|
|
|
0x08080819082b0819, 0x08080819082b1908, 0x0808081919080808, 0x080808191908082b,
|
|
|
|
0x0808081919081919, 0x0808081919082b08, 0x0808081919190819, 0x0808081919191908,
|
|
|
|
0x08080819192b0808, 0x08080819192b2b08, 0x080808192b080819, 0x080808192b081908,
|
|
|
|
0x080808192b190808, 0x0808082b08080808, 0x0808082b0808082b, 0x0808082b08081919,
|
|
|
|
0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908, 0x0808082b082b0808,
|
|
|
|
0x0808082b19080819, 0x0808082b19081908, 0x0808082b19190808, 0x0808082b19191919,
|
|
|
|
0x0808082b2b080808, 0x0808082b2b082b2b, 0x0808190808080819, 0x0808190808081908,
|
|
|
|
0x080819080808192b, 0x0808190808082b19, 0x0808190808190808, 0x080819080819082b,
|
|
|
|
0x0808190808191919, 0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908,
|
|
|
|
0x0808190819080808, 0x080819081908082b, 0x0808190819081919, 0x0808190819082b08,
|
|
|
|
0x0808190819190819, 0x0808190819191908, 0x080819081919192b, 0x08081908192b0808,
|
|
|
|
0x080819082b080819, 0x080819082b081908, 0x080819082b190808, 0x0808191908080808,
|
|
|
|
0x080819190808082b, 0x0808191908081919, 0x0808191908082b08, 0x0808191908190819,
|
|
|
|
0x0808191908191908, 0x08081919082b0808, 0x0808191919080819, 0x0808191919081908,
|
|
|
|
0x0808191919190808, 0x08081919192b0819, 0x080819192b080808, 0x0808192b08080819,
|
|
|
|
0x0808192b08081908, 0x0808192b08190808, 0x0808192b082b192b, 0x0808192b19080808,
|
|
|
|
0x0808192b1908082b, 0x0808192b2b081908, 0x08082b0808080808, 0x08082b080808082b,
|
|
|
|
0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808082b2b, 0x08082b0808190819,
|
|
|
|
0x08082b0808191908, 0x08082b08082b0808, 0x08082b08082b1919, 0x08082b0819080819,
|
|
|
|
0x08082b0819081908, 0x08082b0819190808, 0x08082b0819192b08, 0x08082b082b080808,
|
|
|
|
0x08082b082b2b0808, 0x08082b082b2b2b2b, 0x08082b1908080819, 0x08082b1908081908,
|
|
|
|
0x08082b1908190808, 0x08082b1919080808, 0x08082b192b080819, 0x08082b192b082b19,
|
|
|
|
0x08082b2b08080808, 0x08082b2b082b0808, 0x08082b2b082b2b08, 0x08082b2b2b19192b,
|
|
|
|
0x08082b2b2b2b0808, 0x0819080808080819, 0x0819080808081908, 0x081908080808192b,
|
|
|
|
0x0819080808082b19, 0x0819080808190808, 0x081908080819082b, 0x0819080808191919,
|
|
|
|
0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908, 0x0819080819080808,
|
|
|
|
0x081908081908082b, 0x0819080819081919, 0x0819080819082b08, 0x0819080819190819,
|
|
|
|
0x0819080819191908, 0x08190808192b0808, 0x08190808192b2b2b, 0x081908082b080819,
|
|
|
|
0x081908082b081908, 0x081908082b190808, 0x0819081908080808, 0x081908190808082b,
|
|
|
|
0x0819081908081919, 0x0819081908082b08, 0x0819081908190819, 0x0819081908191908,
|
|
|
|
0x08190819082b0808, 0x0819081919080819, 0x0819081919081908, 0x0819081919190808,
|
|
|
|
0x081908192b080808, 0x081908192b191908, 0x081908192b19192b, 0x0819082b08080819,
|
|
|
|
0x0819082b08081908, 0x0819082b0808192b, 0x0819082b08190808, 0x0819082b19080808,
|
|
|
|
0x0819082b192b0808, 0x0819190808080808, 0x081919080808082b, 0x0819190808081919,
|
|
|
|
0x0819190808082b08, 0x0819190808190819, 0x0819190808191908, 0x08191908082b0808,
|
|
|
|
0x0819190819080819, 0x0819190819081908, 0x0819190819082b19, 0x0819190819190808,
|
|
|
|
0x08191908192b1908, 0x081919082b080808, 0x0819191908080819, 0x0819191908081908,
|
|
|
|
0x0819191908190808, 0x0819191919080808, 0x0819192b08080808, 0x0819192b08191908,
|
|
|
|
0x0819192b19082b19, 0x08192b0808080819, 0x08192b0808081908, 0x08192b0808190808,
|
|
|
|
0x08192b080819082b, 0x08192b0819080808, 0x08192b0819191908, 0x08192b082b08192b,
|
|
|
|
0x08192b1908080808, 0x08192b1908081919, 0x08192b19192b192b, 0x08192b2b19190819,
|
|
|
|
0x08192b2b2b2b2b19, 0x082b080808080808, 0x082b08080808082b, 0x082b080808081919,
|
|
|
|
0x082b080808082b08, 0x082b080808082b2b, 0x082b080808190819, 0x082b080808191908,
|
|
|
|
0x082b0808082b0808, 0x082b080819080819, 0x082b080819081908, 0x082b080819190808,
|
|
|
|
0x082b08082b080808, 0x082b08082b2b0808, 0x082b081908080819, 0x082b081908081908,
|
|
|
|
0x082b081908190808, 0x082b081919080808, 0x082b081919082b08, 0x082b0819192b1919,
|
|
|
|
0x082b082b08080808, 0x082b082b082b082b, 0x082b082b2b080808, 0x082b082b2b2b2b08,
|
|
|
|
0x082b190808080819, 0x082b190808081908, 0x082b190808190808, 0x082b1908082b2b19,
|
|
|
|
0x082b190819080808, 0x082b191908080808, 0x082b191919080819, 0x082b19191919082b,
|
|
|
|
0x082b19192b192b19, 0x082b192b08080819, 0x082b192b08192b2b, 0x082b192b2b2b192b,
|
|
|
|
0x082b2b0808080808, 0x082b2b0808082b08, 0x082b2b0808082b2b, 0x082b2b08082b0808,
|
|
|
|
0x082b2b0819191919, 0x082b2b082b082b08, 0x082b2b082b2b082b, 0x082b2b19192b2b08,
|
|
|
|
0x082b2b192b190808, 0x082b2b2b08082b08, 0x082b2b2b082b0808, 0x082b2b2b2b08082b,
|
|
|
|
0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819, 0x1908080808081908,
|
|
|
|
0x190808080808192b, 0x1908080808082b19, 0x1908080808190808, 0x190808080819082b,
|
|
|
|
0x1908080808191919, 0x1908080808192b08, 0x19080808082b0819, 0x19080808082b1908,
|
|
|
|
0x1908080819080808, 0x190808081908082b, 0x1908080819081919, 0x1908080819082b08,
|
|
|
|
0x1908080819082b2b, 0x1908080819190819, 0x1908080819191908, 0x19080808192b0808,
|
|
|
|
0x19080808192b1919, 0x190808082b080819, 0x190808082b081908, 0x190808082b190808,
|
|
|
|
0x1908081908080808, 0x190808190808082b, 0x1908081908081919, 0x1908081908082b08,
|
|
|
|
0x1908081908190819, 0x1908081908191908, 0x19080819082b0808, 0x1908081919080819,
|
|
|
|
0x1908081919081908, 0x1908081919190808, 0x190808192b080808, 0x190808192b081919,
|
|
|
|
0x190808192b2b082b, 0x1908082b08080819, 0x1908082b08081908, 0x1908082b08190808,
|
|
|
|
0x1908082b0819082b, 0x1908082b082b2b19, 0x1908082b19080808, 0x1908190808080808,
|
|
|
|
0x190819080808082b, 0x1908190808081919, 0x1908190808082b08, 0x1908190808190819,
|
|
|
|
0x1908190808191908, 0x1908190808192b19, 0x19081908082b0808, 0x1908190819080819,
|
|
|
|
0x1908190819081908, 0x1908190819190808, 0x190819082b080808, 0x190819082b191908,
|
|
|
|
0x1908191908080819, 0x1908191908081908, 0x1908191908190808, 0x19081919082b1908,
|
|
|
|
0x1908191919080808, 0x190819192b192b2b, 0x1908192b08080808, 0x1908192b08082b2b,
|
|
|
|
0x1908192b19081908, 0x1908192b19190808, 0x19082b0808080819, 0x19082b0808081908,
|
|
|
|
0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919, 0x19082b0819191908,
|
|
|
|
0x19082b08192b082b, 0x19082b1908080808, 0x19082b1908190819, 0x19082b1919081908,
|
|
|
|
0x19082b1919190808, 0x19082b19192b2b19, 0x19082b2b08081908, 0x1919080808080808,
|
|
|
|
0x191908080808082b, 0x1919080808081919, 0x1919080808082b08, 0x1919080808190819,
|
|
|
|
0x1919080808191908, 0x19190808082b0808, 0x19190808082b2b08, 0x1919080819080819,
|
|
|
|
0x1919080819081908, 0x1919080819190808, 0x191908082b080808, 0x1919081908080819,
|
|
|
|
0x1919081908081908, 0x1919081908190808, 0x1919081908191919, 0x1919081919080808,
|
|
|
|
0x191908191908082b, 0x1919082b08080808, 0x1919082b19081908, 0x1919082b2b2b2b2b,
|
|
|
|
0x1919190808080819, 0x1919190808081908, 0x1919190808190808, 0x19191908082b0819,
|
|
|
|
0x1919190819080808, 0x19191908192b0808, 0x191919082b080819, 0x191919082b2b0819,
|
|
|
|
0x1919191908080808, 0x1919191908082b08, 0x191919192b080808, 0x191919192b082b08,
|
|
|
|
0x1919192b082b0819, 0x1919192b192b2b08, 0x1919192b2b2b0819, 0x19192b0808080808,
|
|
|
|
0x19192b0808191908, 0x19192b0819080819, 0x19192b0819190808, 0x19192b082b192b19,
|
|
|
|
0x19192b1908192b2b, 0x19192b1919080808, 0x19192b191908082b, 0x19192b2b2b081919,
|
|
|
|
0x192b080808080819, 0x192b080808081908, 0x192b080808190808, 0x192b080819080808,
|
|
|
|
0x192b080819191908, 0x192b0808192b082b, 0x192b08082b08192b, 0x192b08082b2b2b19,
|
|
|
|
0x192b081908080808, 0x192b082b082b1908, 0x192b082b19082b2b, 0x192b082b2b19082b,
|
|
|
|
0x192b190808080808, 0x192b19080819192b, 0x192b191908190808, 0x192b191919080808,
|
|
|
|
0x192b191919081919, 0x192b19192b2b1908, 0x192b2b0808080819, 0x192b2b08192b2b2b,
|
|
|
|
0x192b2b19082b1919, 0x192b2b2b0808192b, 0x192b2b2b19191908, 0x192b2b2b192b082b,
|
|
|
|
0x2b08080808080808, 0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08,
|
|
|
|
0x2b08080808190819, 0x2b08080808191908, 0x2b080808082b0808, 0x2b080808082b2b2b,
|
|
|
|
0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808082b080808,
|
|
|
|
0x2b0808082b08082b, 0x2b0808082b2b2b08, 0x2b0808082b2b2b2b, 0x2b08081908080819,
|
|
|
|
0x2b08081908081908, 0x2b0808190808192b, 0x2b08081908190808, 0x2b08081919080808,
|
|
|
|
0x2b08081919190819, 0x2b08081919192b19, 0x2b08082b08080808, 0x2b08082b082b0808,
|
|
|
|
0x2b08082b2b080808, 0x2b08082b2b08082b, 0x2b08082b2b2b0808, 0x2b08082b2b2b2b08,
|
|
|
|
0x2b08190808080819, 0x2b08190808081908, 0x2b08190808190808, 0x2b0819080819082b,
|
|
|
|
0x2b08190808191919, 0x2b08190819080808, 0x2b081908192b0808, 0x2b0819082b082b19,
|
|
|
|
0x2b08191908080808, 0x2b08191919081908, 0x2b0819192b2b1919, 0x2b08192b08192b08,
|
|
|
|
0x2b08192b192b2b2b, 0x2b082b0808080808, 0x2b082b0808082b08, 0x2b082b08082b1919,
|
|
|
|
0x2b082b0819192b2b, 0x2b082b082b080808, 0x2b082b082b08082b, 0x2b082b082b2b2b08,
|
|
|
|
0x2b082b190808192b, 0x2b082b2b082b082b, 0x2b082b2b2b080808, 0x2b082b2b2b082b08,
|
|
|
|
0x2b082b2b2b19192b, 0x2b082b2b2b2b2b08, 0x2b19080808080819, 0x2b19080808081908,
|
|
|
|
0x2b19080808190808, 0x2b19080819080808, 0x2b1908081919192b, 0x2b1908082b081908,
|
|
|
|
0x2b19081908080808, 0x2b190819082b082b, 0x2b190819192b1908, 0x2b19082b1919192b,
|
|
|
|
0x2b19082b2b082b19, 0x2b19190808080808, 0x2b19190808081919, 0x2b19190819081908,
|
|
|
|
0x2b19190819190808, 0x2b19190819192b08, 0x2b191919082b2b19, 0x2b1919192b190808,
|
|
|
|
0x2b1919192b19082b, 0x2b19192b19080819, 0x2b192b0819190819, 0x2b192b082b2b192b,
|
|
|
|
0x2b192b1919082b19, 0x2b192b2b08191919, 0x2b192b2b192b0808, 0x2b2b080808080808,
|
|
|
|
0x2b2b08080808082b, 0x2b2b080808082b08, 0x2b2b080808082b2b, 0x2b2b0808082b0808,
|
|
|
|
0x2b2b0808082b2b2b, 0x2b2b08082b2b0808, 0x2b2b081919190819, 0x2b2b081919192b19,
|
|
|
|
0x2b2b08192b2b192b, 0x2b2b082b08080808, 0x2b2b082b0808082b, 0x2b2b082b08082b08,
|
|
|
|
0x2b2b082b082b2b2b, 0x2b2b082b2b080808, 0x2b2b082b2b2b0808, 0x2b2b190819080808,
|
|
|
|
0x2b2b19082b191919, 0x2b2b192b192b1919, 0x2b2b192b2b192b08, 0x2b2b2b0808082b2b,
|
|
|
|
0x2b2b2b08082b0808, 0x2b2b2b08082b082b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b0808,
|
|
|
|
0x2b2b2b082b2b2b08, 0x2b2b2b1908081908, 0x2b2b2b192b081908, 0x2b2b2b192b08192b,
|
|
|
|
0x2b2b2b2b082b2b08, 0x2b2b2b2b082b2b2b, 0x2b2b2b2b2b190819, 0x2b2b2b2b2b2b2b2b,
|
|
|
|
};
|
|
|
|
|
2024-02-26 16:28:38 +00:00
|
|
|
static const __device__ uint64_t iq2s_grid[1024] = {
|
|
|
|
0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
|
|
|
|
0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b,
|
|
|
|
0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919,
|
|
|
|
0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b,
|
|
|
|
0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919,
|
|
|
|
0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x08080808192b192b,
|
|
|
|
0x08080808192b2b19, 0x080808082b080808, 0x080808082b08082b, 0x080808082b081919,
|
|
|
|
0x080808082b082b08, 0x080808082b190819, 0x080808082b191908, 0x080808082b2b0808,
|
|
|
|
0x080808082b2b1919, 0x080808082b2b2b2b, 0x0808081908080819, 0x0808081908081908,
|
|
|
|
0x080808190808192b, 0x0808081908082b19, 0x0808081908190808, 0x080808190819082b,
|
|
|
|
0x0808081908191919, 0x0808081908192b08, 0x08080819082b0819, 0x08080819082b1908,
|
|
|
|
0x0808081919080808, 0x080808191908082b, 0x0808081919081919, 0x0808081919082b08,
|
|
|
|
0x0808081919190819, 0x0808081919191908, 0x080808191919192b, 0x0808081919192b19,
|
|
|
|
0x08080819192b0808, 0x08080819192b1919, 0x08080819192b2b08, 0x080808192b080819,
|
|
|
|
0x080808192b081908, 0x080808192b190808, 0x080808192b19082b, 0x080808192b191919,
|
|
|
|
0x080808192b2b0819, 0x080808192b2b1908, 0x0808082b08080808, 0x0808082b0808082b,
|
|
|
|
0x0808082b08081919, 0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908,
|
|
|
|
0x0808082b082b0808, 0x0808082b082b2b2b, 0x0808082b19080819, 0x0808082b19081908,
|
|
|
|
0x0808082b1908192b, 0x0808082b19082b19, 0x0808082b19190808, 0x0808082b19191919,
|
|
|
|
0x0808082b2b080808, 0x0808082b2b081919, 0x0808082b2b082b2b, 0x0808082b2b191908,
|
|
|
|
0x0808082b2b2b082b, 0x0808190808080819, 0x0808190808081908, 0x080819080808192b,
|
|
|
|
0x0808190808082b19, 0x0808190808190808, 0x080819080819082b, 0x0808190808191919,
|
|
|
|
0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908, 0x08081908082b192b,
|
|
|
|
0x08081908082b2b19, 0x0808190819080808, 0x080819081908082b, 0x0808190819081919,
|
|
|
|
0x0808190819082b08, 0x0808190819082b2b, 0x0808190819190819, 0x0808190819191908,
|
|
|
|
0x080819081919192b, 0x0808190819192b19, 0x08081908192b0808, 0x08081908192b082b,
|
|
|
|
0x08081908192b1919, 0x080819082b080819, 0x080819082b081908, 0x080819082b08192b,
|
|
|
|
0x080819082b082b19, 0x080819082b190808, 0x080819082b191919, 0x080819082b192b08,
|
|
|
|
0x080819082b2b0819, 0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b,
|
|
|
|
0x0808191908081919, 0x0808191908082b08, 0x0808191908082b2b, 0x0808191908190819,
|
|
|
|
0x0808191908191908, 0x080819190819192b, 0x0808191908192b19, 0x08081919082b0808,
|
|
|
|
0x08081919082b1919, 0x08081919082b2b08, 0x0808191919080819, 0x0808191919081908,
|
|
|
|
0x080819191908192b, 0x0808191919082b19, 0x0808191919190808, 0x080819191919082b,
|
|
|
|
0x0808191919191919, 0x0808191919192b08, 0x08081919192b0819, 0x08081919192b1908,
|
|
|
|
0x080819192b080808, 0x080819192b08082b, 0x080819192b081919, 0x080819192b082b08,
|
|
|
|
0x080819192b190819, 0x080819192b191908, 0x080819192b2b0808, 0x0808192b08080819,
|
|
|
|
0x0808192b08081908, 0x0808192b0808192b, 0x0808192b08082b19, 0x0808192b08190808,
|
|
|
|
0x0808192b08191919, 0x0808192b19080808, 0x0808192b19081919, 0x0808192b19082b08,
|
|
|
|
0x0808192b19190819, 0x0808192b19191908, 0x0808192b192b0808, 0x0808192b2b080819,
|
|
|
|
0x0808192b2b081908, 0x0808192b2b190808, 0x08082b0808080808, 0x08082b080808082b,
|
|
|
|
0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808190819, 0x08082b0808191908,
|
|
|
|
0x08082b080819192b, 0x08082b0808192b19, 0x08082b08082b0808, 0x08082b08082b1919,
|
|
|
|
0x08082b08082b2b2b, 0x08082b0819080819, 0x08082b0819081908, 0x08082b081908192b,
|
|
|
|
0x08082b0819082b19, 0x08082b0819190808, 0x08082b081919082b, 0x08082b0819191919,
|
|
|
|
0x08082b0819192b08, 0x08082b08192b0819, 0x08082b08192b1908, 0x08082b082b080808,
|
|
|
|
0x08082b082b081919, 0x08082b082b191908, 0x08082b082b2b2b2b, 0x08082b1908080819,
|
|
|
|
0x08082b1908081908, 0x08082b1908190808, 0x08082b190819082b, 0x08082b1908191919,
|
|
|
|
0x08082b1908192b08, 0x08082b19082b0819, 0x08082b1919080808, 0x08082b1919081919,
|
|
|
|
0x08082b1919082b08, 0x08082b1919190819, 0x08082b1919191908, 0x08082b19192b0808,
|
|
|
|
0x08082b192b080819, 0x08082b192b190808, 0x08082b2b08080808, 0x08082b2b08190819,
|
|
|
|
0x08082b2b08191908, 0x08082b2b082b082b, 0x08082b2b082b2b08, 0x08082b2b082b2b2b,
|
|
|
|
0x08082b2b19190808, 0x08082b2b2b192b19, 0x0819080808080819, 0x0819080808081908,
|
|
|
|
0x081908080808192b, 0x0819080808082b19, 0x0819080808190808, 0x081908080819082b,
|
|
|
|
0x0819080808191919, 0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908,
|
|
|
|
0x08190808082b192b, 0x0819080819080808, 0x081908081908082b, 0x0819080819081919,
|
|
|
|
0x0819080819082b08, 0x0819080819190819, 0x0819080819191908, 0x081908081919192b,
|
|
|
|
0x0819080819192b19, 0x08190808192b0808, 0x08190808192b082b, 0x08190808192b1919,
|
|
|
|
0x08190808192b2b08, 0x081908082b080819, 0x081908082b081908, 0x081908082b08192b,
|
|
|
|
0x081908082b190808, 0x081908082b191919, 0x081908082b192b08, 0x081908082b2b0819,
|
|
|
|
0x081908082b2b1908, 0x0819081908080808, 0x081908190808082b, 0x0819081908081919,
|
|
|
|
0x0819081908082b08, 0x0819081908082b2b, 0x0819081908190819, 0x0819081908191908,
|
|
|
|
0x081908190819192b, 0x0819081908192b19, 0x08190819082b0808, 0x08190819082b082b,
|
|
|
|
0x08190819082b1919, 0x08190819082b2b08, 0x0819081919080819, 0x0819081919081908,
|
|
|
|
0x081908191908192b, 0x0819081919082b19, 0x0819081919190808, 0x081908191919082b,
|
|
|
|
0x0819081919191919, 0x0819081919192b08, 0x08190819192b0819, 0x08190819192b1908,
|
|
|
|
0x081908192b080808, 0x081908192b08082b, 0x081908192b081919, 0x081908192b082b08,
|
|
|
|
0x081908192b190819, 0x081908192b191908, 0x0819082b08080819, 0x0819082b08081908,
|
|
|
|
0x0819082b08082b19, 0x0819082b08190808, 0x0819082b08191919, 0x0819082b082b0819,
|
|
|
|
0x0819082b082b1908, 0x0819082b19080808, 0x0819082b19081919, 0x0819082b19190819,
|
|
|
|
0x0819082b19191908, 0x0819082b2b080819, 0x0819082b2b081908, 0x0819082b2b190808,
|
|
|
|
0x0819190808080808, 0x081919080808082b, 0x0819190808081919, 0x0819190808082b08,
|
|
|
|
0x0819190808190819, 0x0819190808191908, 0x081919080819192b, 0x0819190808192b19,
|
|
|
|
0x08191908082b0808, 0x08191908082b1919, 0x08191908082b2b08, 0x0819190819080819,
|
|
|
|
0x0819190819081908, 0x081919081908192b, 0x0819190819082b19, 0x0819190819190808,
|
|
|
|
0x081919081919082b, 0x0819190819191919, 0x0819190819192b08, 0x08191908192b0819,
|
|
|
|
0x08191908192b1908, 0x081919082b080808, 0x081919082b08082b, 0x081919082b081919,
|
|
|
|
0x081919082b082b08, 0x081919082b190819, 0x081919082b191908, 0x081919082b2b0808,
|
|
|
|
0x0819191908080819, 0x0819191908081908, 0x081919190808192b, 0x0819191908082b19,
|
|
|
|
0x0819191908190808, 0x081919190819082b, 0x0819191908191919, 0x0819191908192b08,
|
|
|
|
0x08191919082b0819, 0x08191919082b1908, 0x0819191919080808, 0x081919191908082b,
|
|
|
|
0x0819191919081919, 0x0819191919082b08, 0x0819191919190819, 0x0819191919191908,
|
|
|
|
0x08191919192b0808, 0x081919192b080819, 0x081919192b081908, 0x081919192b190808,
|
|
|
|
0x0819192b08080808, 0x0819192b08081919, 0x0819192b08082b08, 0x0819192b08190819,
|
|
|
|
0x0819192b08191908, 0x0819192b082b0808, 0x0819192b19080819, 0x0819192b19081908,
|
|
|
|
0x0819192b19190808, 0x0819192b2b080808, 0x0819192b2b2b2b2b, 0x08192b0808080819,
|
|
|
|
0x08192b0808081908, 0x08192b080808192b, 0x08192b0808082b19, 0x08192b0808190808,
|
|
|
|
0x08192b0808191919, 0x08192b0808192b08, 0x08192b08082b0819, 0x08192b0819080808,
|
|
|
|
0x08192b081908082b, 0x08192b0819081919, 0x08192b0819082b08, 0x08192b0819190819,
|
|
|
|
0x08192b0819191908, 0x08192b08192b0808, 0x08192b082b080819, 0x08192b082b081908,
|
|
|
|
0x08192b1908080808, 0x08192b190808082b, 0x08192b1908081919, 0x08192b1908082b08,
|
|
|
|
0x08192b1908190819, 0x08192b1908191908, 0x08192b19082b0808, 0x08192b1919080819,
|
|
|
|
0x08192b1919081908, 0x08192b1919190808, 0x08192b19192b2b19, 0x08192b192b2b082b,
|
|
|
|
0x08192b2b08081908, 0x08192b2b08190808, 0x08192b2b19080808, 0x08192b2b1919192b,
|
|
|
|
0x082b080808080808, 0x082b08080808082b, 0x082b080808081919, 0x082b080808082b08,
|
|
|
|
0x082b080808190819, 0x082b080808191908, 0x082b08080819192b, 0x082b080808192b19,
|
|
|
|
0x082b0808082b0808, 0x082b0808082b1919, 0x082b0808082b2b2b, 0x082b080819080819,
|
|
|
|
0x082b080819081908, 0x082b080819190808, 0x082b08081919082b, 0x082b080819191919,
|
|
|
|
0x082b0808192b1908, 0x082b08082b080808, 0x082b08082b082b2b, 0x082b08082b191908,
|
|
|
|
0x082b08082b2b2b2b, 0x082b081908080819, 0x082b081908081908, 0x082b081908190808,
|
|
|
|
0x082b08190819082b, 0x082b081908191919, 0x082b0819082b0819, 0x082b081919080808,
|
|
|
|
0x082b08191908082b, 0x082b081919081919, 0x082b081919190819, 0x082b081919191908,
|
|
|
|
0x082b0819192b0808, 0x082b08192b080819, 0x082b08192b081908, 0x082b08192b190808,
|
|
|
|
0x082b082b08080808, 0x082b082b08082b2b, 0x082b082b082b082b, 0x082b082b082b2b08,
|
|
|
|
0x082b082b082b2b2b, 0x082b082b19081908, 0x082b082b19190808, 0x082b082b2b082b08,
|
|
|
|
0x082b082b2b082b2b, 0x082b082b2b2b2b08, 0x082b190808080819, 0x082b190808081908,
|
|
|
|
0x082b19080808192b, 0x082b190808082b19, 0x082b190808190808, 0x082b190808191919,
|
|
|
|
0x082b190808192b08, 0x082b1908082b0819, 0x082b1908082b1908, 0x082b190819080808,
|
|
|
|
0x082b19081908082b, 0x082b190819081919, 0x082b190819082b08, 0x082b190819190819,
|
|
|
|
0x082b190819191908, 0x082b1908192b0808, 0x082b19082b080819, 0x082b19082b081908,
|
|
|
|
0x082b19082b190808, 0x082b191908080808, 0x082b191908081919, 0x082b191908082b08,
|
|
|
|
0x082b191908190819, 0x082b191908191908, 0x082b1919082b0808, 0x082b191919080819,
|
|
|
|
0x082b191919081908, 0x082b191919190808, 0x082b1919192b192b, 0x082b19192b080808,
|
|
|
|
0x082b192b08080819, 0x082b192b08081908, 0x082b192b08190808, 0x082b192b19080808,
|
|
|
|
0x082b192b19192b19, 0x082b2b0808080808, 0x082b2b0808081919, 0x082b2b0808190819,
|
|
|
|
0x082b2b0808191908, 0x082b2b0819080819, 0x082b2b0819081908, 0x082b2b0819190808,
|
|
|
|
0x082b2b082b082b2b, 0x082b2b082b2b2b2b, 0x082b2b1908080819, 0x082b2b1908081908,
|
|
|
|
0x082b2b1908190808, 0x082b2b192b191919, 0x082b2b2b08082b2b, 0x082b2b2b082b082b,
|
|
|
|
0x082b2b2b192b1908, 0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819,
|
|
|
|
0x1908080808081908, 0x190808080808192b, 0x1908080808082b19, 0x1908080808190808,
|
|
|
|
0x190808080819082b, 0x1908080808191919, 0x1908080808192b08, 0x1908080808192b2b,
|
|
|
|
0x19080808082b0819, 0x19080808082b1908, 0x19080808082b192b, 0x1908080819080808,
|
|
|
|
0x190808081908082b, 0x1908080819081919, 0x1908080819082b08, 0x1908080819082b2b,
|
|
|
|
0x1908080819190819, 0x1908080819191908, 0x190808081919192b, 0x1908080819192b19,
|
|
|
|
0x19080808192b0808, 0x19080808192b082b, 0x19080808192b1919, 0x190808082b080819,
|
|
|
|
0x190808082b081908, 0x190808082b190808, 0x190808082b191919, 0x190808082b192b08,
|
|
|
|
0x190808082b2b0819, 0x190808082b2b1908, 0x1908081908080808, 0x190808190808082b,
|
|
|
|
0x1908081908081919, 0x1908081908082b08, 0x1908081908190819, 0x1908081908191908,
|
|
|
|
0x190808190819192b, 0x1908081908192b19, 0x19080819082b0808, 0x19080819082b082b,
|
|
|
|
0x19080819082b1919, 0x1908081919080819, 0x1908081919081908, 0x190808191908192b,
|
|
|
|
0x1908081919082b19, 0x1908081919190808, 0x190808191919082b, 0x1908081919191919,
|
|
|
|
0x1908081919192b08, 0x19080819192b0819, 0x19080819192b1908, 0x190808192b080808,
|
|
|
|
0x190808192b08082b, 0x190808192b081919, 0x190808192b082b08, 0x190808192b190819,
|
|
|
|
0x190808192b191908, 0x190808192b2b0808, 0x1908082b08080819, 0x1908082b08081908,
|
|
|
|
0x1908082b08190808, 0x1908082b0819082b, 0x1908082b08191919, 0x1908082b08192b08,
|
|
|
|
0x1908082b082b1908, 0x1908082b19080808, 0x1908082b19081919, 0x1908082b19082b08,
|
|
|
|
0x1908082b19190819, 0x1908082b19191908, 0x1908082b192b0808, 0x1908082b2b080819,
|
|
|
|
0x1908082b2b081908, 0x1908190808080808, 0x190819080808082b, 0x1908190808081919,
|
|
|
|
0x1908190808082b08, 0x1908190808082b2b, 0x1908190808190819, 0x1908190808191908,
|
|
|
|
0x190819080819192b, 0x1908190808192b19, 0x19081908082b0808, 0x19081908082b082b,
|
|
|
|
0x19081908082b1919, 0x19081908082b2b08, 0x1908190819080819, 0x1908190819081908,
|
|
|
|
0x190819081908192b, 0x1908190819082b19, 0x1908190819190808, 0x190819081919082b,
|
|
|
|
0x1908190819191919, 0x1908190819192b08, 0x19081908192b0819, 0x19081908192b1908,
|
|
|
|
0x190819082b080808, 0x190819082b08082b, 0x190819082b081919, 0x190819082b082b08,
|
|
|
|
0x190819082b190819, 0x190819082b191908, 0x190819082b2b0808, 0x1908191908080819,
|
|
|
|
0x1908191908081908, 0x190819190808192b, 0x1908191908082b19, 0x1908191908190808,
|
|
|
|
0x190819190819082b, 0x1908191908191919, 0x1908191908192b08, 0x19081919082b0819,
|
|
|
|
0x19081919082b1908, 0x1908191919080808, 0x190819191908082b, 0x1908191919081919,
|
|
|
|
0x1908191919082b08, 0x1908191919190819, 0x1908191919191908, 0x19081919192b0808,
|
|
|
|
0x19081919192b2b2b, 0x190819192b080819, 0x190819192b081908, 0x190819192b190808,
|
|
|
|
0x1908192b08080808, 0x1908192b0808082b, 0x1908192b08081919, 0x1908192b08082b08,
|
|
|
|
0x1908192b08190819, 0x1908192b08191908, 0x1908192b082b0808, 0x1908192b19080819,
|
|
|
|
0x1908192b19081908, 0x1908192b19190808, 0x1908192b2b080808, 0x1908192b2b2b1919,
|
|
|
|
0x19082b0808080819, 0x19082b0808081908, 0x19082b0808082b19, 0x19082b0808190808,
|
|
|
|
0x19082b080819082b, 0x19082b0808191919, 0x19082b0808192b08, 0x19082b08082b0819,
|
|
|
|
0x19082b08082b1908, 0x19082b0819080808, 0x19082b081908082b, 0x19082b0819081919,
|
|
|
|
0x19082b0819082b08, 0x19082b0819190819, 0x19082b0819191908, 0x19082b08192b0808,
|
|
|
|
0x19082b082b081908, 0x19082b082b190808, 0x19082b1908080808, 0x19082b190808082b,
|
|
|
|
0x19082b1908081919, 0x19082b1908082b08, 0x19082b1908190819, 0x19082b1908191908,
|
|
|
|
0x19082b19082b0808, 0x19082b1919080819, 0x19082b1919081908, 0x19082b1919190808,
|
|
|
|
0x19082b192b080808, 0x19082b192b19192b, 0x19082b2b08080819, 0x19082b2b08081908,
|
|
|
|
0x19082b2b08190808, 0x19082b2b19080808, 0x1919080808080808, 0x191908080808082b,
|
|
|
|
0x1919080808081919, 0x1919080808082b08, 0x1919080808190819, 0x1919080808191908,
|
|
|
|
0x191908080819192b, 0x1919080808192b19, 0x19190808082b0808, 0x19190808082b082b,
|
|
|
|
0x19190808082b1919, 0x19190808082b2b08, 0x1919080819080819, 0x1919080819081908,
|
|
|
|
0x191908081908192b, 0x1919080819082b19, 0x1919080819190808, 0x191908081919082b,
|
|
|
|
0x1919080819191919, 0x1919080819192b08, 0x19190808192b0819, 0x19190808192b1908,
|
|
|
|
0x191908082b080808, 0x191908082b08082b, 0x191908082b081919, 0x191908082b082b08,
|
|
|
|
0x191908082b190819, 0x191908082b191908, 0x1919081908080819, 0x1919081908081908,
|
|
|
|
0x191908190808192b, 0x1919081908082b19, 0x1919081908190808, 0x191908190819082b,
|
|
|
|
0x1919081908191919, 0x1919081908192b08, 0x19190819082b0819, 0x19190819082b1908,
|
|
|
|
0x1919081919080808, 0x191908191908082b, 0x1919081919081919, 0x1919081919082b08,
|
|
|
|
0x1919081919190819, 0x1919081919191908, 0x19190819192b0808, 0x191908192b080819,
|
|
|
|
0x191908192b081908, 0x191908192b190808, 0x1919082b08080808, 0x1919082b08081919,
|
|
|
|
0x1919082b08082b08, 0x1919082b08190819, 0x1919082b08191908, 0x1919082b082b0808,
|
|
|
|
0x1919082b19080819, 0x1919082b19081908, 0x1919082b19190808, 0x1919082b192b2b19,
|
|
|
|
0x1919082b2b080808, 0x1919190808080819, 0x1919190808081908, 0x191919080808192b,
|
|
|
|
0x1919190808082b19, 0x1919190808190808, 0x191919080819082b, 0x1919190808191919,
|
|
|
|
0x1919190808192b08, 0x19191908082b0819, 0x19191908082b1908, 0x1919190819080808,
|
|
|
|
0x191919081908082b, 0x1919190819081919, 0x1919190819082b08, 0x1919190819190819,
|
|
|
|
0x1919190819191908, 0x19191908192b0808, 0x191919082b080819, 0x191919082b081908,
|
|
|
|
0x191919082b190808, 0x1919191908080808, 0x191919190808082b, 0x1919191908081919,
|
|
|
|
0x1919191908082b08, 0x1919191908190819, 0x1919191908191908, 0x19191919082b0808,
|
|
|
|
0x1919191919080819, 0x1919191919081908, 0x1919191919190808, 0x191919192b080808,
|
|
|
|
0x1919192b08080819, 0x1919192b08081908, 0x1919192b08190808, 0x1919192b082b192b,
|
|
|
|
0x1919192b19080808, 0x19192b0808080808, 0x19192b080808082b, 0x19192b0808081919,
|
|
|
|
0x19192b0808082b08, 0x19192b0808190819, 0x19192b0808191908, 0x19192b08082b0808,
|
|
|
|
0x19192b0819080819, 0x19192b0819081908, 0x19192b0819190808, 0x19192b0819192b2b,
|
|
|
|
0x19192b082b080808, 0x19192b1908080819, 0x19192b1908081908, 0x19192b1908190808,
|
|
|
|
0x19192b1919080808, 0x19192b2b08080808, 0x19192b2b08192b19, 0x19192b2b2b081919,
|
|
|
|
0x19192b2b2b2b2b08, 0x192b080808080819, 0x192b080808081908, 0x192b08080808192b,
|
|
|
|
0x192b080808190808, 0x192b08080819082b, 0x192b080808191919, 0x192b080808192b08,
|
|
|
|
0x192b0808082b0819, 0x192b0808082b1908, 0x192b080819080808, 0x192b080819081919,
|
|
|
|
0x192b080819082b08, 0x192b080819190819, 0x192b080819191908, 0x192b0808192b0808,
|
|
|
|
0x192b08082b081908, 0x192b08082b190808, 0x192b081908080808, 0x192b08190808082b,
|
|
|
|
0x192b081908081919, 0x192b081908082b08, 0x192b081908190819, 0x192b081908191908,
|
|
|
|
0x192b0819082b0808, 0x192b081919080819, 0x192b081919081908, 0x192b081919190808,
|
|
|
|
0x192b08192b080808, 0x192b08192b192b19, 0x192b082b08081908, 0x192b082b08190808,
|
|
|
|
0x192b082b19080808, 0x192b082b1919192b, 0x192b082b2b2b0819, 0x192b190808080808,
|
|
|
|
0x192b190808081919, 0x192b190808082b08, 0x192b190808190819, 0x192b190808191908,
|
|
|
|
0x192b1908082b0808, 0x192b190819080819, 0x192b190819081908, 0x192b190819190808,
|
|
|
|
0x192b19082b080808, 0x192b191908080819, 0x192b191908081908, 0x192b191908190808,
|
|
|
|
0x192b191919080808, 0x192b191919082b2b, 0x192b1919192b2b08, 0x192b19192b19082b,
|
|
|
|
0x192b192b08080808, 0x192b192b2b191908, 0x192b2b0808080819, 0x192b2b0808081908,
|
|
|
|
0x192b2b0808190808, 0x192b2b08192b1919, 0x192b2b082b192b08, 0x192b2b1908080808,
|
|
|
|
0x192b2b19082b2b2b, 0x192b2b2b1908082b, 0x192b2b2b2b2b0819, 0x2b08080808080808,
|
|
|
|
0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08, 0x2b08080808190819,
|
|
|
|
0x2b08080808191908, 0x2b08080808192b19, 0x2b080808082b0808, 0x2b080808082b1919,
|
|
|
|
0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808081919082b,
|
|
|
|
0x2b08080819191919, 0x2b08080819192b08, 0x2b080808192b0819, 0x2b0808082b080808,
|
|
|
|
0x2b0808082b081919, 0x2b0808082b190819, 0x2b0808082b191908, 0x2b08081908080819,
|
|
|
|
0x2b08081908081908, 0x2b08081908082b19, 0x2b08081908190808, 0x2b0808190819082b,
|
|
|
|
0x2b08081908191919, 0x2b08081908192b08, 0x2b080819082b0819, 0x2b080819082b1908,
|
|
|
|
0x2b08081919080808, 0x2b0808191908082b, 0x2b08081919081919, 0x2b08081919082b08,
|
|
|
|
0x2b08081919190819, 0x2b08081919191908, 0x2b0808192b080819, 0x2b0808192b081908,
|
|
|
|
0x2b0808192b190808, 0x2b0808192b2b2b19, 0x2b08082b08080808, 0x2b08082b08081919,
|
|
|
|
0x2b08082b08082b2b, 0x2b08082b08190819, 0x2b08082b08191908, 0x2b08082b19080819,
|
|
|
|
0x2b08082b19081908, 0x2b08082b19190808, 0x2b08190808080819, 0x2b08190808081908,
|
|
|
|
0x2b0819080808192b, 0x2b08190808082b19, 0x2b08190808190808, 0x2b0819080819082b,
|
|
|
|
0x2b08190808191919, 0x2b08190808192b08, 0x2b081908082b0819, 0x2b08190819080808,
|
|
|
|
0x2b0819081908082b, 0x2b08190819081919, 0x2b08190819082b08, 0x2b08190819190819,
|
|
|
|
0x2b08190819191908, 0x2b081908192b0808, 0x2b0819082b080819, 0x2b0819082b081908,
|
|
|
|
0x2b0819082b190808, 0x2b08191908080808, 0x2b0819190808082b, 0x2b08191908081919,
|
|
|
|
0x2b08191908082b08, 0x2b08191908190819, 0x2b08191908191908, 0x2b081919082b0808,
|
|
|
|
0x2b08191919080819, 0x2b08191919081908, 0x2b08191919190808, 0x2b0819192b080808,
|
|
|
|
0x2b0819192b082b2b, 0x2b08192b08080819, 0x2b08192b08081908, 0x2b08192b08190808,
|
|
|
|
0x2b08192b082b2b19, 0x2b08192b19080808, 0x2b082b0808080808, 0x2b082b0808081919,
|
|
|
|
0x2b082b0808190819, 0x2b082b0808191908, 0x2b082b0819080819, 0x2b082b0819081908,
|
|
|
|
0x2b082b0819190808, 0x2b082b082b2b082b, 0x2b082b1908080819, 0x2b082b1908081908,
|
|
|
|
0x2b082b1919080808, 0x2b082b19192b1919, 0x2b082b2b082b082b, 0x2b082b2b19192b08,
|
|
|
|
0x2b082b2b19192b2b, 0x2b082b2b2b08082b, 0x2b082b2b2b2b082b, 0x2b19080808080819,
|
|
|
|
0x2b19080808081908, 0x2b19080808082b19, 0x2b19080808190808, 0x2b1908080819082b,
|
|
|
|
0x2b19080808191919, 0x2b19080808192b08, 0x2b190808082b1908, 0x2b19080819080808,
|
|
|
|
0x2b1908081908082b, 0x2b19080819081919, 0x2b19080819082b08, 0x2b19080819190819,
|
|
|
|
0x2b19080819191908, 0x2b190808192b0808, 0x2b1908082b080819, 0x2b1908082b081908,
|
|
|
|
0x2b1908082b190808, 0x2b19081908080808, 0x2b19081908081919, 0x2b19081908190819,
|
|
|
|
0x2b19081908191908, 0x2b19081919080819, 0x2b19081919081908, 0x2b19081919190808,
|
|
|
|
0x2b19081919192b2b, 0x2b19082b08080819, 0x2b19082b08081908, 0x2b19082b08190808,
|
|
|
|
0x2b19082b19080808, 0x2b19082b2b2b192b, 0x2b19190808080808, 0x2b1919080808082b,
|
|
|
|
0x2b19190808081919, 0x2b19190808082b08, 0x2b19190808190819, 0x2b19190808191908,
|
|
|
|
0x2b191908082b0808, 0x2b19190819080819, 0x2b19190819081908, 0x2b19190819190808,
|
|
|
|
0x2b1919082b080808, 0x2b1919082b19192b, 0x2b19191908080819, 0x2b19191908081908,
|
|
|
|
0x2b19191908190808, 0x2b19191919080808, 0x2b1919192b192b08, 0x2b1919192b2b0819,
|
|
|
|
0x2b19192b08080808, 0x2b19192b1908192b, 0x2b19192b192b1908, 0x2b192b0808080819,
|
|
|
|
0x2b192b0808081908, 0x2b192b0808190808, 0x2b192b08082b192b, 0x2b192b0819080808,
|
|
|
|
0x2b192b082b2b2b19, 0x2b192b1908080808, 0x2b192b1919082b19, 0x2b192b191919082b,
|
|
|
|
0x2b192b2b2b190808, 0x2b2b080808080808, 0x2b2b080808081919, 0x2b2b080808082b2b,
|
|
|
|
0x2b2b080808191908, 0x2b2b0808082b082b, 0x2b2b0808082b2b2b, 0x2b2b080819080819,
|
|
|
|
0x2b2b080819081908, 0x2b2b080819190808, 0x2b2b08082b2b082b, 0x2b2b08082b2b2b2b,
|
|
|
|
0x2b2b081919080808, 0x2b2b0819192b1919, 0x2b2b082b0808082b, 0x2b2b082b08082b2b,
|
|
|
|
0x2b2b082b082b082b, 0x2b2b082b082b2b08, 0x2b2b082b082b2b2b, 0x2b2b082b2b08082b,
|
|
|
|
0x2b2b082b2b082b08, 0x2b2b082b2b082b2b, 0x2b2b082b2b2b2b08, 0x2b2b190808080819,
|
|
|
|
0x2b2b190808081908, 0x2b2b190808190808, 0x2b2b190819080808, 0x2b2b19082b082b19,
|
|
|
|
0x2b2b19082b2b1908, 0x2b2b191908080808, 0x2b2b191908192b19, 0x2b2b192b19190819,
|
|
|
|
0x2b2b2b0808082b2b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b082b, 0x2b2b2b1919191908,
|
|
|
|
0x2b2b2b192b08192b, 0x2b2b2b2b08082b08, 0x2b2b2b2b08082b2b, 0x2b2b2b2b082b0808,
|
|
|
|
0x2b2b2b2b082b082b, 0x2b2b2b2b082b2b08, 0x2b2b2b2b2b082b08, 0x2b2b2b2b2b2b2b2b,
|
|
|
|
};
|
|
|
|
|
2024-01-30 13:14:12 +00:00
|
|
|
static const __device__ uint32_t iq3xxs_grid[256] = {
|
|
|
|
0x04040404, 0x04040414, 0x04040424, 0x04040c0c, 0x04040c1c, 0x04040c3e, 0x04041404, 0x04041414,
|
|
|
|
0x04041c0c, 0x04042414, 0x04043e1c, 0x04043e2c, 0x040c040c, 0x040c041c, 0x040c0c04, 0x040c0c14,
|
|
|
|
0x040c140c, 0x040c142c, 0x040c1c04, 0x040c1c14, 0x040c240c, 0x040c2c24, 0x040c3e04, 0x04140404,
|
|
|
|
0x04140414, 0x04140424, 0x04140c0c, 0x04141404, 0x04141414, 0x04141c0c, 0x04141c1c, 0x04141c3e,
|
|
|
|
0x04142c0c, 0x04142c3e, 0x04143e2c, 0x041c040c, 0x041c043e, 0x041c0c04, 0x041c0c14, 0x041c142c,
|
|
|
|
0x041c3e04, 0x04240c1c, 0x04241c3e, 0x04242424, 0x04242c3e, 0x04243e1c, 0x04243e2c, 0x042c040c,
|
|
|
|
0x042c043e, 0x042c1c14, 0x042c2c14, 0x04341c2c, 0x04343424, 0x043e0c04, 0x043e0c24, 0x043e0c34,
|
|
|
|
0x043e241c, 0x043e340c, 0x0c04040c, 0x0c04041c, 0x0c040c04, 0x0c040c14, 0x0c04140c, 0x0c04141c,
|
|
|
|
0x0c041c04, 0x0c041c14, 0x0c041c24, 0x0c04243e, 0x0c042c04, 0x0c0c0404, 0x0c0c0414, 0x0c0c0c0c,
|
|
|
|
0x0c0c1404, 0x0c0c1414, 0x0c14040c, 0x0c14041c, 0x0c140c04, 0x0c140c14, 0x0c14140c, 0x0c141c04,
|
|
|
|
0x0c143e14, 0x0c1c0404, 0x0c1c0414, 0x0c1c1404, 0x0c1c1c0c, 0x0c1c2434, 0x0c1c3434, 0x0c24040c,
|
|
|
|
0x0c24042c, 0x0c242c04, 0x0c2c1404, 0x0c2c1424, 0x0c2c2434, 0x0c2c3e0c, 0x0c34042c, 0x0c3e1414,
|
|
|
|
0x0c3e2404, 0x14040404, 0x14040414, 0x14040c0c, 0x14040c1c, 0x14041404, 0x14041414, 0x14041434,
|
|
|
|
0x14041c0c, 0x14042414, 0x140c040c, 0x140c041c, 0x140c042c, 0x140c0c04, 0x140c0c14, 0x140c140c,
|
|
|
|
0x140c1c04, 0x140c341c, 0x140c343e, 0x140c3e04, 0x14140404, 0x14140414, 0x14140c0c, 0x14140c3e,
|
|
|
|
0x14141404, 0x14141414, 0x14141c3e, 0x14142404, 0x14142c2c, 0x141c040c, 0x141c0c04, 0x141c0c24,
|
|
|
|
0x141c3e04, 0x141c3e24, 0x14241c2c, 0x14242c1c, 0x142c041c, 0x142c143e, 0x142c240c, 0x142c3e24,
|
|
|
|
0x143e040c, 0x143e041c, 0x143e0c34, 0x143e242c, 0x1c04040c, 0x1c040c04, 0x1c040c14, 0x1c04140c,
|
|
|
|
0x1c04141c, 0x1c042c04, 0x1c04342c, 0x1c043e14, 0x1c0c0404, 0x1c0c0414, 0x1c0c1404, 0x1c0c1c0c,
|
|
|
|
0x1c0c2424, 0x1c0c2434, 0x1c14040c, 0x1c14041c, 0x1c140c04, 0x1c14142c, 0x1c142c14, 0x1c143e14,
|
|
|
|
0x1c1c0c0c, 0x1c1c1c1c, 0x1c241c04, 0x1c24243e, 0x1c243e14, 0x1c2c0404, 0x1c2c0434, 0x1c2c1414,
|
|
|
|
0x1c2c2c2c, 0x1c340c24, 0x1c341c34, 0x1c34341c, 0x1c3e1c1c, 0x1c3e3404, 0x24040424, 0x24040c3e,
|
|
|
|
0x24041c2c, 0x24041c3e, 0x24042c1c, 0x24042c3e, 0x240c3e24, 0x24141404, 0x24141c3e, 0x24142404,
|
|
|
|
0x24143404, 0x24143434, 0x241c043e, 0x241c242c, 0x24240424, 0x24242c0c, 0x24243424, 0x242c142c,
|
|
|
|
0x242c241c, 0x242c3e04, 0x243e042c, 0x243e0c04, 0x243e0c14, 0x243e1c04, 0x2c040c14, 0x2c04240c,
|
|
|
|
0x2c043e04, 0x2c0c0404, 0x2c0c0434, 0x2c0c1434, 0x2c0c2c2c, 0x2c140c24, 0x2c141c14, 0x2c143e14,
|
|
|
|
0x2c1c0414, 0x2c1c2c1c, 0x2c240c04, 0x2c24141c, 0x2c24143e, 0x2c243e14, 0x2c2c0414, 0x2c2c1c0c,
|
|
|
|
0x2c342c04, 0x2c3e1424, 0x2c3e2414, 0x34041424, 0x34042424, 0x34042434, 0x34043424, 0x340c140c,
|
|
|
|
0x340c340c, 0x34140c3e, 0x34143424, 0x341c1c04, 0x341c1c34, 0x34242424, 0x342c042c, 0x342c2c14,
|
|
|
|
0x34341c1c, 0x343e041c, 0x343e140c, 0x3e04041c, 0x3e04042c, 0x3e04043e, 0x3e040c04, 0x3e041c14,
|
|
|
|
0x3e042c14, 0x3e0c1434, 0x3e0c2404, 0x3e140c14, 0x3e14242c, 0x3e142c14, 0x3e1c0404, 0x3e1c0c2c,
|
|
|
|
0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04,
|
|
|
|
};
|
|
|
|
|
2024-02-24 14:23:52 +00:00
|
|
|
static const __device__ uint32_t iq3xs_grid[512] = {
|
|
|
|
0x04040404, 0x0404040c, 0x04040414, 0x0404042c, 0x0404043e, 0x04040c04, 0x04040c0c, 0x04040c14,
|
|
|
|
0x04040c24, 0x04040c34, 0x04041404, 0x0404140c, 0x0404142c, 0x04041c1c, 0x04042404, 0x04042414,
|
|
|
|
0x0404242c, 0x0404243e, 0x04042c0c, 0x04042c1c, 0x04043404, 0x04043414, 0x04043e0c, 0x04043e24,
|
|
|
|
0x04043e3e, 0x040c0404, 0x040c040c, 0x040c0414, 0x040c0424, 0x040c0c04, 0x040c0c0c, 0x040c0c2c,
|
|
|
|
0x040c1404, 0x040c141c, 0x040c143e, 0x040c1c0c, 0x040c1c2c, 0x040c2424, 0x040c340c, 0x040c342c,
|
|
|
|
0x040c3e14, 0x04140404, 0x0414040c, 0x0414042c, 0x0414043e, 0x04140c04, 0x04140c1c, 0x04140c34,
|
|
|
|
0x0414140c, 0x0414142c, 0x04141c04, 0x04141c24, 0x04142414, 0x0414242c, 0x0414243e, 0x04142c0c,
|
|
|
|
0x04142c1c, 0x04143e04, 0x04143e1c, 0x041c041c, 0x041c0c0c, 0x041c0c2c, 0x041c1404, 0x041c1414,
|
|
|
|
0x041c1c0c, 0x041c1c1c, 0x041c1c34, 0x041c2424, 0x041c2c04, 0x041c2c14, 0x041c343e, 0x041c3e0c,
|
|
|
|
0x041c3e2c, 0x04240404, 0x04240c1c, 0x04240c3e, 0x0424140c, 0x04241424, 0x04241c14, 0x04242404,
|
|
|
|
0x0424241c, 0x04242c0c, 0x04243e04, 0x042c0414, 0x042c0424, 0x042c1404, 0x042c1414, 0x042c1434,
|
|
|
|
0x042c1c1c, 0x042c240c, 0x042c242c, 0x042c243e, 0x042c3434, 0x042c3e1c, 0x04340434, 0x04340c0c,
|
|
|
|
0x04340c1c, 0x04341c0c, 0x04342c14, 0x04343e0c, 0x043e0404, 0x043e0414, 0x043e0424, 0x043e1404,
|
|
|
|
0x043e1414, 0x043e1434, 0x043e1c1c, 0x043e2c04, 0x043e2c24, 0x0c040404, 0x0c04040c, 0x0c040414,
|
|
|
|
0x0c040424, 0x0c040c04, 0x0c040c0c, 0x0c040c1c, 0x0c040c2c, 0x0c040c3e, 0x0c041404, 0x0c041414,
|
|
|
|
0x0c041c0c, 0x0c041c24, 0x0c041c34, 0x0c042c24, 0x0c042c34, 0x0c04340c, 0x0c043e14, 0x0c0c0404,
|
|
|
|
0x0c0c040c, 0x0c0c041c, 0x0c0c0434, 0x0c0c0c04, 0x0c0c0c24, 0x0c0c140c, 0x0c0c1c04, 0x0c0c1c1c,
|
|
|
|
0x0c0c240c, 0x0c0c2c04, 0x0c0c2c14, 0x0c0c3e04, 0x0c0c3e34, 0x0c140404, 0x0c140c14, 0x0c140c2c,
|
|
|
|
0x0c140c3e, 0x0c141404, 0x0c141424, 0x0c141c14, 0x0c142404, 0x0c14241c, 0x0c142c2c, 0x0c143404,
|
|
|
|
0x0c143e14, 0x0c1c040c, 0x0c1c0424, 0x0c1c043e, 0x0c1c0c04, 0x0c1c0c1c, 0x0c1c140c, 0x0c1c143e,
|
|
|
|
0x0c1c1c04, 0x0c1c1c24, 0x0c1c240c, 0x0c1c3414, 0x0c1c3e04, 0x0c24041c, 0x0c24042c, 0x0c240c14,
|
|
|
|
0x0c240c24, 0x0c241c0c, 0x0c241c1c, 0x0c242414, 0x0c242434, 0x0c242c04, 0x0c242c24, 0x0c2c040c,
|
|
|
|
0x0c2c0c04, 0x0c2c0c1c, 0x0c2c140c, 0x0c2c1c04, 0x0c2c1c14, 0x0c2c2c0c, 0x0c341404, 0x0c341424,
|
|
|
|
0x0c34143e, 0x0c342424, 0x0c342434, 0x0c3e040c, 0x0c3e041c, 0x0c3e0c04, 0x0c3e0c14, 0x0c3e140c,
|
|
|
|
0x0c3e1c2c, 0x0c3e240c, 0x0c3e3414, 0x0c3e3e04, 0x14040404, 0x1404040c, 0x1404041c, 0x1404042c,
|
|
|
|
0x1404043e, 0x14040c04, 0x14040c14, 0x14040c24, 0x14040c34, 0x1404140c, 0x1404141c, 0x1404143e,
|
|
|
|
0x14041c04, 0x14041c14, 0x1404240c, 0x1404241c, 0x1404242c, 0x14042c04, 0x14042c14, 0x1404343e,
|
|
|
|
0x14043e04, 0x14043e1c, 0x14043e2c, 0x140c0404, 0x140c0414, 0x140c0c04, 0x140c0c1c, 0x140c0c3e,
|
|
|
|
0x140c1414, 0x140c142c, 0x140c1c0c, 0x140c1c24, 0x140c2414, 0x140c2c0c, 0x1414040c, 0x14140424,
|
|
|
|
0x1414043e, 0x1414140c, 0x1414141c, 0x14141c04, 0x14141c3e, 0x1414240c, 0x14142c1c, 0x14142c3e,
|
|
|
|
0x14143e0c, 0x14143e24, 0x141c0404, 0x141c0414, 0x141c042c, 0x141c0c0c, 0x141c1414, 0x141c1424,
|
|
|
|
0x141c1c0c, 0x141c1c1c, 0x141c2414, 0x141c2c04, 0x141c3434, 0x1424040c, 0x1424043e, 0x14241404,
|
|
|
|
0x1424141c, 0x14241c14, 0x14241c2c, 0x1424240c, 0x14243e14, 0x14243e2c, 0x142c0424, 0x142c0c0c,
|
|
|
|
0x142c1414, 0x142c1c3e, 0x142c2404, 0x142c2c1c, 0x142c3e04, 0x14340404, 0x14340414, 0x1434043e,
|
|
|
|
0x1434140c, 0x14342c2c, 0x1434340c, 0x143e042c, 0x143e0c0c, 0x143e1434, 0x143e1c04, 0x143e241c,
|
|
|
|
0x143e2c04, 0x1c040414, 0x1c040c0c, 0x1c040c1c, 0x1c040c2c, 0x1c040c3e, 0x1c041414, 0x1c041c0c,
|
|
|
|
0x1c041c1c, 0x1c041c2c, 0x1c042414, 0x1c042424, 0x1c04243e, 0x1c042c0c, 0x1c04341c, 0x1c043e0c,
|
|
|
|
0x1c0c040c, 0x1c0c041c, 0x1c0c042c, 0x1c0c0c24, 0x1c0c140c, 0x1c0c141c, 0x1c0c2404, 0x1c0c3404,
|
|
|
|
0x1c0c3e14, 0x1c0c3e34, 0x1c140404, 0x1c140c14, 0x1c141404, 0x1c141c14, 0x1c141c24, 0x1c142c04,
|
|
|
|
0x1c1c040c, 0x1c1c0c04, 0x1c1c0c24, 0x1c1c140c, 0x1c1c141c, 0x1c1c143e, 0x1c1c1c04, 0x1c1c240c,
|
|
|
|
0x1c1c241c, 0x1c1c243e, 0x1c1c2c2c, 0x1c1c3e1c, 0x1c24041c, 0x1c240c0c, 0x1c240c34, 0x1c241414,
|
|
|
|
0x1c241c0c, 0x1c242c14, 0x1c243404, 0x1c243424, 0x1c2c040c, 0x1c2c0c04, 0x1c2c0c14, 0x1c2c142c,
|
|
|
|
0x1c2c1c14, 0x1c2c2424, 0x1c2c2c34, 0x1c2c3e1c, 0x1c340c34, 0x1c34240c, 0x1c3e040c, 0x1c3e041c,
|
|
|
|
0x1c3e1404, 0x1c3e1414, 0x1c3e1c2c, 0x24040404, 0x24040424, 0x24040c14, 0x24041404, 0x24041424,
|
|
|
|
0x2404143e, 0x24041c14, 0x2404240c, 0x24042c04, 0x24043e04, 0x240c0414, 0x240c043e, 0x240c0c0c,
|
|
|
|
0x240c0c1c, 0x240c1414, 0x240c1c04, 0x240c1c2c, 0x240c241c, 0x240c2c0c, 0x240c2c2c, 0x2414040c,
|
|
|
|
0x2414041c, 0x24140c04, 0x24140c2c, 0x2414140c, 0x24141c1c, 0x24142404, 0x24142c3e, 0x24143414,
|
|
|
|
0x24143e04, 0x241c0424, 0x241c0c0c, 0x241c0c1c, 0x241c1404, 0x241c1414, 0x241c1c0c, 0x241c1c2c,
|
|
|
|
0x24240404, 0x24240414, 0x24241424, 0x24241c3e, 0x24242404, 0x24243e0c, 0x242c042c, 0x242c043e,
|
|
|
|
0x242c140c, 0x242c3414, 0x24340c1c, 0x24341c24, 0x24343404, 0x243e0c04, 0x243e0c2c, 0x243e1c04,
|
|
|
|
0x243e241c, 0x243e2c0c, 0x2c040414, 0x2c040c04, 0x2c040c24, 0x2c041414, 0x2c042404, 0x2c042424,
|
|
|
|
0x2c04243e, 0x2c042c14, 0x2c043434, 0x2c043e24, 0x2c0c040c, 0x2c0c041c, 0x2c0c042c, 0x2c0c0c14,
|
|
|
|
0x2c0c140c, 0x2c0c1c14, 0x2c0c3e14, 0x2c140404, 0x2c140c0c, 0x2c14141c, 0x2c141c04, 0x2c141c34,
|
|
|
|
0x2c142c1c, 0x2c1c0414, 0x2c1c043e, 0x2c1c0c04, 0x2c1c143e, 0x2c1c2424, 0x2c1c2c0c, 0x2c1c342c,
|
|
|
|
0x2c1c3e1c, 0x2c24040c, 0x2c240424, 0x2c241404, 0x2c241c14, 0x2c242434, 0x2c2c0c14, 0x2c2c1434,
|
|
|
|
0x2c2c2c0c, 0x2c2c2c1c, 0x2c342414, 0x2c3e0414, 0x2c3e0424, 0x2c3e1414, 0x34040c0c, 0x34040c1c,
|
|
|
|
0x34040c2c, 0x34041c0c, 0x34041c1c, 0x34043404, 0x340c0404, 0x340c1404, 0x340c143e, 0x340c3424,
|
|
|
|
0x34140c14, 0x34141c24, 0x34142414, 0x34142c2c, 0x34143414, 0x34143e04, 0x341c0404, 0x341c0c24,
|
|
|
|
0x341c140c, 0x341c2404, 0x3424142c, 0x3424241c, 0x34243414, 0x342c0404, 0x342c041c, 0x342c1c24,
|
|
|
|
0x342c3404, 0x3434042c, 0x34342404, 0x343e0c0c, 0x343e0c1c, 0x3e040404, 0x3e040424, 0x3e04043e,
|
|
|
|
0x3e041404, 0x3e041414, 0x3e041c34, 0x3e042404, 0x3e042c24, 0x3e043414, 0x3e0c0414, 0x3e0c0c0c,
|
|
|
|
0x3e0c1424, 0x3e0c241c, 0x3e0c242c, 0x3e14040c, 0x3e140424, 0x3e140c04, 0x3e140c34, 0x3e14140c,
|
|
|
|
0x3e141c04, 0x3e142c0c, 0x3e1c0414, 0x3e1c1c14, 0x3e1c1c2c, 0x3e1c2c1c, 0x3e24040c, 0x3e24042c,
|
|
|
|
0x3e240c1c, 0x3e241404, 0x3e242c04, 0x3e2c1414, 0x3e2c2414, 0x3e340414, 0x3e341c0c, 0x3e3e0404,
|
|
|
|
};
|
|
|
|
|
|
|
|
|
2024-02-18 16:16:55 +00:00
|
|
|
static const __device__ uint64_t iq1s_grid[512] = {
|
|
|
|
0xffffffffffff0101, 0xffffffffff01ff00, 0xffffffffff010100, 0xffffffff00000000,
|
|
|
|
0xffffffff01ff00ff, 0xffffffff01ff0001, 0xffffffff0101ffff, 0xffffffff0101ff01,
|
|
|
|
0xffffff00ff000000, 0xffffff000000ff00, 0xffffff00000000ff, 0xffffff0000000100,
|
|
|
|
0xffffff0000010000, 0xffffff0001000000, 0xffffff01ffff00ff, 0xffffff01ff01ff00,
|
|
|
|
0xffffff01ff010100, 0xffffff0100000001, 0xffffff0101ffff00, 0xffffff0101ff0101,
|
|
|
|
0xffffff0101010100, 0xffff00ffff00ff01, 0xffff00ffff0000ff, 0xffff00ff00ff0100,
|
|
|
|
0xffff00ff0100ff00, 0xffff00ff010001ff, 0xffff0000ff0101ff, 0xffff000000ffff00,
|
|
|
|
0xffff000000000000, 0xffff00000001ff01, 0xffff000001000101, 0xffff0000010100ff,
|
|
|
|
0xffff0001ffff0100, 0xffff00010000ff00, 0xffff000100010101, 0xffff000101000000,
|
|
|
|
0xffff01ffffff0000, 0xffff01ffff01ffff, 0xffff01ffff010100, 0xffff01ff00000000,
|
|
|
|
0xffff01ff01ffffff, 0xffff01ff01ff0001, 0xffff01ff0101ffff, 0xffff01ff01010001,
|
|
|
|
0xffff0100ffffff01, 0xffff01000000ffff, 0xffff010000000100, 0xffff010001ff01ff,
|
|
|
|
0xffff010001000000, 0xffff0101ff000000, 0xffff0101000101ff, 0xffff010101ffff01,
|
|
|
|
0xffff01010101ff00, 0xff00ffffff000000, 0xff00ffff00ffff00, 0xff00ffff00000001,
|
|
|
|
0xff00ffff000001ff, 0xff00ffff01010000, 0xff00ff00ffff0000, 0xff00ff00ff00ff00,
|
|
|
|
0xff00ff00ff0000ff, 0xff00ff00ff000100, 0xff00ff00ff010001, 0xff00ff0000ff0001,
|
|
|
|
0xff00ff000000ffff, 0xff00ff0000000000, 0xff00ff000001ff00, 0xff00ff0000010100,
|
|
|
|
0xff00ff0001ff0000, 0xff00ff000100ff00, 0xff00ff0001000100, 0xff00ff01ff000000,
|
|
|
|
0xff00ff0100ff0000, 0xff00ff01000001ff, 0xff00ff0101010001, 0xff0000ff00000000,
|
|
|
|
0xff0000ff0001ff00, 0xff0000ff00010100, 0xff000000ffff0101, 0xff000000ff000000,
|
|
|
|
0xff000000ff01ff00, 0xff00000000ff0000, 0xff0000000000ff00, 0xff000000000000ff,
|
|
|
|
0xff00000000000000, 0xff00000000000001, 0xff00000000000100, 0xff0000000001ffff,
|
|
|
|
0xff00000000010000, 0xff00000001000000, 0xff00000001010100, 0xff000001ff00ff01,
|
|
|
|
0xff000001ff0100ff, 0xff00000100000000, 0xff0000010001ff00, 0xff00000101ff0100,
|
|
|
|
0xff0000010100ff00, 0xff0001ff00ff00ff, 0xff0001ff00000101, 0xff0001ff000100ff,
|
|
|
|
0xff0001ff01000000, 0xff000100ff0001ff, 0xff0001000000ff01, 0xff00010000000000,
|
|
|
|
0xff00010000010001, 0xff00010000010100, 0xff00010001ffff00, 0xff00010001ff0101,
|
|
|
|
0xff00010001010000, 0xff000101ffffffff, 0xff000101ff000101, 0xff00010101ff00ff,
|
|
|
|
0xff00010101000001, 0xff000101010100ff, 0xff01ffffff000101, 0xff01ffffff01ffff,
|
|
|
|
0xff01ffffff01ff01, 0xff01ffffff0101ff, 0xff01ffff00000000, 0xff01ffff01ff0001,
|
|
|
|
0xff01ffff0101ff01, 0xff01ff00ff000000, 0xff01ff0000ff0100, 0xff01ff000000ff01,
|
|
|
|
0xff01ff0000010000, 0xff01ff00010000ff, 0xff01ff01ff01ff00, 0xff01ff0100000101,
|
|
|
|
0xff0100ffffff0000, 0xff0100ffff010000, 0xff0100ff01ff00ff, 0xff0100ff01000100,
|
|
|
|
0xff0100ff010100ff, 0xff010000ffffff01, 0xff01000000000000, 0xff0100000101ff00,
|
|
|
|
0xff010001ffff00ff, 0xff010001ff000100, 0xff01000100ffff00, 0xff01000100010001,
|
|
|
|
0xff01000101ff0001, 0xff010001010001ff, 0xff0101ffffffffff, 0xff0101ffff01ffff,
|
|
|
|
0xff0101ffff010101, 0xff0101ff0000ff00, 0xff0101ff01010001, 0xff010100ff000000,
|
|
|
|
0xff010100ff01ff01, 0xff01010000ff0001, 0xff01010000000100, 0xff01010001000000,
|
|
|
|
0xff0101010100ffff, 0x00ffffff0000ff01, 0x00ffffff000000ff, 0x00ffffff00000100,
|
|
|
|
0x00ffffff00010000, 0x00ffff00ffff0001, 0x00ffff00ff0000ff, 0x00ffff00ff000100,
|
|
|
|
0x00ffff0000000000, 0x00ffff0001000100, 0x00ffff0001010001, 0x00ffff01ff00ff01,
|
|
|
|
0x00ffff0100ff0100, 0x00ffff010000ff00, 0x00ffff01000100ff, 0x00ffff0101ff00ff,
|
|
|
|
0x00ffff010101ff00, 0x00ff00ffffffffff, 0x00ff00ffffff01ff, 0x00ff00ffff000101,
|
|
|
|
0x00ff00ff00000000, 0x00ff00ff000101ff, 0x00ff00ff01010101, 0x00ff0000ff000000,
|
|
|
|
0x00ff0000ff01ffff, 0x00ff000000ff0000, 0x00ff00000000ff00, 0x00ff0000000000ff,
|
|
|
|
0x00ff000000000000, 0x00ff000000000001, 0x00ff000000000100, 0x00ff000000010000,
|
|
|
|
0x00ff000001ffff01, 0x00ff000001000000, 0x00ff0001ff000101, 0x00ff000100ffffff,
|
|
|
|
0x00ff000100000000, 0x00ff0001010001ff, 0x00ff01ffff000000, 0x00ff01ff0001ff00,
|
|
|
|
0x00ff01ff01ff0100, 0x00ff0100ff01ff01, 0x00ff010000ff00ff, 0x00ff010000ff0101,
|
|
|
|
0x00ff010000000000, 0x00ff010000010101, 0x00ff01000100ff00, 0x00ff010001010000,
|
|
|
|
0x00ff0101ffffff00, 0x00ff01010000ff01, 0x00ff010100000100, 0x00ff010101ff0000,
|
|
|
|
0x0000ffffffff0100, 0x0000ffffff00ff00, 0x0000ffffff0000ff, 0x0000ffffff010000,
|
|
|
|
0x0000ffff00000000, 0x0000ffff00010101, 0x0000ffff01ffff01, 0x0000ffff01000100,
|
|
|
|
0x0000ff00ff000000, 0x0000ff00ff01ff00, 0x0000ff00ff0101ff, 0x0000ff0000ff0000,
|
|
|
|
0x0000ff000000ff00, 0x0000ff00000000ff, 0x0000ff0000000000, 0x0000ff0000000001,
|
|
|
|
0x0000ff0000000100, 0x0000ff0000010000, 0x0000ff0001ffffff, 0x0000ff0001ff01ff,
|
|
|
|
0x0000ff0001000000, 0x0000ff000101ffff, 0x0000ff01ffff0101, 0x0000ff01ff010000,
|
|
|
|
0x0000ff0100000000, 0x0000ff0101000101, 0x000000ffffff0001, 0x000000ffff000000,
|
|
|
|
0x000000ff00ff0000, 0x000000ff0000ff00, 0x000000ff000000ff, 0x000000ff00000000,
|
|
|
|
0x000000ff00000001, 0x000000ff00000100, 0x000000ff00010000, 0x000000ff01000000,
|
|
|
|
0x000000ff0101ff00, 0x00000000ffff0000, 0x00000000ff00ff00, 0x00000000ff0000ff,
|
|
|
|
0x00000000ff000000, 0x00000000ff000001, 0x00000000ff000100, 0x00000000ff010000,
|
|
|
|
0x0000000000ffff00, 0x0000000000ff00ff, 0x0000000000ff0000, 0x0000000000ff0001,
|
|
|
|
0x0000000000ff0100, 0x000000000000ffff, 0x000000000000ff00, 0x000000000000ff01,
|
|
|
|
0x00000000000000ff, 0x0000000000000001, 0x00000000000001ff, 0x0000000000000100,
|
|
|
|
0x0000000000000101, 0x000000000001ff00, 0x00000000000100ff, 0x0000000000010000,
|
|
|
|
0x0000000000010001, 0x0000000000010100, 0x0000000001ff0000, 0x000000000100ff00,
|
|
|
|
0x00000000010000ff, 0x0000000001000000, 0x0000000001000001, 0x0000000001000100,
|
|
|
|
0x0000000001010000, 0x00000001ffff01ff, 0x00000001ff000000, 0x0000000100ff0000,
|
|
|
|
0x000000010000ff00, 0x00000001000000ff, 0x0000000100000000, 0x0000000100000001,
|
|
|
|
0x0000000100000100, 0x0000000100010000, 0x0000000101000000, 0x000001ffff00ff00,
|
|
|
|
0x000001ffff010001, 0x000001ffff0101ff, 0x000001ff00ffff01, 0x000001ff0000ffff,
|
|
|
|
0x000001ff00000000, 0x000001ff010000ff, 0x000001ff01010100, 0x00000100ffff0100,
|
|
|
|
0x00000100ff000000, 0x0000010000ff0000, 0x000001000000ff00, 0x00000100000000ff,
|
|
|
|
0x0000010000000000, 0x0000010000000001, 0x0000010000000100, 0x0000010000010000,
|
|
|
|
0x0000010001000000, 0x000001000101ff01, 0x00000101ffff0001, 0x00000101ff01ffff,
|
|
|
|
0x0000010100000000, 0x0000010101010100, 0x0001ffffff000000, 0x0001ffff00ffffff,
|
|
|
|
0x0001ffff00000100, 0x0001ffff0001ff00, 0x0001ffff01000000, 0x0001ff00ffffff00,
|
|
|
|
0x0001ff00ffff01ff, 0x0001ff00ff010000, 0x0001ff0000000000, 0x0001ff0000010001,
|
|
|
|
0x0001ff0001ff0000, 0x0001ff0001010100, 0x0001ff01ff0000ff, 0x0001ff01ff000001,
|
|
|
|
0x0001ff0100ffffff, 0x0001ff010001ffff, 0x0001ff01000101ff, 0x0001ff010100ff01,
|
|
|
|
0x000100ffff00ffff, 0x000100ffff00ff01, 0x000100ffff000100, 0x000100ff00000000,
|
|
|
|
0x000100ff000101ff, 0x000100ff01ff0101, 0x000100ff0100ffff, 0x000100ff01010101,
|
|
|
|
0x00010000ff000000, 0x00010000ff010100, 0x0001000000ff0000, 0x000100000000ff00,
|
|
|
|
0x00010000000000ff, 0x0001000000000000, 0x0001000000000001, 0x0001000000000100,
|
|
|
|
0x0001000000010000, 0x0001000001ffff01, 0x0001000001000000, 0x0001000100ff0101,
|
|
|
|
0x0001000100000000, 0x00010001010100ff, 0x000101ffffff01ff, 0x000101ffffff0101,
|
|
|
|
0x000101ff00010000, 0x000101ff01ff0000, 0x000101ff0100ff01, 0x00010100ffff0000,
|
|
|
|
0x0001010000000000, 0x000101000001ffff, 0x0001010000010101, 0x00010100010001ff,
|
|
|
|
0x00010101ff00ff00, 0x00010101ff010001, 0x0001010100ffffff, 0x0001010100ff01ff,
|
|
|
|
0x00010101000101ff, 0x0001010101ff0000, 0x000101010100ff01, 0x0001010101000101,
|
|
|
|
0x01ffffffffff0101, 0x01ffffffff01ffff, 0x01ffffffff01ff01, 0x01ffffffff0101ff,
|
|
|
|
0x01ffffffff010101, 0x01ffffff00000000, 0x01ffffff01ff01ff, 0x01ffffff01000101,
|
|
|
|
0x01ffffff0101ff01, 0x01ffffff010100ff, 0x01ffff000000ff00, 0x01ffff0000000001,
|
|
|
|
0x01ffff00000001ff, 0x01ffff0000010000, 0x01ffff0001ff0000, 0x01ffff01ffffffff,
|
|
|
|
0x01ffff01ffff01ff, 0x01ffff01ff000000, 0x01ffff01ff01ffff, 0x01ffff01ff0101ff,
|
|
|
|
0x01ffff010100ffff, 0x01ff00ffffff0000, 0x01ff00ffff010000, 0x01ff00ff00ffff01,
|
|
|
|
0x01ff0000ff0000ff, 0x01ff000000000000, 0x01ff00000001ff01, 0x01ff000001ffffff,
|
|
|
|
0x01ff000001010100, 0x01ff0001ffffff01, 0x01ff0001ff010001, 0x01ff000101ff0100,
|
|
|
|
0x01ff000101000001, 0x01ff0001010100ff, 0x01ff01ffff00ffff, 0x01ff01ff00010001,
|
|
|
|
0x01ff01ff01000000, 0x01ff01ff010101ff, 0x01ff0100ff000001, 0x01ff010000ffff00,
|
|
|
|
0x01ff010000000100, 0x01ff010001ff01ff, 0x01ff01000101ffff, 0x01ff0101ffff00ff,
|
|
|
|
0x01ff0101ffff0101, 0x01ff0101ff0101ff, 0x01ff010100010000, 0x0100ffff00ff00ff,
|
|
|
|
0x0100ffff00ff0001, 0x0100ffff00000100, 0x0100ffff0100ff00, 0x0100ff00ffff0000,
|
|
|
|
0x0100ff00ff00ffff, 0x0100ff00ff00ff01, 0x0100ff00ff000100, 0x0100ff00ff010000,
|
|
|
|
0x0100ff0000000000, 0x0100ff00000100ff, 0x0100ff0001ff0101, 0x0100ff0001010101,
|
|
|
|
0x0100ff0100ff00ff, 0x0100ff0100ff0001, 0x0100ff0100000100, 0x0100ff0100010001,
|
|
|
|
0x0100ff0101000000, 0x010000ffff00ff00, 0x010000ff0000ffff, 0x010000ff00000000,
|
|
|
|
0x010000ff010001ff, 0x010000ff01010001, 0x01000000ffffff00, 0x01000000ffff0101,
|
|
|
|
0x01000000ff000000, 0x01000000ff0100ff, 0x01000000ff010101, 0x0100000000ff0000,
|
|
|
|
0x010000000000ff00, 0x01000000000000ff, 0x0100000000000000, 0x0100000000000001,
|
|
|
|
0x0100000000000100, 0x0100000000010000, 0x0100000001000000, 0x0100000100000000,
|
|
|
|
0x01000001000101ff, 0x0100000101ffff01, 0x010001ffff000101, 0x010001ff00ff0100,
|
|
|
|
0x010001ff0000ff00, 0x010001ff000100ff, 0x010001ff01ffffff, 0x01000100ffff0000,
|
|
|
|
0x01000100ff0001ff, 0x0100010000000000, 0x010001000001ff00, 0x0100010001ff0000,
|
|
|
|
0x01000100010000ff, 0x0100010001000101, 0x01000101ff00ff01, 0x0100010100ff0100,
|
|
|
|
0x010001010000ffff, 0x0100010101010001, 0x0101ffffffff0101, 0x0101ffffff0001ff,
|
|
|
|
0x0101ffffff01ffff, 0x0101ffffff010101, 0x0101ffff00000000, 0x0101ffff0101ffff,
|
|
|
|
0x0101ffff010101ff, 0x0101ff00ff000000, 0x0101ff0000ff0100, 0x0101ff000000ff00,
|
|
|
|
0x0101ff0000010000, 0x0101ff00010000ff, 0x0101ff0001000001, 0x0101ff01ff010101,
|
|
|
|
0x0101ff0100000000, 0x0101ff010101ff00, 0x010100ffffff0000, 0x010100ffff010000,
|
|
|
|
0x010100ff00ff01ff, 0x010100ff000000ff, 0x010100ff00000101, 0x010100ff01ffff00,
|
|
|
|
0x01010000ffffff01, 0x01010000ff000100, 0x01010000ff01ff01, 0x0101000000000000,
|
|
|
|
0x01010000000100ff, 0x010100000101ff01, 0x01010001ffff0000, 0x01010001ff00ffff,
|
|
|
|
0x01010001ff010000, 0x0101000101ffffff, 0x0101000101ff01ff, 0x0101000101010101,
|
|
|
|
0x010101ffff01ffff, 0x010101ff00000000, 0x010101ff0001ff01, 0x010101ff0101ffff,
|
|
|
|
0x010101ff010101ff, 0x01010100ffffffff, 0x01010100ff000001, 0x010101000000ff00,
|
|
|
|
0x0101010001010000, 0x0101010100ff0001, 0x010101010001ff01, 0x010101010101ffff,
|
|
|
|
};
|
|
|
|
|
2024-01-08 15:02:32 +00:00
|
|
|
static const __device__ uint8_t ksigns_iq2xs[128] = {
|
|
|
|
0, 129, 130, 3, 132, 5, 6, 135, 136, 9, 10, 139, 12, 141, 142, 15,
|
|
|
|
144, 17, 18, 147, 20, 149, 150, 23, 24, 153, 154, 27, 156, 29, 30, 159,
|
|
|
|
160, 33, 34, 163, 36, 165, 166, 39, 40, 169, 170, 43, 172, 45, 46, 175,
|
|
|
|
48, 177, 178, 51, 180, 53, 54, 183, 184, 57, 58, 187, 60, 189, 190, 63,
|
|
|
|
192, 65, 66, 195, 68, 197, 198, 71, 72, 201, 202, 75, 204, 77, 78, 207,
|
|
|
|
80, 209, 210, 83, 212, 85, 86, 215, 216, 89, 90, 219, 92, 221, 222, 95,
|
|
|
|
96, 225, 226, 99, 228, 101, 102, 231, 232, 105, 106, 235, 108, 237, 238, 111,
|
|
|
|
240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255,
|
|
|
|
};
|
|
|
|
|
2024-01-30 13:14:12 +00:00
|
|
|
//#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
|
|
|
|
static const __device__ uint64_t ksigns64[128] = {
|
|
|
|
0x0000000000000000, 0xff000000000000ff, 0xff0000000000ff00, 0x000000000000ffff,
|
|
|
|
0xff00000000ff0000, 0x0000000000ff00ff, 0x0000000000ffff00, 0xff00000000ffffff,
|
|
|
|
0xff000000ff000000, 0x00000000ff0000ff, 0x00000000ff00ff00, 0xff000000ff00ffff,
|
|
|
|
0x00000000ffff0000, 0xff000000ffff00ff, 0xff000000ffffff00, 0x00000000ffffffff,
|
|
|
|
0xff0000ff00000000, 0x000000ff000000ff, 0x000000ff0000ff00, 0xff0000ff0000ffff,
|
|
|
|
0x000000ff00ff0000, 0xff0000ff00ff00ff, 0xff0000ff00ffff00, 0x000000ff00ffffff,
|
|
|
|
0x000000ffff000000, 0xff0000ffff0000ff, 0xff0000ffff00ff00, 0x000000ffff00ffff,
|
|
|
|
0xff0000ffffff0000, 0x000000ffffff00ff, 0x000000ffffffff00, 0xff0000ffffffffff,
|
|
|
|
0xff00ff0000000000, 0x0000ff00000000ff, 0x0000ff000000ff00, 0xff00ff000000ffff,
|
|
|
|
0x0000ff0000ff0000, 0xff00ff0000ff00ff, 0xff00ff0000ffff00, 0x0000ff0000ffffff,
|
|
|
|
0x0000ff00ff000000, 0xff00ff00ff0000ff, 0xff00ff00ff00ff00, 0x0000ff00ff00ffff,
|
|
|
|
0xff00ff00ffff0000, 0x0000ff00ffff00ff, 0x0000ff00ffffff00, 0xff00ff00ffffffff,
|
|
|
|
0x0000ffff00000000, 0xff00ffff000000ff, 0xff00ffff0000ff00, 0x0000ffff0000ffff,
|
|
|
|
0xff00ffff00ff0000, 0x0000ffff00ff00ff, 0x0000ffff00ffff00, 0xff00ffff00ffffff,
|
|
|
|
0xff00ffffff000000, 0x0000ffffff0000ff, 0x0000ffffff00ff00, 0xff00ffffff00ffff,
|
|
|
|
0x0000ffffffff0000, 0xff00ffffffff00ff, 0xff00ffffffffff00, 0x0000ffffffffffff,
|
|
|
|
0xffff000000000000, 0x00ff0000000000ff, 0x00ff00000000ff00, 0xffff00000000ffff,
|
|
|
|
0x00ff000000ff0000, 0xffff000000ff00ff, 0xffff000000ffff00, 0x00ff000000ffffff,
|
|
|
|
0x00ff0000ff000000, 0xffff0000ff0000ff, 0xffff0000ff00ff00, 0x00ff0000ff00ffff,
|
|
|
|
0xffff0000ffff0000, 0x00ff0000ffff00ff, 0x00ff0000ffffff00, 0xffff0000ffffffff,
|
|
|
|
0x00ff00ff00000000, 0xffff00ff000000ff, 0xffff00ff0000ff00, 0x00ff00ff0000ffff,
|
|
|
|
0xffff00ff00ff0000, 0x00ff00ff00ff00ff, 0x00ff00ff00ffff00, 0xffff00ff00ffffff,
|
|
|
|
0xffff00ffff000000, 0x00ff00ffff0000ff, 0x00ff00ffff00ff00, 0xffff00ffff00ffff,
|
|
|
|
0x00ff00ffffff0000, 0xffff00ffffff00ff, 0xffff00ffffffff00, 0x00ff00ffffffffff,
|
|
|
|
0x00ffff0000000000, 0xffffff00000000ff, 0xffffff000000ff00, 0x00ffff000000ffff,
|
|
|
|
0xffffff0000ff0000, 0x00ffff0000ff00ff, 0x00ffff0000ffff00, 0xffffff0000ffffff,
|
|
|
|
0xffffff00ff000000, 0x00ffff00ff0000ff, 0x00ffff00ff00ff00, 0xffffff00ff00ffff,
|
|
|
|
0x00ffff00ffff0000, 0xffffff00ffff00ff, 0xffffff00ffffff00, 0x00ffff00ffffffff,
|
|
|
|
0xffffffff00000000, 0x00ffffff000000ff, 0x00ffffff0000ff00, 0xffffffff0000ffff,
|
|
|
|
0x00ffffff00ff0000, 0xffffffff00ff00ff, 0xffffffff00ffff00, 0x00ffffff00ffffff,
|
|
|
|
0x00ffffffff000000, 0xffffffffff0000ff, 0xffffffffff00ff00, 0x00ffffffff00ffff,
|
|
|
|
0xffffffffffff0000, 0x00ffffffffff00ff, 0x00ffffffffffff00, 0xffffffffffffffff,
|
|
|
|
};
|
|
|
|
//#endif
|
|
|
|
|
2024-01-08 15:02:32 +00:00
|
|
|
static const __device__ uint8_t kmask_iq2xs[8] = {1, 2, 4, 8, 16, 32, 64, 128};
|
|
|
|
|
|
|
|
inline bool ggml_cuda_supports_mmq(enum ggml_type type) {
|
|
|
|
switch (type) {
|
|
|
|
case GGML_TYPE_Q4_0:
|
|
|
|
case GGML_TYPE_Q4_1:
|
|
|
|
case GGML_TYPE_Q5_0:
|
|
|
|
case GGML_TYPE_Q5_1:
|
|
|
|
case GGML_TYPE_Q8_0:
|
|
|
|
case GGML_TYPE_Q2_K:
|
|
|
|
case GGML_TYPE_Q3_K:
|
|
|
|
case GGML_TYPE_Q4_K:
|
|
|
|
case GGML_TYPE_Q5_K:
|
|
|
|
case GGML_TYPE_Q6_K:
|
|
|
|
return true;
|
|
|
|
default:
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
template<typename dst_t>
|
|
|
|
static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
|
|
|
|
|
|
|
|
const int i = blockIdx.x;
|
|
|
|
const block_iq2_xxs * x = (const block_iq2_xxs *) vx;
|
|
|
|
|
|
|
|
const int tid = threadIdx.x;
|
|
|
|
#if QK_K == 256
|
|
|
|
const int il = tid/8; // 0...3
|
|
|
|
const int ib = tid%8; // 0...7
|
|
|
|
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
|
|
|
|
const uint16_t * q2 = x[i].qs + 4*ib;
|
|
|
|
const uint8_t * aux8 = (const uint8_t *)q2;
|
2024-01-11 19:39:39 +00:00
|
|
|
const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[il]);
|
2024-01-08 15:02:32 +00:00
|
|
|
const uint32_t aux32 = q2[2] | (q2[3] << 16);
|
|
|
|
const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.25f;
|
|
|
|
const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
|
|
|
|
for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
|
|
|
|
#else
|
|
|
|
assert(false);
|
|
|
|
#endif
|
|
|
|
|
|
|
|
}
|
|
|
|
|
2024-01-11 19:39:39 +00:00
|
|
|
template<typename dst_t>
|
|
|
|
static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
|
|
|
|
|
|
|
|
const int i = blockIdx.x;
|
|
|
|
const block_iq2_xs * x = (const block_iq2_xs *) vx;
|
|
|
|
|
|
|
|
const int tid = threadIdx.x;
|
|
|
|
#if QK_K == 256
|
|
|
|
const int il = tid/8; // 0...3
|
|
|
|
const int ib = tid%8; // 0...7
|
|
|
|
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
|
|
|
|
const uint16_t * q2 = x[i].qs + 4*ib;
|
|
|
|
const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[il] & 511));
|
|
|
|
const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
|
|
|
|
const uint8_t signs = ksigns_iq2xs[q2[il] >> 9];
|
|
|
|
for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
|
|
|
|
#else
|
|
|
|
assert(false);
|
|
|
|
#endif
|
|
|
|
|
|
|
|
}
|
|
|
|
|
2024-02-26 16:28:38 +00:00
|
|
|
template<typename dst_t>
|
|
|
|
static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
|
|
|
|
|
|
|
|
const int i = blockIdx.x;
|
|
|
|
const block_iq2_s * x = (const block_iq2_s *) vx;
|
|
|
|
|
|
|
|
const int tid = threadIdx.x;
|
|
|
|
#if QK_K == 256
|
|
|
|
const int il = tid/8; // 0...3
|
|
|
|
const int ib = tid%8; // 0...7
|
|
|
|
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
|
|
|
|
const uint8_t * grid = (const uint8_t *)(iq2s_grid + (x[i].qs[4*ib+il] | ((x[i].qh[ib] << (8-2*il)) & 0x300)));
|
|
|
|
const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
|
|
|
|
const uint8_t signs = x[i].qs[QK_K/8+4*ib+il];
|
|
|
|
for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
|
|
|
|
#else
|
|
|
|
assert(false);
|
|
|
|
#endif
|
|
|
|
|
|
|
|
}
|
|
|
|
|
2024-01-30 13:14:12 +00:00
|
|
|
template<typename dst_t>
|
|
|
|
static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
|
|
|
|
|
|
|
|
const int i = blockIdx.x;
|
|
|
|
const block_iq3_xxs * x = (const block_iq3_xxs *) vx;
|
|
|
|
|
|
|
|
const int tid = threadIdx.x;
|
|
|
|
#if QK_K == 256
|
|
|
|
const int il = tid/8; // 0...3
|
|
|
|
const int ib = tid%8; // 0...7
|
|
|
|
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
|
|
|
|
const uint8_t * q3 = x[i].qs + 8*ib;
|
|
|
|
const uint16_t * gas = (const uint16_t *)(x[i].qs + QK_K/4) + 2*ib;
|
|
|
|
const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*il+0]);
|
|
|
|
const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*il+1]);
|
|
|
|
const uint32_t aux32 = gas[0] | (gas[1] << 16);
|
|
|
|
const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.5f;
|
|
|
|
const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
|
|
|
|
for (int j = 0; j < 4; ++j) {
|
|
|
|
y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
|
|
|
|
y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
|
|
|
|
}
|
|
|
|
#else
|
|
|
|
assert(false);
|
|
|
|
#endif
|
|
|
|
|
|
|
|
}
|
|
|
|
|
2024-02-24 14:23:52 +00:00
|
|
|
template<typename dst_t>
|
|
|
|
static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
|
|
|
|
|
|
|
|
const int i = blockIdx.x;
|
|
|
|
const block_iq3_s * x = (const block_iq3_s *) vx;
|
|
|
|
|
|
|
|
const int tid = threadIdx.x;
|
|
|
|
#if QK_K == 256
|
|
|
|
const int il = tid/8; // 0...3
|
|
|
|
const int ib = tid%8; // 0...7
|
|
|
|
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
|
|
|
|
const uint8_t * qs = x[i].qs + 8*ib;
|
|
|
|
const uint8_t * grid1 = (const uint8_t *)(iq3xs_grid + (qs[2*il+0] | ((x[i].qh[ib] << (8-2*il)) & 256)));
|
|
|
|
const uint8_t * grid2 = (const uint8_t *)(iq3xs_grid + (qs[2*il+1] | ((x[i].qh[ib] << (7-2*il)) & 256)));
|
|
|
|
const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib/2] >> 4*(ib%2)) & 0xf)) * 0.5f;
|
|
|
|
const uint8_t signs = x[i].signs[4*ib + il];
|
|
|
|
for (int j = 0; j < 4; ++j) {
|
|
|
|
y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
|
|
|
|
y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
|
|
|
|
}
|
|
|
|
#else
|
|
|
|
assert(false);
|
|
|
|
#endif
|
|
|
|
|
|
|
|
}
|
|
|
|
|
2024-02-18 16:16:55 +00:00
|
|
|
template<typename dst_t>
|
|
|
|
static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
|
|
|
|
|
|
|
|
const int i = blockIdx.x;
|
|
|
|
const block_iq1_s * x = (const block_iq1_s *) vx;
|
|
|
|
|
|
|
|
const int tid = threadIdx.x;
|
|
|
|
#if QK_K == 256
|
|
|
|
const int il = tid/8; // 0...3
|
|
|
|
const int ib = tid%8; // 0...7
|
|
|
|
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
|
|
|
|
const int i8 = 4*ib+il;
|
|
|
|
uint8_t h = x[i].scales[i8/2] >> 4*(i8%2);
|
|
|
|
const int8_t * grid = (const int8_t *)(iq1s_grid + (x[i].qs[i8] | ((h & 8) << 5)));
|
|
|
|
const float d = (float)x[i].d * (2*(h & 7) + 1);
|
|
|
|
for (int j = 0; j < 8; ++j) y[j] = d * grid[j];
|
|
|
|
#else
|
|
|
|
assert(false);
|
|
|
|
#endif
|
|
|
|
|
|
|
|
}
|
|
|
|
|
2024-02-21 09:39:52 +00:00
|
|
|
static const __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
|
|
|
|
|
|
|
|
template<typename dst_t>
|
|
|
|
static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst_t * __restrict__ yy) {
|
|
|
|
|
|
|
|
const int i = blockIdx.x;
|
|
|
|
const block_iq4_nl * x = (const block_iq4_nl *) vx + i*(QK_K/QK4_NL);
|
|
|
|
|
|
|
|
const int tid = threadIdx.x;
|
|
|
|
const int il = tid/8; // 0...3
|
|
|
|
const int ib = tid%8; // 0...7
|
|
|
|
dst_t * y = yy + i*QK_K + 32*ib + 4*il;
|
|
|
|
const uint8_t * q4 = x[ib].qs + 4*il;
|
|
|
|
const float d = (float)x[ib].d;
|
|
|
|
for (int j = 0; j < 4; ++j) {
|
|
|
|
y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
|
|
|
|
y[j+16] = d * kvalues_iq4nl[q4[j] >> 4];
|
|
|
|
}
|
|
|
|
|
|
|
|
}
|
2024-02-18 16:16:55 +00:00
|
|
|
|
2024-02-28 08:37:02 +00:00
|
|
|
#if QK_K != 64
|
2024-02-27 14:34:24 +00:00
|
|
|
template<typename dst_t>
|
|
|
|
static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
|
|
|
|
const int i = blockIdx.x;
|
|
|
|
const block_iq4_xs * x = (const block_iq4_xs *)vx;
|
|
|
|
|
|
|
|
const int tid = threadIdx.x;
|
|
|
|
const int il = tid/8; // 0...3
|
|
|
|
const int ib = tid%8; // 0...7
|
|
|
|
dst_t * y = yy + i*QK_K + 32*ib + 4*il;
|
|
|
|
const uint8_t * q4 = x[i].qs + 16*ib + 4*il;
|
|
|
|
const float d = (float)x[i].d * ((((x[i].scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((x[i].scales_h >> 2*ib) & 3) << 4)) - 32);
|
|
|
|
for (int j = 0; j < 4; ++j) {
|
|
|
|
y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
|
|
|
|
y[j+16] = d * kvalues_iq4nl[q4[j] >> 4];
|
|
|
|
}
|
|
|
|
}
|
2024-02-28 08:37:02 +00:00
|
|
|
#endif
|
2024-02-27 14:34:24 +00:00
|
|
|
|
2023-07-07 22:25:15 +00:00
|
|
|
static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
|
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
|
|
|
|
2023-06-16 17:08:44 +00:00
|
|
|
static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
|
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
|
|
|
|
2023-11-03 11:13:09 +00:00
|
|
|
const int row = blockIdx.x*blockDim.y + threadIdx.y;
|
2023-06-16 17:08:44 +00:00
|
|
|
if (row > nrows) return;
|
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
|
|
|
|
2023-06-16 17:08:44 +00:00
|
|
|
const int num_blocks_per_row = ncols / QK_K;
|
|
|
|
const int ib0 = row*num_blocks_per_row;
|
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
|
|
|
|
2023-06-16 17:08:44 +00:00
|
|
|
const block_q2_K * x = (const block_q2_K *)vx + ib0;
|
|
|
|
|
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 tmp = 0; // partial sum for thread in warp
|
|
|
|
|
|
|
|
#if QK_K == 256
|
2023-06-19 15:14:09 +00:00
|
|
|
const int tid = threadIdx.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...15
|
|
|
|
const int ix = threadIdx.x%K_QUANTS_PER_ITERATION; // 0 or 0,1
|
2023-06-16 17:08:44 +00:00
|
|
|
|
|
|
|
const int step = 16/K_QUANTS_PER_ITERATION;
|
|
|
|
|
2023-06-19 15:14:09 +00:00
|
|
|
const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128...
|
|
|
|
const int in = tid - step*im; // 0...15 or 0...7
|
2023-06-16 17:08:44 +00:00
|
|
|
|
2023-06-19 15:14:09 +00:00
|
|
|
const int l0 = K_QUANTS_PER_ITERATION*in; // 0...15 or 0...14 in steps of 2
|
2023-06-16 17:08:44 +00:00
|
|
|
const int q_offset = 32*im + l0;
|
|
|
|
const int s_offset = 8*im;
|
|
|
|
const int y_offset = 128*im + l0;
|
|
|
|
|
|
|
|
uint32_t aux[4];
|
|
|
|
const uint8_t * d = (const uint8_t *)aux;
|
|
|
|
const uint8_t * m = (const uint8_t *)(aux + 2);
|
|
|
|
|
|
|
|
for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
|
|
|
|
|
|
|
|
const float * y = yy + i * QK_K + y_offset;
|
|
|
|
const uint8_t * q = x[i].qs + q_offset;
|
|
|
|
|
2023-08-25 09:09:42 +00:00
|
|
|
const float dall = __low2half(x[i].dm);
|
|
|
|
const float dmin = __high2half(x[i].dm);
|
2023-06-16 17:08:44 +00:00
|
|
|
|
|
|
|
const uint32_t * a = (const uint32_t *)(x[i].scales + s_offset);
|
|
|
|
aux[0] = a[0] & 0x0f0f0f0f;
|
|
|
|
aux[1] = a[1] & 0x0f0f0f0f;
|
|
|
|
aux[2] = (a[0] >> 4) & 0x0f0f0f0f;
|
|
|
|
aux[3] = (a[1] >> 4) & 0x0f0f0f0f;
|
|
|
|
|
|
|
|
float sum1 = 0, sum2 = 0;
|
|
|
|
for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
|
|
|
|
sum1 += y[l+ 0] * d[0] * ((q[l+ 0] >> 0) & 3)
|
|
|
|
+ y[l+32] * d[2] * ((q[l+ 0] >> 2) & 3)
|
|
|
|
+ y[l+64] * d[4] * ((q[l+ 0] >> 4) & 3)
|
|
|
|
+ y[l+96] * d[6] * ((q[l+ 0] >> 6) & 3)
|
|
|
|
+ y[l+16] * d[1] * ((q[l+16] >> 0) & 3)
|
|
|
|
+ y[l+48] * d[3] * ((q[l+16] >> 2) & 3)
|
|
|
|
+ y[l+80] * d[5] * ((q[l+16] >> 4) & 3)
|
|
|
|
+y[l+112] * d[7] * ((q[l+16] >> 6) & 3);
|
|
|
|
sum2 += y[l+ 0] * m[0] + y[l+32] * m[2] + y[l+64] * m[4] + y[ l+96] * m[6]
|
|
|
|
+ y[l+16] * m[1] + y[l+48] * m[3] + y[l+80] * m[5] + y[l+112] * m[7];
|
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
|
|
|
|
2023-06-16 17:08:44 +00:00
|
|
|
}
|
|
|
|
tmp += dall * sum1 - dmin * sum2;
|
|
|
|
|
|
|
|
}
|
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
|
|
|
|
const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...15 or 0...7
|
|
|
|
const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); // 0....1 or 0...3
|
|
|
|
const int offset = tid * K_QUANTS_PER_ITERATION;
|
|
|
|
|
|
|
|
uint32_t uaux[2];
|
|
|
|
const uint8_t * d = (const uint8_t *)uaux;
|
|
|
|
|
|
|
|
for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
|
|
|
|
|
|
|
|
const float * y = yy + i * QK_K + offset;
|
|
|
|
const uint8_t * q = x[i].qs + offset;
|
|
|
|
const uint32_t * s = (const uint32_t *)x[i].scales;
|
|
|
|
|
|
|
|
uaux[0] = s[0] & 0x0f0f0f0f;
|
|
|
|
uaux[1] = (s[0] >> 4) & 0x0f0f0f0f;
|
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
const float2 dall = __half22float2(x[i].dm);
|
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 sum1 = 0, sum2 = 0;
|
|
|
|
for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
|
|
|
|
const uint8_t ql = q[l];
|
|
|
|
sum1 += y[l+ 0] * d[0] * ((ql >> 0) & 3)
|
|
|
|
+ y[l+16] * d[1] * ((ql >> 2) & 3)
|
|
|
|
+ y[l+32] * d[2] * ((ql >> 4) & 3)
|
|
|
|
+ y[l+48] * d[3] * ((ql >> 6) & 3);
|
|
|
|
sum2 += y[l+0] * d[4] + y[l+16] * d[5] + y[l+32] * d[6] + y[l+48] * d[7];
|
|
|
|
}
|
|
|
|
tmp += dall.x * sum1 - dall.y * sum2;
|
|
|
|
}
|
|
|
|
#endif
|
2023-06-16 17:08:44 +00:00
|
|
|
|
|
|
|
// sum up partial sums and write back result
|
2024-02-27 13:22:45 +00:00
|
|
|
tmp = warp_reduce_sum(tmp);
|
2023-06-16 17:08:44 +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 (threadIdx.x == 0) {
|
2023-06-16 17:08:44 +00:00
|
|
|
dst[row] = tmp;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-07-07 22:25:15 +00:00
|
|
|
static __global__ void dequantize_mul_mat_vec_q3_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
|
2023-06-16 17:08:44 +00:00
|
|
|
|
2023-11-03 11:13:09 +00:00
|
|
|
const int row = blockIdx.x*blockDim.y + threadIdx.y;
|
2023-06-19 15:14:09 +00:00
|
|
|
if (row > nrows) return;
|
|
|
|
|
2023-06-16 17:08:44 +00:00
|
|
|
const int num_blocks_per_row = ncols / QK_K;
|
|
|
|
const int ib0 = row*num_blocks_per_row;
|
|
|
|
|
|
|
|
const block_q3_K * x = (const block_q3_K *)vx + ib0;
|
|
|
|
|
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 tmp = 0; // partial sum for thread in warp
|
|
|
|
|
|
|
|
#if QK_K == 256
|
|
|
|
|
|
|
|
const uint16_t kmask1 = 0x0303;
|
|
|
|
const uint16_t kmask2 = 0x0f0f;
|
|
|
|
|
2023-06-19 15:14:09 +00:00
|
|
|
const int tid = threadIdx.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...16
|
|
|
|
const int ix = threadIdx.x%K_QUANTS_PER_ITERATION; // 0 or 0,1
|
2023-06-16 17:08:44 +00:00
|
|
|
|
2023-06-19 15:14:09 +00:00
|
|
|
const int n = K_QUANTS_PER_ITERATION; // iterations in the inner loop
|
|
|
|
const int step = 16/K_QUANTS_PER_ITERATION;
|
|
|
|
const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128...
|
|
|
|
const int in = tid - step*im; // 0....15 or 0...7
|
2023-06-16 17:08:44 +00:00
|
|
|
|
|
|
|
const uint8_t m = 1 << (4*im);
|
|
|
|
|
2023-06-19 15:14:09 +00:00
|
|
|
const int l0 = n*in; // 0...15 or 0...14 in steps of 2
|
2023-06-16 17:08:44 +00:00
|
|
|
const int q_offset = 32*im + l0;
|
|
|
|
const int y_offset = 128*im + l0;
|
|
|
|
|
|
|
|
uint16_t utmp[4];
|
|
|
|
const int8_t * s = (const int8_t *)utmp;
|
|
|
|
|
|
|
|
const uint16_t s_shift = 4*im;
|
|
|
|
|
2023-06-19 15:14:09 +00:00
|
|
|
for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
|
2023-06-16 17:08:44 +00:00
|
|
|
|
|
|
|
const float * y = yy + i * QK_K + y_offset;
|
|
|
|
const uint8_t * q = x[i].qs + q_offset;
|
|
|
|
const uint8_t * h = x[i].hmask + l0;
|
|
|
|
|
|
|
|
const uint16_t * a = (const uint16_t *)x[i].scales;
|
|
|
|
utmp[0] = ((a[0] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 0)) & kmask1) << 4);
|
|
|
|
utmp[1] = ((a[1] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 0)) & kmask1) << 4);
|
|
|
|
utmp[2] = ((a[2] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 2)) & kmask1) << 4);
|
|
|
|
utmp[3] = ((a[3] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 2)) & kmask1) << 4);
|
|
|
|
|
|
|
|
const float d = x[i].d;
|
|
|
|
|
|
|
|
float sum = 0;
|
|
|
|
for (int l = 0; l < n; ++l) {
|
|
|
|
sum += y[l+ 0] * (s[0] - 32) * (((q[l] >> 0) & 3) - (h[l] & (m << 0) ? 0 : 4))
|
|
|
|
+ y[l+32] * (s[2] - 32) * (((q[l] >> 2) & 3) - (h[l] & (m << 1) ? 0 : 4))
|
|
|
|
+ y[l+64] * (s[4] - 32) * (((q[l] >> 4) & 3) - (h[l] & (m << 2) ? 0 : 4))
|
|
|
|
+ y[l+96] * (s[6] - 32) * (((q[l] >> 6) & 3) - (h[l] & (m << 3) ? 0 : 4));
|
|
|
|
sum += y[l+16] * (s[1] - 32) * (((q[l+16] >> 0) & 3) - (h[l+16] & (m << 0) ? 0 : 4))
|
|
|
|
+ y[l+48] * (s[3] - 32) * (((q[l+16] >> 2) & 3) - (h[l+16] & (m << 1) ? 0 : 4))
|
|
|
|
+ y[l+80] * (s[5] - 32) * (((q[l+16] >> 4) & 3) - (h[l+16] & (m << 2) ? 0 : 4))
|
|
|
|
+ y[l+112] * (s[7] - 32) * (((q[l+16] >> 6) & 3) - (h[l+16] & (m << 3) ? 0 : 4));
|
|
|
|
}
|
|
|
|
tmp += d * 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
|
|
|
|
2023-06-16 17:08:44 +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
|
|
|
#else
|
|
|
|
|
|
|
|
const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...15 or 0...7
|
|
|
|
const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); // 0....1 or 0...3
|
|
|
|
const int offset = tid * K_QUANTS_PER_ITERATION; // 0...15 or 0...14
|
|
|
|
const int in = offset/8; // 0 or 1
|
|
|
|
const int im = offset%8; // 0...7
|
|
|
|
|
|
|
|
for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
|
|
|
|
|
|
|
|
const float * y = yy + i * QK_K + offset;
|
|
|
|
const uint8_t * q = x[i].qs + offset;
|
|
|
|
const uint8_t * s = x[i].scales;
|
|
|
|
|
|
|
|
const float dall = (float)x[i].d;
|
|
|
|
|
|
|
|
float sum = 0;
|
|
|
|
for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
|
|
|
|
const uint8_t hl = x[i].hmask[im+l] >> in;
|
|
|
|
const uint8_t ql = q[l];
|
|
|
|
sum += y[l+ 0] * dall * ((s[0] & 0xF) - 8) * ((int8_t)((ql >> 0) & 3) - ((hl >> 0) & 1 ? 0 : 4))
|
|
|
|
+ y[l+16] * dall * ((s[0] >> 4) - 8) * ((int8_t)((ql >> 2) & 3) - ((hl >> 2) & 1 ? 0 : 4))
|
|
|
|
+ y[l+32] * dall * ((s[1] & 0xF) - 8) * ((int8_t)((ql >> 4) & 3) - ((hl >> 4) & 1 ? 0 : 4))
|
|
|
|
+ y[l+48] * dall * ((s[1] >> 4) - 8) * ((int8_t)((ql >> 6) & 3) - ((hl >> 6) & 1 ? 0 : 4));
|
|
|
|
}
|
|
|
|
tmp += sum;
|
|
|
|
}
|
|
|
|
#endif
|
2023-06-16 17:08:44 +00:00
|
|
|
|
|
|
|
// sum up partial sums and write back result
|
2024-02-27 13:22:45 +00:00
|
|
|
tmp = warp_reduce_sum(tmp);
|
2023-06-16 17:08:44 +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 (threadIdx.x == 0) {
|
2023-06-16 17:08:44 +00:00
|
|
|
dst[row] = tmp;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-07-07 22:25:15 +00:00
|
|
|
static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
|
2023-06-16 17:08:44 +00:00
|
|
|
|
2023-11-03 11:13:09 +00:00
|
|
|
const int row = blockIdx.x*blockDim.y + threadIdx.y;
|
2023-06-19 15:14:09 +00:00
|
|
|
if (row > nrows) return;
|
2023-06-16 17:08:44 +00:00
|
|
|
const int num_blocks_per_row = ncols / QK_K;
|
|
|
|
const int ib0 = row*num_blocks_per_row;
|
|
|
|
|
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_q4_K * x = (const block_q4_K *)vx + ib0;
|
|
|
|
|
|
|
|
#if QK_K == 256
|
|
|
|
const uint16_t kmask1 = 0x3f3f;
|
|
|
|
const uint16_t kmask2 = 0x0f0f;
|
|
|
|
const uint16_t kmask3 = 0xc0c0;
|
|
|
|
|
2023-06-19 15:14:09 +00:00
|
|
|
const int tid = threadIdx.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...16
|
|
|
|
const int ix = threadIdx.x%K_QUANTS_PER_ITERATION; // 0 or 0,1
|
2023-06-16 17:08:44 +00:00
|
|
|
|
2023-06-19 15:14:09 +00:00
|
|
|
const int step = 8/K_QUANTS_PER_ITERATION; // 8 or 4
|
|
|
|
|
|
|
|
const int il = tid/step; // 0...3
|
|
|
|
const int ir = tid - step*il; // 0...7 or 0...3
|
|
|
|
const int n = 2 * K_QUANTS_PER_ITERATION; // 2 or 4
|
2023-06-16 17:08:44 +00:00
|
|
|
|
|
|
|
const int im = il/2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
|
|
|
|
const int in = il%2;
|
|
|
|
|
|
|
|
const int l0 = n*(2*ir + in);
|
|
|
|
const int q_offset = 32*im + l0;
|
|
|
|
const int y_offset = 64*im + l0;
|
|
|
|
|
|
|
|
uint16_t aux[4];
|
|
|
|
const uint8_t * sc = (const uint8_t *)aux;
|
|
|
|
|
2023-07-23 05:49:20 +00:00
|
|
|
#if K_QUANTS_PER_ITERATION == 2
|
|
|
|
uint32_t q32[4];
|
|
|
|
const uint8_t * q4 = (const uint8_t *)q32;
|
|
|
|
#else
|
|
|
|
uint16_t q16[4];
|
|
|
|
const uint8_t * q4 = (const uint8_t *)q16;
|
|
|
|
#endif
|
|
|
|
|
2023-06-16 17:08:44 +00:00
|
|
|
float tmp = 0; // partial sum for thread in warp
|
|
|
|
|
2023-06-19 15:14:09 +00:00
|
|
|
for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
|
2023-06-16 17:08:44 +00:00
|
|
|
|
|
|
|
const float * y1 = yy + i*QK_K + y_offset;
|
|
|
|
const float * y2 = y1 + 128;
|
|
|
|
|
2023-08-25 09:09:42 +00:00
|
|
|
const float dall = __low2half(x[i].dm);
|
|
|
|
const float dmin = __high2half(x[i].dm);
|
2023-06-16 17:08:44 +00:00
|
|
|
|
|
|
|
const uint16_t * a = (const uint16_t *)x[i].scales;
|
|
|
|
aux[0] = a[im+0] & kmask1;
|
|
|
|
aux[1] = a[im+2] & kmask1;
|
|
|
|
aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
|
|
|
|
aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
|
|
|
|
|
2023-07-23 05:49:20 +00:00
|
|
|
#if K_QUANTS_PER_ITERATION == 2
|
|
|
|
const uint32_t * q1 = (const uint32_t *)(x[i].qs + q_offset);
|
|
|
|
const uint32_t * q2 = q1 + 16;
|
|
|
|
|
|
|
|
q32[0] = q1[0] & 0x0f0f0f0f;
|
|
|
|
q32[1] = q1[0] & 0xf0f0f0f0;
|
|
|
|
q32[2] = q2[0] & 0x0f0f0f0f;
|
|
|
|
q32[3] = q2[0] & 0xf0f0f0f0;
|
|
|
|
|
2023-06-16 17:08:44 +00:00
|
|
|
float4 s = {0.f, 0.f, 0.f, 0.f};
|
|
|
|
float smin = 0;
|
2023-07-23 05:49:20 +00:00
|
|
|
for (int l = 0; l < 4; ++l) {
|
|
|
|
s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+ 4];
|
|
|
|
s.z += y2[l] * q4[l+8]; s.w += y2[l+32] * q4[l+12];
|
2023-06-16 17:08:44 +00:00
|
|
|
smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
|
|
|
|
}
|
2023-07-23 05:49:20 +00:00
|
|
|
tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin;
|
|
|
|
#else
|
|
|
|
const uint16_t * q1 = (const uint16_t *)(x[i].qs + q_offset);
|
|
|
|
const uint16_t * q2 = q1 + 32;
|
|
|
|
|
|
|
|
q16[0] = q1[0] & 0x0f0f;
|
|
|
|
q16[1] = q1[0] & 0xf0f0;
|
|
|
|
q16[2] = q2[0] & 0x0f0f;
|
|
|
|
q16[3] = q2[0] & 0xf0f0;
|
|
|
|
|
|
|
|
float4 s = {0.f, 0.f, 0.f, 0.f};
|
|
|
|
float smin = 0;
|
|
|
|
for (int l = 0; l < 2; ++l) {
|
|
|
|
s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+2];
|
|
|
|
s.z += y2[l] * q4[l+4]; s.w += y2[l+32] * q4[l+6];
|
|
|
|
smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
|
|
|
|
}
|
|
|
|
tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin;
|
|
|
|
#endif
|
2023-06-16 17:08:44 +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
|
|
|
#else
|
|
|
|
const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...15
|
|
|
|
const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION);
|
|
|
|
|
|
|
|
const int step = tid * K_QUANTS_PER_ITERATION;
|
|
|
|
|
|
|
|
uint16_t aux16[2];
|
|
|
|
const uint8_t * s = (const uint8_t *)aux16;
|
|
|
|
|
|
|
|
float tmp = 0;
|
|
|
|
|
|
|
|
for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
|
|
|
|
const uint8_t * q = x[i].qs + step;
|
|
|
|
const float * y = yy + i*QK_K + step;
|
|
|
|
const uint16_t * a = (const uint16_t *)x[i].scales;
|
|
|
|
aux16[0] = a[0] & 0x0f0f;
|
|
|
|
aux16[1] = (a[0] >> 4) & 0x0f0f;
|
2023-08-27 12:19:59 +00:00
|
|
|
const float d = (float)x[i].dm[0];
|
|
|
|
const float m = (float)x[i].dm[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
|
|
|
float sum = 0.f;
|
|
|
|
for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
|
|
|
|
sum += y[j+ 0] * (d * s[0] * (q[j+ 0] & 0xF) - m * s[2])
|
|
|
|
+ y[j+16] * (d * s[0] * (q[j+16] & 0xF) - m * s[2])
|
|
|
|
+ y[j+32] * (d * s[1] * (q[j+ 0] >> 4) - m * s[3])
|
|
|
|
+ y[j+48] * (d * s[1] * (q[j+16] >> 4) - m * s[3]);
|
|
|
|
}
|
|
|
|
tmp += sum;
|
|
|
|
}
|
|
|
|
|
|
|
|
#endif
|
2023-06-16 17:08:44 +00:00
|
|
|
|
|
|
|
// sum up partial sums and write back result
|
2024-02-27 13:22:45 +00:00
|
|
|
tmp = warp_reduce_sum(tmp);
|
2023-06-16 17:08:44 +00:00
|
|
|
|
|
|
|
if (tid == 0) {
|
|
|
|
dst[row] = tmp;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-07-07 22:25:15 +00:00
|
|
|
static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols) {
|
2023-06-16 17:08:44 +00:00
|
|
|
|
|
|
|
const int row = blockIdx.x;
|
|
|
|
const int num_blocks_per_row = ncols / QK_K;
|
|
|
|
const int ib0 = row*num_blocks_per_row;
|
|
|
|
|
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_q5_K * x = (const block_q5_K *)vx + ib0;
|
|
|
|
|
|
|
|
float tmp = 0; // partial sum for thread in warp
|
|
|
|
|
|
|
|
#if QK_K == 256
|
|
|
|
const uint16_t kmask1 = 0x3f3f;
|
|
|
|
const uint16_t kmask2 = 0x0f0f;
|
|
|
|
const uint16_t kmask3 = 0xc0c0;
|
|
|
|
|
2023-06-16 17:08:44 +00:00
|
|
|
const int tid = threadIdx.x/2; // 0...15
|
|
|
|
const int ix = threadIdx.x%2;
|
|
|
|
|
|
|
|
const int il = tid/4; // 0...3
|
|
|
|
const int ir = tid - 4*il;// 0...3
|
2023-06-19 15:14:09 +00:00
|
|
|
const int n = 2;
|
2023-06-16 17:08:44 +00:00
|
|
|
|
|
|
|
const int im = il/2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
|
|
|
|
const int in = il%2;
|
|
|
|
|
|
|
|
const int l0 = n*(2*ir + in);
|
|
|
|
const int q_offset = 32*im + l0;
|
|
|
|
const int y_offset = 64*im + l0;
|
|
|
|
|
|
|
|
const uint8_t hm1 = 1 << (2*im);
|
|
|
|
const uint8_t hm2 = hm1 << 4;
|
|
|
|
|
|
|
|
uint16_t aux[4];
|
|
|
|
const uint8_t * sc = (const uint8_t *)aux;
|
|
|
|
|
2023-07-23 21:19:47 +00:00
|
|
|
uint16_t q16[8];
|
|
|
|
const uint8_t * q4 = (const uint8_t *)q16;
|
|
|
|
|
2023-06-16 17:08:44 +00:00
|
|
|
for (int i = ix; i < num_blocks_per_row; i += 2) {
|
|
|
|
|
|
|
|
const uint8_t * ql1 = x[i].qs + q_offset;
|
|
|
|
const uint8_t * qh = x[i].qh + l0;
|
|
|
|
const float * y1 = yy + i*QK_K + y_offset;
|
|
|
|
const float * y2 = y1 + 128;
|
|
|
|
|
2023-08-25 09:09:42 +00:00
|
|
|
const float dall = __low2half(x[i].dm);
|
|
|
|
const float dmin = __high2half(x[i].dm);
|
2023-06-16 17:08:44 +00:00
|
|
|
|
|
|
|
const uint16_t * a = (const uint16_t *)x[i].scales;
|
|
|
|
aux[0] = a[im+0] & kmask1;
|
|
|
|
aux[1] = a[im+2] & kmask1;
|
|
|
|
aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
|
|
|
|
aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
|
|
|
|
|
|
|
|
float4 sum = {0.f, 0.f, 0.f, 0.f};
|
|
|
|
float smin = 0;
|
2023-07-23 21:19:47 +00:00
|
|
|
const uint16_t * q1 = (const uint16_t *)ql1;
|
|
|
|
const uint16_t * q2 = q1 + 32;
|
|
|
|
q16[0] = q1[0] & 0x0f0f;
|
|
|
|
q16[1] = q1[8] & 0x0f0f;
|
|
|
|
q16[2] = (q1[0] >> 4) & 0x0f0f;
|
|
|
|
q16[3] = (q1[8] >> 4) & 0x0f0f;
|
|
|
|
q16[4] = q2[0] & 0x0f0f;
|
|
|
|
q16[5] = q2[8] & 0x0f0f;
|
|
|
|
q16[6] = (q2[0] >> 4) & 0x0f0f;
|
|
|
|
q16[7] = (q2[8] >> 4) & 0x0f0f;
|
2023-06-16 17:08:44 +00:00
|
|
|
for (int l = 0; l < n; ++l) {
|
2023-07-23 21:19:47 +00:00
|
|
|
sum.x += y1[l+ 0] * (q4[l +0] + (qh[l+ 0] & (hm1 << 0) ? 16 : 0))
|
|
|
|
+ y1[l+16] * (q4[l +2] + (qh[l+16] & (hm1 << 0) ? 16 : 0));
|
|
|
|
sum.y += y1[l+32] * (q4[l +4] + (qh[l+ 0] & (hm1 << 1) ? 16 : 0))
|
|
|
|
+ y1[l+48] * (q4[l +6] + (qh[l+16] & (hm1 << 1) ? 16 : 0));
|
|
|
|
sum.z += y2[l+ 0] * (q4[l +8] + (qh[l+ 0] & (hm2 << 0) ? 16 : 0))
|
|
|
|
+ y2[l+16] * (q4[l+10] + (qh[l+16] & (hm2 << 0) ? 16 : 0));
|
|
|
|
sum.w += y2[l+32] * (q4[l+12] + (qh[l+ 0] & (hm2 << 1) ? 16 : 0))
|
|
|
|
+ y2[l+48] * (q4[l+14] + (qh[l+16] & (hm2 << 1) ? 16 : 0));
|
2023-06-19 15:14:09 +00:00
|
|
|
smin += (y1[l] + y1[l+16]) * sc[2] + (y1[l+32] + y1[l+48]) * sc[3]
|
|
|
|
+ (y2[l] + y2[l+16]) * sc[6] + (y2[l+32] + y2[l+48]) * sc[7];
|
2023-06-16 17:08:44 +00:00
|
|
|
}
|
|
|
|
tmp += dall * (sum.x * sc[0] + sum.y * sc[1] + sum.z * sc[4] + sum.w * sc[5]) - dmin * smin;
|
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
|
|
|
}
|
2023-06-16 17:08:44 +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
|
|
|
#else
|
|
|
|
const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...15
|
|
|
|
const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION);
|
|
|
|
const int step = tid * K_QUANTS_PER_ITERATION;
|
|
|
|
const int im = step/8;
|
|
|
|
const int in = step%8;
|
|
|
|
|
|
|
|
for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
|
|
|
|
const uint8_t * q = x[i].qs + step;
|
|
|
|
const int8_t * s = x[i].scales;
|
|
|
|
const float * y = yy + i*QK_K + step;
|
|
|
|
const float d = x[i].d;
|
|
|
|
float sum = 0.f;
|
|
|
|
for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
|
|
|
|
const uint8_t h = x[i].qh[in+j] >> im;
|
|
|
|
sum += y[j+ 0] * d * s[0] * ((q[j+ 0] & 0xF) - ((h >> 0) & 1 ? 0 : 16))
|
|
|
|
+ y[j+16] * d * s[1] * ((q[j+16] & 0xF) - ((h >> 2) & 1 ? 0 : 16))
|
|
|
|
+ y[j+32] * d * s[2] * ((q[j+ 0] >> 4) - ((h >> 4) & 1 ? 0 : 16))
|
|
|
|
+ y[j+48] * d * s[3] * ((q[j+16] >> 4) - ((h >> 6) & 1 ? 0 : 16));
|
|
|
|
}
|
|
|
|
tmp += sum;
|
2023-06-16 17:08:44 +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
|
|
|
#endif
|
2023-06-16 17:08:44 +00:00
|
|
|
|
|
|
|
// sum up partial sums and write back result
|
2024-02-27 13:22:45 +00:00
|
|
|
tmp = warp_reduce_sum(tmp);
|
2023-06-16 17:08:44 +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 (threadIdx.x == 0) {
|
2023-06-16 17:08:44 +00:00
|
|
|
dst[row] = tmp;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-07-07 22:25:15 +00:00
|
|
|
static __global__ void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
|
2023-06-16 17:08:44 +00:00
|
|
|
|
|
|
|
static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
|
|
|
|
|
2023-11-03 11:13:09 +00:00
|
|
|
const int row = blockIdx.x*blockDim.y + threadIdx.y;
|
2023-06-16 17:08:44 +00:00
|
|
|
if (row > nrows) return;
|
|
|
|
|
|
|
|
const int num_blocks_per_row = ncols / QK_K;
|
|
|
|
const int ib0 = row*num_blocks_per_row;
|
|
|
|
|
|
|
|
const block_q6_K * x = (const block_q6_K *)vx + ib0;
|
|
|
|
|
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
|
|
|
|
|
2023-06-16 17:08:44 +00:00
|
|
|
const int tid = threadIdx.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...16
|
|
|
|
const int ix = threadIdx.x%K_QUANTS_PER_ITERATION; // 0 or 0, 1
|
|
|
|
|
|
|
|
const int step = 16/K_QUANTS_PER_ITERATION; // 16 or 8
|
|
|
|
|
|
|
|
const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128...
|
|
|
|
const int in = tid - step*im; // 0...15 or 0...7
|
|
|
|
|
|
|
|
#if K_QUANTS_PER_ITERATION == 1
|
|
|
|
const int l0 = K_QUANTS_PER_ITERATION*in; // 0...15
|
|
|
|
const int is = 0;
|
|
|
|
#else
|
|
|
|
const int l0 = 4 * in; // 0, 4, 8, ..., 28
|
|
|
|
const int is = in / 4;
|
|
|
|
#endif
|
|
|
|
const int ql_offset = 64*im + l0;
|
|
|
|
const int qh_offset = 32*im + l0;
|
|
|
|
const int s_offset = 8*im + is;
|
|
|
|
const int y_offset = 128*im + l0;
|
|
|
|
|
|
|
|
float tmp = 0; // partial sum for thread in warp
|
|
|
|
|
|
|
|
for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
|
|
|
|
|
|
|
|
const float * y = yy + i * QK_K + y_offset;
|
|
|
|
const uint8_t * ql = x[i].ql + ql_offset;
|
|
|
|
const uint8_t * qh = x[i].qh + qh_offset;
|
|
|
|
const int8_t * s = x[i].scales + s_offset;
|
|
|
|
|
|
|
|
const float d = x[i].d;
|
|
|
|
|
|
|
|
#if K_QUANTS_PER_ITERATION == 1
|
|
|
|
float sum = y[ 0] * s[0] * d * ((int8_t)((ql[ 0] & 0xF) | ((qh[ 0] & 0x03) << 4)) - 32)
|
|
|
|
+ y[16] * s[1] * d * ((int8_t)((ql[16] & 0xF) | ((qh[16] & 0x03) << 4)) - 32)
|
|
|
|
+ y[32] * s[2] * d * ((int8_t)((ql[32] & 0xF) | ((qh[ 0] & 0x0c) << 2)) - 32)
|
|
|
|
+ y[48] * s[3] * d * ((int8_t)((ql[48] & 0xF) | ((qh[16] & 0x0c) << 2)) - 32)
|
|
|
|
+ y[64] * s[4] * d * ((int8_t)((ql[ 0] >> 4) | ((qh[ 0] & 0x30) >> 0)) - 32)
|
|
|
|
+ y[80] * s[5] * d * ((int8_t)((ql[16] >> 4) | ((qh[16] & 0x30) >> 0)) - 32)
|
|
|
|
+ y[96] * s[6] * d * ((int8_t)((ql[32] >> 4) | ((qh[ 0] & 0xc0) >> 2)) - 32)
|
|
|
|
+y[112] * s[7] * d * ((int8_t)((ql[48] >> 4) | ((qh[16] & 0xc0) >> 2)) - 32);
|
|
|
|
tmp += sum;
|
|
|
|
#else
|
|
|
|
float sum = 0;
|
|
|
|
for (int l = 0; l < 4; ++l) {
|
|
|
|
sum += y[l+ 0] * s[0] * d * ((int8_t)((ql[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32)
|
|
|
|
+ y[l+32] * s[2] * d * ((int8_t)((ql[l+32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32)
|
|
|
|
+ y[l+64] * s[4] * d * ((int8_t)((ql[l+ 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32)
|
|
|
|
+ y[l+96] * s[6] * d * ((int8_t)((ql[l+32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32);
|
|
|
|
}
|
|
|
|
tmp += sum;
|
|
|
|
#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
|
|
|
|
|
|
|
|
const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...7
|
|
|
|
const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); // 0...3
|
|
|
|
|
|
|
|
const int step = tid * K_QUANTS_PER_ITERATION;
|
|
|
|
|
|
|
|
float tmp = 0; // partial sum for thread in warp
|
|
|
|
|
|
|
|
for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
|
|
|
|
|
|
|
|
const float * y = yy + i * QK_K + step;
|
|
|
|
const uint8_t * ql = x[i].ql + step;
|
|
|
|
const uint8_t * qh = x[i].qh + step;
|
|
|
|
const int8_t * s = x[i].scales;
|
|
|
|
|
|
|
|
const float d = x[i+0].d;
|
|
|
|
|
|
|
|
float sum = 0;
|
|
|
|
for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
|
|
|
|
sum += y[j+ 0] * s[0] * d * ((int8_t)((ql[j+ 0] & 0xF) | ((qh[j] & 0x03) << 4)) - 32)
|
|
|
|
+ y[j+16] * s[1] * d * ((int8_t)((ql[j+16] & 0xF) | ((qh[j] & 0x0c) << 2)) - 32)
|
|
|
|
+ y[j+32] * s[2] * d * ((int8_t)((ql[j+ 0] >> 4) | ((qh[j] & 0x30) >> 0)) - 32)
|
|
|
|
+ y[j+48] * s[3] * d * ((int8_t)((ql[j+16] >> 4) | ((qh[j] & 0xc0) >> 2)) - 32);
|
|
|
|
}
|
|
|
|
tmp += sum;
|
|
|
|
|
|
|
|
}
|
|
|
|
|
|
|
|
#endif
|
|
|
|
|
2023-06-16 17:08:44 +00:00
|
|
|
// sum up partial sums and write back result
|
2024-02-27 13:22:45 +00:00
|
|
|
tmp = warp_reduce_sum(tmp);
|
2023-06-16 17:08:44 +00:00
|
|
|
|
|
|
|
if (tid == 0) {
|
|
|
|
dst[row] = tmp;
|
|
|
|
}
|
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
|
|
|
}
|
|
|
|
|
2023-06-19 08:23:56 +00:00
|
|
|
static __device__ void convert_f16(const void * vx, const int ib, const int iqs, dfloat2 & v){
|
2023-05-13 13:38:36 +00:00
|
|
|
const half * x = (const half *) vx;
|
|
|
|
|
2023-06-19 08:23:56 +00:00
|
|
|
// automatic half -> float type cast if dfloat == float
|
|
|
|
v.x = x[ib + iqs + 0];
|
|
|
|
v.y = x[ib + iqs + 1];
|
2023-05-13 13:38:36 +00:00
|
|
|
}
|
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
static __global__ void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, const int kx, const int kx_padded) {
|
|
|
|
const int ix = blockDim.x*blockIdx.x + threadIdx.x;
|
2023-07-05 12:19:42 +00:00
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
if (ix >= kx_padded) {
|
2023-07-05 12:19:42 +00:00
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
const int iy = blockDim.y*blockIdx.y + threadIdx.y;
|
|
|
|
|
|
|
|
const int i_padded = iy*kx_padded + ix;
|
|
|
|
|
2023-07-05 12:19:42 +00:00
|
|
|
block_q8_1 * y = (block_q8_1 *) vy;
|
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
const int ib = i_padded / QK8_1; // block index
|
|
|
|
const int iqs = i_padded % QK8_1; // quant index
|
2023-07-05 12:19:42 +00:00
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
const float xi = ix < kx ? x[iy*kx + ix] : 0.0f;
|
2023-07-05 12:19:42 +00:00
|
|
|
float amax = fabsf(xi);
|
|
|
|
float sum = xi;
|
|
|
|
|
2024-02-27 13:22:45 +00:00
|
|
|
amax = warp_reduce_max(amax);
|
|
|
|
sum = warp_reduce_sum(sum);
|
2023-07-05 12:19:42 +00:00
|
|
|
|
|
|
|
const float d = amax / 127;
|
|
|
|
const int8_t q = amax == 0.0f ? 0 : roundf(xi / d);
|
|
|
|
|
|
|
|
y[ib].qs[iqs] = q;
|
|
|
|
|
|
|
|
if (iqs > 0) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2023-08-25 09:09:42 +00:00
|
|
|
reinterpret_cast<half&>(y[ib].ds.x) = d;
|
|
|
|
reinterpret_cast<half&>(y[ib].ds.y) = sum;
|
2023-07-05 12:19:42 +00:00
|
|
|
}
|
|
|
|
|
2023-10-08 17:19:14 +00:00
|
|
|
template<int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
|
2023-12-13 12:04:25 +00:00
|
|
|
static __global__ void k_get_rows(
|
|
|
|
const void * src0, const int32_t * src1, dst_t * dst,
|
|
|
|
int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/
|
|
|
|
/*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/
|
|
|
|
/*size_t s0,*/ size_t s1, size_t s2, size_t s3,
|
|
|
|
/*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03,
|
|
|
|
size_t s10, size_t s11, size_t s12/*, size_t s13*/) {
|
|
|
|
|
|
|
|
const int i00 = (blockIdx.x*blockDim.x + threadIdx.x)*2;
|
|
|
|
const int i10 = blockDim.y*blockIdx.y + threadIdx.y;
|
|
|
|
const int i11 = (blockIdx.z*blockDim.z + threadIdx.z)/ne12;
|
|
|
|
const int i12 = (blockIdx.z*blockDim.z + threadIdx.z)%ne12;
|
|
|
|
|
|
|
|
if (i00 >= ne00) {
|
2023-10-08 17:19:14 +00:00
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2023-12-13 12:04:25 +00:00
|
|
|
const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
|
2023-10-08 17:19:14 +00:00
|
|
|
|
2023-12-13 12:04:25 +00:00
|
|
|
dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
|
|
|
|
const void * src0_row = (const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03;
|
2023-10-08 17:19:14 +00:00
|
|
|
|
2023-12-13 12:04:25 +00:00
|
|
|
const int ib = i00/qk; // block index
|
|
|
|
const int iqs = (i00%qk)/qr; // quant index
|
|
|
|
const int iybs = i00 - i00%qk; // dst block start index
|
2023-10-08 17:19:14 +00:00
|
|
|
const int y_offset = qr == 1 ? 1 : qk/2;
|
|
|
|
|
|
|
|
// dequantize
|
|
|
|
dfloat2 v;
|
2023-12-13 12:04:25 +00:00
|
|
|
dequantize_kernel(src0_row, ib, iqs, v);
|
2023-10-08 17:19:14 +00:00
|
|
|
|
2023-12-13 12:04:25 +00:00
|
|
|
dst_row[iybs + iqs + 0] = v.x;
|
|
|
|
dst_row[iybs + iqs + y_offset] = v.y;
|
|
|
|
}
|
|
|
|
|
|
|
|
template<typename src0_t, typename dst_t>
|
|
|
|
static __global__ void k_get_rows_float(
|
|
|
|
const src0_t * src0, const int32_t * src1, dst_t * dst,
|
|
|
|
int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/
|
|
|
|
/*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/
|
|
|
|
/*size_t s0,*/ size_t s1, size_t s2, size_t s3,
|
|
|
|
/*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03,
|
|
|
|
size_t s10, size_t s11, size_t s12/*, size_t s13*/) {
|
|
|
|
|
|
|
|
const int i00 = blockIdx.x*blockDim.x + threadIdx.x;
|
|
|
|
const int i10 = blockDim.y*blockIdx.y + threadIdx.y;
|
|
|
|
const int i11 = (blockIdx.z*blockDim.z + threadIdx.z)/ne12;
|
|
|
|
const int i12 = (blockIdx.z*blockDim.z + threadIdx.z)%ne12;
|
|
|
|
|
|
|
|
if (i00 >= ne00) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
|
|
|
|
|
|
|
|
dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
|
|
|
|
const src0_t * src0_row = (const src0_t *)((const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03);
|
|
|
|
|
|
|
|
dst_row[i00] = src0_row[i00];
|
2023-10-08 17:19:14 +00:00
|
|
|
}
|
|
|
|
|
2023-09-28 10:08:28 +00:00
|
|
|
template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
|
|
|
|
static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y, const int k) {
|
2024-01-07 16:24:08 +00:00
|
|
|
const int i = 2*(blockDim.x*blockIdx.x + threadIdx.x);
|
2023-05-11 21:23:08 +00:00
|
|
|
|
2023-05-14 18:53:23 +00:00
|
|
|
if (i >= k) {
|
|
|
|
return;
|
2023-04-26 20:14:13 +00:00
|
|
|
}
|
|
|
|
|
2023-05-14 18:53:23 +00:00
|
|
|
const int ib = i/qk; // block index
|
|
|
|
const int iqs = (i%qk)/qr; // quant index
|
|
|
|
const int iybs = i - i%qk; // y block start index
|
|
|
|
const int y_offset = qr == 1 ? 1 : qk/2;
|
2023-04-25 20:40:51 +00:00
|
|
|
|
2023-05-14 18:53:23 +00:00
|
|
|
// dequantize
|
2023-06-19 08:23:56 +00:00
|
|
|
dfloat2 v;
|
|
|
|
dequantize_kernel(vx, ib, iqs, v);
|
|
|
|
|
|
|
|
y[iybs + iqs + 0] = v.x;
|
|
|
|
y[iybs + iqs + y_offset] = v.y;
|
2023-04-25 20:40:51 +00:00
|
|
|
}
|
|
|
|
|
2024-01-07 16:24:08 +00:00
|
|
|
template <typename src_t, typename dst_t>
|
|
|
|
static __global__ void convert_unary(const void * __restrict__ vx, dst_t * __restrict__ y, const int k) {
|
|
|
|
const int i = blockDim.x*blockIdx.x + threadIdx.x;
|
|
|
|
|
|
|
|
if (i >= k) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
const src_t * x = (src_t *) vx;
|
|
|
|
|
|
|
|
y[i] = x[i];
|
|
|
|
}
|
|
|
|
|
2024-01-12 19:38:54 +00:00
|
|
|
template <bool need_check>
|
|
|
|
static __global__ void dequantize_block_q8_0_f16(const void * __restrict__ vx, half * __restrict__ y, const int k) {
|
|
|
|
#if __CUDA_ARCH__ >= CC_PASCAL
|
|
|
|
constexpr int nint = CUDA_Q8_0_NE_ALIGN/sizeof(int) + WARP_SIZE;
|
|
|
|
|
|
|
|
const int i0 = CUDA_Q8_0_NE_ALIGN*blockIdx.x;
|
|
|
|
const int * x0 = ((int *) vx) + blockIdx.x * nint;
|
|
|
|
half2 * y2 = (half2 *) (y + i0);
|
|
|
|
|
|
|
|
__shared__ int vals[nint];
|
|
|
|
|
|
|
|
#pragma unroll
|
|
|
|
for (int ix0 = 0; ix0 < nint; ix0 += WARP_SIZE) {
|
|
|
|
if (need_check && i0*sizeof(block_q8_0)/QK8_0 + sizeof(int)*(ix0 + threadIdx.x) >= k*sizeof(block_q8_0)/QK8_0) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
const int ix = ix0 + threadIdx.x;
|
|
|
|
vals[ix] = x0[ix];
|
|
|
|
}
|
|
|
|
|
|
|
|
#pragma unroll
|
|
|
|
for (int iy = 0; iy < CUDA_Q8_0_NE_ALIGN; iy += 2*WARP_SIZE) {
|
|
|
|
if (need_check && i0 + iy + 2*threadIdx.x >= k) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
const half * b0 = ((const half *) vals) + (sizeof(block_q8_0)/sizeof(half)) * ((iy + 2*threadIdx.x)/QK8_0);
|
|
|
|
const half d = *b0;
|
|
|
|
const char2 qs = ((const char2 *) (b0 + 1))[threadIdx.x % (QK8_0/2)];
|
|
|
|
|
|
|
|
y2[iy/2 + threadIdx.x] = __hmul2(make_half2(qs.x, qs.y), __half2half2(d));
|
|
|
|
}
|
|
|
|
#else
|
|
|
|
(void) vx; (void) y; (void) k;
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2024-01-12 19:38:54 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= CC_PASCAL
|
|
|
|
}
|
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
// VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called
|
2023-08-02 16:04:04 +00:00
|
|
|
// MMVQ = mul_mat_vec_q, MMQ = mul_mat_q
|
2023-07-05 12:19:42 +00:00
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
#define VDR_Q4_0_Q8_1_MMVQ 2
|
|
|
|
#define VDR_Q4_0_Q8_1_MMQ 4
|
2023-07-05 12:19:42 +00:00
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
template <int vdr> static __device__ __forceinline__ float vec_dot_q4_0_q8_1_impl(
|
|
|
|
const int * v, const int * u, const float & d4, const half2 & ds8) {
|
2023-07-05 12:19:42 +00:00
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
|
2023-08-02 16:04:04 +00:00
|
|
|
int sumi = 0;
|
2023-07-05 12:19:42 +00:00
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
#pragma unroll
|
|
|
|
for (int i = 0; i < vdr; ++i) {
|
|
|
|
const int vi0 = (v[i] >> 0) & 0x0F0F0F0F;
|
|
|
|
const int vi1 = (v[i] >> 4) & 0x0F0F0F0F;
|
|
|
|
|
|
|
|
// SIMD dot product of quantized values
|
|
|
|
sumi = __dp4a(vi0, u[2*i+0], sumi);
|
|
|
|
sumi = __dp4a(vi1, u[2*i+1], sumi);
|
|
|
|
}
|
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
const float2 ds8f = __half22float2(ds8);
|
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
// second part effectively subtracts 8 from each quant value
|
2023-08-05 16:20:44 +00:00
|
|
|
return d4 * (sumi * ds8f.x - (8*vdr/QI4_0) * ds8f.y);
|
2023-08-02 16:04:04 +00:00
|
|
|
#else
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2023-08-02 16:04:04 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
|
|
|
|
}
|
|
|
|
|
|
|
|
#define VDR_Q4_1_Q8_1_MMVQ 2
|
|
|
|
#define VDR_Q4_1_Q8_1_MMQ 4
|
|
|
|
|
|
|
|
template <int vdr> static __device__ __forceinline__ float vec_dot_q4_1_q8_1_impl(
|
|
|
|
const int * v, const int * u, const half2 & dm4, const half2 & ds8) {
|
|
|
|
|
|
|
|
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
|
|
|
|
int sumi = 0;
|
|
|
|
|
|
|
|
#pragma unroll
|
|
|
|
for (int i = 0; i < vdr; ++i) {
|
|
|
|
const int vi0 = (v[i] >> 0) & 0x0F0F0F0F;
|
|
|
|
const int vi1 = (v[i] >> 4) & 0x0F0F0F0F;
|
|
|
|
|
|
|
|
// SIMD dot product of quantized values
|
|
|
|
sumi = __dp4a(vi0, u[2*i+0], sumi);
|
|
|
|
sumi = __dp4a(vi1, u[2*i+1], sumi);
|
|
|
|
}
|
|
|
|
|
|
|
|
#ifdef GGML_CUDA_F16
|
2023-08-05 16:20:44 +00:00
|
|
|
const float2 tmp = __half22float2(__hmul2(dm4, ds8));
|
|
|
|
const float d4d8 = tmp.x;
|
|
|
|
const float m4s8 = tmp.y;
|
2023-08-02 16:04:04 +00:00
|
|
|
#else
|
2023-08-05 16:20:44 +00:00
|
|
|
const float2 dm4f = __half22float2(dm4);
|
|
|
|
const float2 ds8f = __half22float2(ds8);
|
|
|
|
const float d4d8 = dm4f.x * ds8f.x;
|
|
|
|
const float m4s8 = dm4f.y * ds8f.y;
|
2023-08-02 16:04:04 +00:00
|
|
|
#endif // GGML_CUDA_F16
|
|
|
|
|
|
|
|
// scale second part of sum by QI8_1/(vdr * QR4_1) to compensate for multiple threads adding it
|
|
|
|
return sumi * d4d8 + m4s8 / (QI8_1 / (vdr * QR4_1));
|
|
|
|
#else
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2023-08-02 16:04:04 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
|
|
|
|
}
|
|
|
|
|
|
|
|
#define VDR_Q5_0_Q8_1_MMVQ 2
|
|
|
|
#define VDR_Q5_0_Q8_1_MMQ 4
|
|
|
|
|
|
|
|
template <int vdr> static __device__ __forceinline__ float vec_dot_q5_0_q8_1_impl(
|
|
|
|
const int * vl, const int * vh, const int * u, const float & d5, const half2 & ds8) {
|
|
|
|
|
|
|
|
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
|
|
|
|
int sumi = 0;
|
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
#pragma unroll
|
2023-08-02 16:04:04 +00:00
|
|
|
for (int i = 0; i < vdr; ++i) {
|
|
|
|
int vi0 = (vl[i] >> 0) & 0x0F0F0F0F; // lower 4 qs bits, still need qh as 5th bits
|
|
|
|
vi0 |= (vh[i] << 4) & 0x00000010; // 0 -> 4
|
|
|
|
vi0 |= (vh[i] << 11) & 0x00001000; // 1 -> 12
|
|
|
|
vi0 |= (vh[i] << 18) & 0x00100000; // 2 -> 20
|
|
|
|
vi0 |= (vh[i] << 25) & 0x10000000; // 3 -> 28
|
|
|
|
sumi = __dp4a(vi0, u[2*i+0], sumi); // SIMD dot product of quantized values
|
|
|
|
|
|
|
|
int vi1 = (vl[i] >> 4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits
|
|
|
|
vi1 |= (vh[i] >> 12) & 0x00000010; // 16 -> 4
|
|
|
|
vi1 |= (vh[i] >> 5) & 0x00001000; // 17 -> 12
|
|
|
|
vi1 |= (vh[i] << 2) & 0x00100000; // 18 -> 20
|
|
|
|
vi1 |= (vh[i] << 9) & 0x10000000; // 19 -> 28
|
|
|
|
sumi = __dp4a(vi1, u[2*i+1], sumi); // SIMD dot product of quantized values
|
|
|
|
}
|
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
const float2 ds8f = __half22float2(ds8);
|
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
// second part effectively subtracts 16 from each quant value
|
2023-08-05 16:20:44 +00:00
|
|
|
return d5 * (sumi * ds8f.x - (16*vdr/QI5_0) * ds8f.y);
|
2023-08-02 16:04:04 +00:00
|
|
|
#else
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2023-08-02 16:04:04 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
|
|
|
|
}
|
|
|
|
|
|
|
|
#define VDR_Q5_1_Q8_1_MMVQ 2
|
|
|
|
#define VDR_Q5_1_Q8_1_MMQ 4
|
|
|
|
|
|
|
|
template <int vdr> static __device__ __forceinline__ float vec_dot_q5_1_q8_1_impl(
|
|
|
|
const int * vl, const int * vh, const int * u, const half2 & dm5, const half2 & ds8) {
|
|
|
|
|
|
|
|
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
|
|
|
|
int sumi = 0;
|
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
#pragma unroll
|
2023-08-02 16:04:04 +00:00
|
|
|
for (int i = 0; i < vdr; ++i) {
|
|
|
|
int vi0 = (vl[i] >> 0) & 0x0F0F0F0F; // lower 4 qs bits, still need qh as 5th bits
|
|
|
|
vi0 |= (vh[i] << 4) & 0x00000010; // 0 -> 4
|
|
|
|
vi0 |= (vh[i] << 11) & 0x00001000; // 1 -> 12
|
|
|
|
vi0 |= (vh[i] << 18) & 0x00100000; // 2 -> 20
|
|
|
|
vi0 |= (vh[i] << 25) & 0x10000000; // 3 -> 28
|
|
|
|
sumi = __dp4a(vi0, u[2*i+0], sumi); // SIMD dot product of quantized values
|
|
|
|
|
|
|
|
int vi1 = (vl[i] >> 4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits
|
|
|
|
vi1 |= (vh[i] >> 12) & 0x00000010; // 16 -> 4
|
|
|
|
vi1 |= (vh[i] >> 5) & 0x00001000; // 17 -> 12
|
|
|
|
vi1 |= (vh[i] << 2) & 0x00100000; // 18 -> 20
|
|
|
|
vi1 |= (vh[i] << 9) & 0x10000000; // 19 -> 28
|
|
|
|
sumi = __dp4a(vi1, u[2*i+1], sumi); // SIMD dot product of quantized values
|
|
|
|
}
|
|
|
|
|
|
|
|
#ifdef GGML_CUDA_F16
|
2023-08-05 16:20:44 +00:00
|
|
|
const float2 tmp = __half22float2(__hmul2(dm5, ds8));
|
|
|
|
const float d5d8 = tmp.x;
|
|
|
|
const float m5s8 = tmp.y;
|
2023-08-02 16:04:04 +00:00
|
|
|
#else
|
2023-08-05 16:20:44 +00:00
|
|
|
const float2 dm5f = __half22float2(dm5);
|
|
|
|
const float2 ds8f = __half22float2(ds8);
|
|
|
|
const float d5d8 = dm5f.x * ds8f.x;
|
|
|
|
const float m5s8 = dm5f.y * ds8f.y;
|
2023-08-02 16:04:04 +00:00
|
|
|
#endif // GGML_CUDA_F16
|
|
|
|
|
|
|
|
// scale second part of sum by QI5_1 / vdr to compensate for multiple threads adding it
|
|
|
|
return sumi*d5d8 + m5s8 / (QI5_1 / vdr);
|
2023-07-05 12:19:42 +00:00
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
#else
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2023-08-02 16:04:04 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
|
|
|
|
}
|
|
|
|
|
|
|
|
#define VDR_Q8_0_Q8_1_MMVQ 2
|
|
|
|
#define VDR_Q8_0_Q8_1_MMQ 8
|
|
|
|
|
|
|
|
template <int vdr> static __device__ __forceinline__ float vec_dot_q8_0_q8_1_impl(
|
2023-08-05 16:20:44 +00:00
|
|
|
const int * v, const int * u, const float & d8_0, const float & d8_1) {
|
2023-08-02 16:04:04 +00:00
|
|
|
|
|
|
|
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
|
|
|
|
int sumi = 0;
|
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
#pragma unroll
|
2023-08-02 16:04:04 +00:00
|
|
|
for (int i = 0; i < vdr; ++i) {
|
|
|
|
// SIMD dot product of quantized values
|
|
|
|
sumi = __dp4a(v[i], u[i], sumi);
|
|
|
|
}
|
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
return d8_0*d8_1 * sumi;
|
2023-08-02 16:04:04 +00:00
|
|
|
#else
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2023-08-02 16:04:04 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
|
|
|
|
}
|
|
|
|
|
|
|
|
template <int vdr> static __device__ __forceinline__ float vec_dot_q8_1_q8_1_impl(
|
|
|
|
const int * v, const int * u, const half2 & dm8, const half2 & ds8) {
|
|
|
|
|
|
|
|
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
|
|
|
|
int sumi = 0;
|
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
#pragma unroll
|
2023-08-02 16:04:04 +00:00
|
|
|
for (int i = 0; i < vdr; ++i) {
|
|
|
|
// SIMD dot product of quantized values
|
|
|
|
sumi = __dp4a(v[i], u[i], sumi);
|
|
|
|
}
|
|
|
|
|
|
|
|
#ifdef GGML_CUDA_F16
|
2023-08-05 16:20:44 +00:00
|
|
|
const float2 tmp = __half22float2(__hmul2(dm8, ds8));
|
|
|
|
const float d8d8 = tmp.x;
|
|
|
|
const float m8s8 = tmp.y;
|
2023-08-02 16:04:04 +00:00
|
|
|
#else
|
2023-08-05 16:20:44 +00:00
|
|
|
const float2 dm8f = __half22float2(dm8);
|
|
|
|
const float2 ds8f = __half22float2(ds8);
|
2023-08-09 07:42:34 +00:00
|
|
|
const float d8d8 = dm8f.x * ds8f.x;
|
|
|
|
const float m8s8 = dm8f.y * ds8f.y;
|
2023-08-02 16:04:04 +00:00
|
|
|
#endif // GGML_CUDA_F16
|
|
|
|
|
|
|
|
// scale second part of sum by QI8_1/ vdr to compensate for multiple threads adding it
|
|
|
|
return sumi*d8d8 + m8s8 / (QI8_1 / vdr);
|
2023-07-05 12:19:42 +00:00
|
|
|
#else
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2023-07-14 17:44:08 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
|
2023-07-05 12:19:42 +00:00
|
|
|
}
|
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
#define VDR_Q2_K_Q8_1_MMVQ 1
|
|
|
|
#define VDR_Q2_K_Q8_1_MMQ 2
|
|
|
|
|
|
|
|
// contiguous v/x values
|
|
|
|
static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmvq(
|
|
|
|
const int & v, const int * __restrict__ u, const uint8_t * __restrict__ scales,
|
|
|
|
const half2 & dm2, const float * __restrict__ d8) {
|
|
|
|
|
|
|
|
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
|
|
|
|
float sumf_d = 0.0f;
|
|
|
|
float sumf_m = 0.0f;
|
|
|
|
|
|
|
|
#pragma unroll
|
|
|
|
for (int i = 0; i < QR2_K; ++i) {
|
|
|
|
const int sc = scales[2*i];
|
|
|
|
|
|
|
|
const int vi = (v >> (2*i)) & 0x03030303;
|
|
|
|
|
|
|
|
sumf_d += d8[i] * (__dp4a(vi, u[i], 0) * (sc & 0xF)); // SIMD dot product
|
|
|
|
|
|
|
|
// fill int with 4x m
|
|
|
|
int m = sc >> 4;
|
|
|
|
m |= m << 8;
|
|
|
|
m |= m << 16;
|
|
|
|
sumf_m += d8[i] * __dp4a(m, u[i], 0); // multiply constant q2_K part with sum of q8_1 values
|
|
|
|
}
|
|
|
|
|
|
|
|
const float2 dm2f = __half22float2(dm2);
|
|
|
|
|
|
|
|
return dm2f.x*sumf_d - dm2f.y*sumf_m;
|
|
|
|
#else
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2023-08-05 16:20:44 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
|
|
|
|
}
|
|
|
|
|
|
|
|
// contiguous u/y values
|
|
|
|
static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmq(
|
|
|
|
const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ scales,
|
|
|
|
const half2 & dm2, const float & d8) {
|
|
|
|
|
|
|
|
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
|
|
|
|
int sumi_d = 0;
|
|
|
|
int sumi_m = 0;
|
|
|
|
|
|
|
|
#pragma unroll
|
|
|
|
for (int i0 = 0; i0 < QI8_1; i0 += QI8_1/2) {
|
|
|
|
int sumi_d_sc = 0;
|
|
|
|
|
|
|
|
const int sc = scales[i0 / (QI8_1/2)];
|
|
|
|
|
|
|
|
// fill int with 4x m
|
|
|
|
int m = sc >> 4;
|
|
|
|
m |= m << 8;
|
|
|
|
m |= m << 16;
|
|
|
|
|
|
|
|
#pragma unroll
|
|
|
|
for (int i = i0; i < i0 + QI8_1/2; ++i) {
|
|
|
|
sumi_d_sc = __dp4a(v[i], u[i], sumi_d_sc); // SIMD dot product
|
|
|
|
sumi_m = __dp4a(m, u[i], sumi_m); // multiply sum of q8_1 values with m
|
|
|
|
}
|
|
|
|
|
|
|
|
sumi_d += sumi_d_sc * (sc & 0xF);
|
|
|
|
}
|
|
|
|
|
|
|
|
const float2 dm2f = __half22float2(dm2);
|
|
|
|
|
|
|
|
return d8 * (dm2f.x*sumi_d - dm2f.y*sumi_m);
|
|
|
|
#else
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2023-08-05 16:20:44 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
|
|
|
|
}
|
|
|
|
|
|
|
|
#define VDR_Q3_K_Q8_1_MMVQ 1
|
|
|
|
#define VDR_Q3_K_Q8_1_MMQ 2
|
|
|
|
|
|
|
|
// contiguous v/x values
|
|
|
|
static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmvq(
|
|
|
|
const int & vl, const int & vh, const int * __restrict__ u, const uint8_t * __restrict__ scales,
|
|
|
|
const int & scale_offset, const float & d3, const float * __restrict__ d8) {
|
|
|
|
|
|
|
|
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
|
|
|
|
float sumf = 0.0f;
|
|
|
|
|
|
|
|
#pragma unroll
|
|
|
|
for (int i = 0; i < QR3_K; ++i) {
|
|
|
|
const int isc = scale_offset + 2*i;
|
|
|
|
|
|
|
|
const int isc_low = isc % (QK_K/32);
|
|
|
|
const int sc_shift_low = 4 * (isc / (QK_K/32));
|
|
|
|
const int sc_low = (scales[isc_low] >> sc_shift_low) & 0xF;
|
|
|
|
|
|
|
|
const int isc_high = isc % (QK_K/64);
|
|
|
|
const int sc_shift_high = 2 * (isc / (QK_K/64));
|
|
|
|
const int sc_high = ((scales[(QK_K/32) + isc_high] >> sc_shift_high) & 3) << 4;
|
|
|
|
|
|
|
|
const int sc = (sc_low | sc_high) - 32;
|
|
|
|
|
|
|
|
const int vil = (vl >> (2*i)) & 0x03030303;
|
|
|
|
|
|
|
|
const int vih = ((vh >> i) << 2) & 0x04040404;
|
|
|
|
|
|
|
|
const int vi = __vsubss4(vil, vih);
|
|
|
|
|
|
|
|
sumf += d8[i] * (__dp4a(vi, u[i], 0) * sc); // SIMD dot product
|
|
|
|
}
|
|
|
|
|
|
|
|
return d3 * sumf;
|
|
|
|
#else
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2023-08-05 16:20:44 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
|
|
|
|
}
|
|
|
|
|
|
|
|
// contiguous u/y values
|
|
|
|
static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmq(
|
|
|
|
const int * __restrict__ v, const int * __restrict__ u, const int8_t * __restrict__ scales,
|
|
|
|
const float & d3, const float & d8) {
|
|
|
|
|
|
|
|
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
|
|
|
|
int sumi = 0;
|
|
|
|
|
|
|
|
#pragma unroll
|
|
|
|
for (int i0 = 0; i0 < QR3_K*VDR_Q3_K_Q8_1_MMQ; i0 += QI8_1/2) {
|
|
|
|
int sumi_sc = 0;
|
|
|
|
|
|
|
|
for (int i = i0; i < i0 + QI8_1/2; ++i) {
|
|
|
|
sumi_sc = __dp4a(v[i], u[i], sumi_sc); // SIMD dot product
|
|
|
|
}
|
|
|
|
|
|
|
|
sumi += sumi_sc * scales[i0 / (QI8_1/2)];
|
|
|
|
}
|
|
|
|
|
|
|
|
return d3*d8 * sumi;
|
|
|
|
#else
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2023-08-05 16:20:44 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
|
|
|
|
}
|
|
|
|
|
|
|
|
#define VDR_Q4_K_Q8_1_MMVQ 2
|
|
|
|
#define VDR_Q4_K_Q8_1_MMQ 8
|
|
|
|
|
|
|
|
// contiguous v/x values
|
|
|
|
static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_vmmq(
|
|
|
|
const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc,
|
|
|
|
const uint8_t * __restrict__ m, const half2 & dm4, const float * __restrict__ d8) {
|
|
|
|
|
|
|
|
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
|
|
|
|
float sumf_d = 0.0f;
|
|
|
|
float sumf_m = 0.0f;
|
|
|
|
|
|
|
|
#pragma unroll
|
|
|
|
for (int i = 0; i < QR4_K; ++i) {
|
|
|
|
const int v0i = (v[0] >> (4*i)) & 0x0F0F0F0F;
|
|
|
|
const int v1i = (v[1] >> (4*i)) & 0x0F0F0F0F;
|
|
|
|
|
|
|
|
const int dot1 = __dp4a(v1i, u[2*i+1], __dp4a(v0i, u[2*i+0], 0)); // SIMD dot product
|
|
|
|
const int dot2 = __dp4a(0x01010101, u[2*i+1], __dp4a(0x01010101, u[2*i+0], 0)); // sum of u
|
|
|
|
|
|
|
|
sumf_d += d8[i] * (dot1 * sc[i]);
|
|
|
|
sumf_m += d8[i] * (dot2 * m[i]); // multiply constant part of q4_K with sum of q8_1 values
|
|
|
|
}
|
|
|
|
|
|
|
|
const float2 dm4f = __half22float2(dm4);
|
|
|
|
|
|
|
|
return dm4f.x*sumf_d - dm4f.y*sumf_m;
|
|
|
|
|
|
|
|
#else
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2023-08-05 16:20:44 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
|
|
|
|
}
|
|
|
|
|
|
|
|
// contiguous u/y values
|
|
|
|
static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_mmq(
|
|
|
|
const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc,
|
|
|
|
const uint8_t * __restrict__ m, const half2 & dm4, const half2 * __restrict__ ds8) {
|
|
|
|
|
|
|
|
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
|
|
|
|
float sumf_d = 0.0f;
|
|
|
|
float sumf_m = 0.0f;
|
|
|
|
|
|
|
|
#pragma unroll
|
2023-08-14 08:41:22 +00:00
|
|
|
for (int i = 0; i < QR4_K*VDR_Q4_K_Q8_1_MMQ/QI8_1; ++i) {
|
2023-08-05 16:20:44 +00:00
|
|
|
int sumi_d = 0;
|
|
|
|
|
|
|
|
#pragma unroll
|
2023-08-14 08:41:22 +00:00
|
|
|
for (int j = 0; j < QI8_1; ++j) {
|
|
|
|
sumi_d = __dp4a((v[j] >> (4*i)) & 0x0F0F0F0F, u[i*QI8_1 + j], sumi_d); // SIMD dot product
|
2023-08-05 16:20:44 +00:00
|
|
|
}
|
|
|
|
|
2023-08-14 08:41:22 +00:00
|
|
|
const float2 ds8f = __half22float2(ds8[i]);
|
2023-08-05 16:20:44 +00:00
|
|
|
|
2023-08-14 08:41:22 +00:00
|
|
|
sumf_d += ds8f.x * (sc[i] * sumi_d);
|
|
|
|
sumf_m += ds8f.y * m[i]; // sum of q8_1 block * q4_K min val
|
2023-08-05 16:20:44 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
const float2 dm4f = __half22float2(dm4);
|
|
|
|
|
|
|
|
return dm4f.x*sumf_d - dm4f.y*sumf_m;
|
|
|
|
|
|
|
|
#else
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2023-08-05 16:20:44 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
|
|
|
|
}
|
|
|
|
|
|
|
|
#define VDR_Q5_K_Q8_1_MMVQ 2
|
|
|
|
#define VDR_Q5_K_Q8_1_MMQ 8
|
|
|
|
|
|
|
|
// contiguous v/x values
|
2023-08-14 08:41:22 +00:00
|
|
|
static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_vmmq(
|
2023-08-05 16:20:44 +00:00
|
|
|
const int * __restrict__ vl, const int * __restrict__ vh, const int * __restrict__ u, const uint8_t * __restrict__ sc,
|
|
|
|
const uint8_t * __restrict__ m, const half2 & dm5, const float * __restrict__ d8) {
|
|
|
|
|
|
|
|
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
|
|
|
|
float sumf_d = 0.0f;
|
|
|
|
float sumf_m = 0.0f;
|
|
|
|
|
|
|
|
#pragma unroll
|
|
|
|
for (int i = 0; i < QR5_K; ++i) {
|
|
|
|
const int vl0i = (vl[0] >> (4*i)) & 0x0F0F0F0F;
|
|
|
|
const int vl1i = (vl[1] >> (4*i)) & 0x0F0F0F0F;
|
|
|
|
|
|
|
|
const int vh0i = ((vh[0] >> i) << 4) & 0x10101010;
|
|
|
|
const int vh1i = ((vh[1] >> i) << 4) & 0x10101010;
|
|
|
|
|
|
|
|
const int v0i = vl0i | vh0i;
|
|
|
|
const int v1i = vl1i | vh1i;
|
|
|
|
|
|
|
|
const int dot1 = __dp4a(v0i, u[2*i+0], __dp4a(v1i, u[2*i+1], 0)); // SIMD dot product
|
|
|
|
const int dot2 = __dp4a(0x01010101, u[2*i+0], __dp4a(0x01010101, u[2*i+1], 0)); // sum of u
|
|
|
|
|
|
|
|
sumf_d += d8[i] * (dot1 * sc[i]);
|
|
|
|
sumf_m += d8[i] * (dot2 * m[i]);
|
|
|
|
|
|
|
|
}
|
|
|
|
|
|
|
|
const float2 dm5f = __half22float2(dm5);
|
|
|
|
|
|
|
|
return dm5f.x*sumf_d - dm5f.y*sumf_m;
|
|
|
|
|
|
|
|
#else
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2023-08-05 16:20:44 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
|
|
|
|
}
|
|
|
|
|
2023-08-14 08:41:22 +00:00
|
|
|
// contiguous u/y values
|
|
|
|
static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_mmq(
|
|
|
|
const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc,
|
|
|
|
const uint8_t * __restrict__ m, const half2 & dm4, const half2 * __restrict__ ds8) {
|
|
|
|
|
|
|
|
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
|
|
|
|
float sumf_d = 0.0f;
|
|
|
|
float sumf_m = 0.0f;
|
|
|
|
|
|
|
|
#pragma unroll
|
|
|
|
for (int i = 0; i < QR5_K*VDR_Q5_K_Q8_1_MMQ/QI8_1; ++i) {
|
|
|
|
int sumi_d = 0;
|
|
|
|
|
|
|
|
#pragma unroll
|
|
|
|
for (int j = 0; j < QI8_1; ++j) {
|
|
|
|
sumi_d = __dp4a(v[i*QI8_1 + j], u[i*QI8_1 + j], sumi_d); // SIMD dot product
|
|
|
|
}
|
|
|
|
|
|
|
|
const float2 ds8f = __half22float2(ds8[i]);
|
|
|
|
|
|
|
|
sumf_d += ds8f.x * (sc[i] * sumi_d);
|
|
|
|
sumf_m += ds8f.y * m[i]; // sum of q8_1 block * q4_K min val
|
|
|
|
}
|
|
|
|
|
|
|
|
const float2 dm4f = __half22float2(dm4);
|
|
|
|
|
|
|
|
return dm4f.x*sumf_d - dm4f.y*sumf_m;
|
|
|
|
|
|
|
|
#else
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2023-08-14 08:41:22 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
|
|
|
|
}
|
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
#define VDR_Q6_K_Q8_1_MMVQ 1
|
|
|
|
#define VDR_Q6_K_Q8_1_MMQ 8
|
|
|
|
|
|
|
|
// contiguous v/x values
|
|
|
|
static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmvq(
|
|
|
|
const int & vl, const int & vh, const int * __restrict__ u, const int8_t * __restrict__ scales,
|
|
|
|
const float & d, const float * __restrict__ d8) {
|
|
|
|
|
|
|
|
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
|
|
|
|
float sumf = 0.0f;
|
|
|
|
|
|
|
|
#pragma unroll
|
|
|
|
for (int i = 0; i < QR6_K; ++i) {
|
|
|
|
const int sc = scales[4*i];
|
|
|
|
|
|
|
|
const int vil = (vl >> (4*i)) & 0x0F0F0F0F;
|
|
|
|
|
|
|
|
const int vih = ((vh >> (4*i)) << 4) & 0x30303030;
|
|
|
|
|
|
|
|
const int vi = __vsubss4((vil | vih), 0x20202020); // vi = (vil | vih) - 32
|
|
|
|
|
|
|
|
sumf += d8[i] * (__dp4a(vi, u[i], 0) * sc); // SIMD dot product
|
|
|
|
}
|
|
|
|
|
|
|
|
return d*sumf;
|
|
|
|
#else
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2023-08-05 16:20:44 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
|
|
|
|
}
|
|
|
|
|
|
|
|
// contiguous u/y values
|
|
|
|
static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmq(
|
|
|
|
const int * __restrict__ v, const int * __restrict__ u, const int8_t * __restrict__ sc,
|
|
|
|
const float & d6, const float * __restrict__ d8) {
|
|
|
|
|
|
|
|
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
|
|
|
|
float sumf_d = 0.0f;
|
|
|
|
|
|
|
|
#pragma unroll
|
|
|
|
for (int i0 = 0; i0 < VDR_Q6_K_Q8_1_MMQ; i0 += 4) {
|
|
|
|
int2 sumi_d = {0, 0}; // 2 q6_K scales per q8_1 scale
|
|
|
|
|
|
|
|
#pragma unroll
|
|
|
|
for (int i = i0; i < i0 + 2; ++i) {
|
|
|
|
sumi_d.x = __dp4a(v[2*i+0], u[2*i+0], sumi_d.x); // SIMD dot product
|
|
|
|
sumi_d.x = __dp4a(v[2*i+1], u[2*i+1], sumi_d.x); // SIMD dot product
|
|
|
|
|
|
|
|
sumi_d.y = __dp4a(v[2*i+4], u[2*i+4], sumi_d.y); // SIMD dot product
|
|
|
|
sumi_d.y = __dp4a(v[2*i+5], u[2*i+5], sumi_d.y); // SIMD dot product
|
|
|
|
}
|
|
|
|
|
|
|
|
sumf_d += d8[i0/4] * (sc[i0/2+0]*sumi_d.x + sc[i0/2+1]*sumi_d.y);
|
|
|
|
}
|
|
|
|
|
|
|
|
return d6 * sumf_d;
|
|
|
|
|
|
|
|
#else
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2023-08-05 16:20:44 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
|
|
|
|
}
|
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
static __device__ __forceinline__ float vec_dot_q4_0_q8_1(
|
|
|
|
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
|
|
|
|
|
|
|
|
const block_q4_0 * bq4_0 = (const block_q4_0 *) vbq;
|
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
int v[VDR_Q4_0_Q8_1_MMVQ];
|
|
|
|
int u[2*VDR_Q4_0_Q8_1_MMVQ];
|
2023-07-05 12:19:42 +00:00
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
#pragma unroll
|
|
|
|
for (int i = 0; i < VDR_Q4_0_Q8_1_MMVQ; ++i) {
|
|
|
|
v[i] = get_int_from_uint8(bq4_0->qs, iqs + i);
|
|
|
|
u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
|
|
|
|
u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_0);
|
|
|
|
}
|
|
|
|
|
|
|
|
return vec_dot_q4_0_q8_1_impl<VDR_Q4_0_Q8_1_MMVQ>(v, u, bq4_0->d, bq8_1->ds);
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q4_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
|
2023-11-18 15:11:18 +00:00
|
|
|
(void)x_qh; (void)x_sc;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
__shared__ int tile_x_qs[mmq_y * (WARP_SIZE) + mmq_y];
|
|
|
|
__shared__ float tile_x_d[mmq_y * (WARP_SIZE/QI4_0) + mmq_y/QI4_0];
|
2023-07-29 21:04:44 +00:00
|
|
|
|
|
|
|
*x_ql = tile_x_qs;
|
2023-08-02 16:04:04 +00:00
|
|
|
*x_dm = (half2 *) tile_x_d;
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
2023-07-05 12:19:42 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q4_0(
|
2023-07-29 21:04:44 +00:00
|
|
|
const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
|
2023-08-02 14:48:10 +00:00
|
|
|
int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
|
2023-11-18 15:11:18 +00:00
|
|
|
(void)x_qh; (void)x_sc;
|
2023-09-16 14:55:43 +00:00
|
|
|
GGML_CUDA_ASSUME(i_offset >= 0);
|
|
|
|
GGML_CUDA_ASSUME(i_offset < nwarps);
|
|
|
|
GGML_CUDA_ASSUME(k >= 0);
|
|
|
|
GGML_CUDA_ASSUME(k < WARP_SIZE);
|
2023-07-05 12:19:42 +00:00
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
const int kbx = k / QI4_0;
|
|
|
|
const int kqsx = k % QI4_0;
|
|
|
|
|
2023-11-18 15:11:18 +00:00
|
|
|
const block_q4_0 * bx0 = (const block_q4_0 *) vx;
|
2023-07-31 11:18:51 +00:00
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
float * x_dmf = (float *) x_dm;
|
|
|
|
|
2023-07-31 11:18:51 +00:00
|
|
|
#pragma unroll
|
2023-08-09 07:42:34 +00:00
|
|
|
for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
|
2023-08-02 14:48:10 +00:00
|
|
|
int i = i0 + i_offset;
|
|
|
|
|
|
|
|
if (need_check) {
|
|
|
|
i = min(i, i_max);
|
|
|
|
}
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
const block_q4_0 * bxi = bx0 + i*blocks_per_row + kbx;
|
|
|
|
|
|
|
|
x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8(bxi->qs, kqsx);
|
2023-08-09 07:42:34 +00:00
|
|
|
// x_dmf[i * (WARP_SIZE/QI4_0) + i / QI4_0 + kbx] = bxi->d;
|
2023-07-31 11:18:51 +00:00
|
|
|
}
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
const int blocks_per_tile_x_row = WARP_SIZE / QI4_0;
|
|
|
|
const int kbxd = k % blocks_per_tile_x_row;
|
2023-07-31 11:18:51 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
#pragma unroll
|
|
|
|
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_0) {
|
|
|
|
int i = i0 + i_offset * QI4_0 + k / blocks_per_tile_x_row;
|
2023-07-31 11:18:51 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
if (need_check) {
|
|
|
|
i = min(i, i_max);
|
|
|
|
}
|
2023-07-31 11:18:51 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
const block_q4_0 * bxi = bx0 + i*blocks_per_row + kbxd;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
x_dmf[i * (WARP_SIZE/QI4_0) + i / QI4_0 + kbxd] = bxi->d;
|
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static __device__ __forceinline__ float vec_dot_q4_0_q8_1_mul_mat(
|
|
|
|
const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
|
|
|
|
const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
|
2023-11-18 15:11:18 +00:00
|
|
|
(void)x_qh; (void)x_sc;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
|
|
|
const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2));
|
2023-11-18 15:11:18 +00:00
|
|
|
const float * x_dmf = (const float *) x_dm;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
int u[2*VDR_Q4_0_Q8_1_MMQ];
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
#pragma unroll
|
|
|
|
for (int l = 0; l < VDR_Q4_0_Q8_1_MMQ; ++l) {
|
2023-08-09 07:42:34 +00:00
|
|
|
u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l) % WARP_SIZE];
|
|
|
|
u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI4_0) % WARP_SIZE];
|
2023-08-02 16:04:04 +00:00
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
return vec_dot_q4_0_q8_1_impl<VDR_Q4_0_Q8_1_MMQ>
|
|
|
|
(&x_ql[i * (WARP_SIZE + 1) + k], u, x_dmf[i * (WARP_SIZE/QI4_0) + i/QI4_0 + k/QI4_0],
|
2023-08-09 07:42:34 +00:00
|
|
|
y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]);
|
2023-07-05 12:19:42 +00:00
|
|
|
}
|
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
static __device__ __forceinline__ float vec_dot_q4_1_q8_1(
|
|
|
|
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
|
|
|
|
|
|
|
|
const block_q4_1 * bq4_1 = (const block_q4_1 *) vbq;
|
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
int v[VDR_Q4_1_Q8_1_MMVQ];
|
|
|
|
int u[2*VDR_Q4_1_Q8_1_MMVQ];
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
#pragma unroll
|
|
|
|
for (int i = 0; i < VDR_Q4_1_Q8_1_MMVQ; ++i) {
|
|
|
|
v[i] = get_int_from_uint8_aligned(bq4_1->qs, iqs + i);
|
|
|
|
u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
|
|
|
|
u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_1);
|
|
|
|
}
|
|
|
|
|
|
|
|
return vec_dot_q4_1_q8_1_impl<VDR_Q4_1_Q8_1_MMVQ>(v, u, bq4_1->dm, bq8_1->ds);
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q4_1(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
|
2023-11-18 15:11:18 +00:00
|
|
|
(void)x_qh; (void)x_sc;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
__shared__ int tile_x_qs[mmq_y * (WARP_SIZE) + + mmq_y];
|
|
|
|
__shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI4_1) + mmq_y/QI4_1];
|
2023-07-29 21:04:44 +00:00
|
|
|
|
|
|
|
*x_ql = tile_x_qs;
|
|
|
|
*x_dm = tile_x_dm;
|
|
|
|
}
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q4_1(
|
2023-07-29 21:04:44 +00:00
|
|
|
const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
|
2023-08-02 14:48:10 +00:00
|
|
|
int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
|
2023-11-18 15:11:18 +00:00
|
|
|
(void)x_qh; (void)x_sc;
|
2023-07-31 11:18:51 +00:00
|
|
|
|
2023-09-16 14:55:43 +00:00
|
|
|
GGML_CUDA_ASSUME(i_offset >= 0);
|
|
|
|
GGML_CUDA_ASSUME(i_offset < nwarps);
|
|
|
|
GGML_CUDA_ASSUME(k >= 0);
|
|
|
|
GGML_CUDA_ASSUME(k < WARP_SIZE);
|
2023-07-29 21:04:44 +00:00
|
|
|
|
|
|
|
const int kbx = k / QI4_1;
|
|
|
|
const int kqsx = k % QI4_1;
|
|
|
|
|
2023-11-18 15:11:18 +00:00
|
|
|
const block_q4_1 * bx0 = (const block_q4_1 *) vx;
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
#pragma unroll
|
2023-08-09 07:42:34 +00:00
|
|
|
for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
|
2023-08-02 14:48:10 +00:00
|
|
|
int i = i0 + i_offset;
|
|
|
|
|
|
|
|
if (need_check) {
|
|
|
|
i = min(i, i_max);
|
|
|
|
}
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
const block_q4_1 * bxi = bx0 + i*blocks_per_row + kbx;
|
|
|
|
|
|
|
|
x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx);
|
|
|
|
}
|
|
|
|
|
|
|
|
const int blocks_per_tile_x_row = WARP_SIZE / QI4_1;
|
|
|
|
const int kbxd = k % blocks_per_tile_x_row;
|
|
|
|
|
|
|
|
#pragma unroll
|
2023-08-09 07:42:34 +00:00
|
|
|
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_1) {
|
2023-08-02 14:48:10 +00:00
|
|
|
int i = i0 + i_offset * QI4_1 + k / blocks_per_tile_x_row;
|
|
|
|
|
|
|
|
if (need_check) {
|
|
|
|
i = min(i, i_max);
|
|
|
|
}
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
const block_q4_1 * bxi = bx0 + i*blocks_per_row + kbxd;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-07-31 11:18:51 +00:00
|
|
|
x_dm[i * (WARP_SIZE/QI4_1) + i / QI4_1 + kbxd] = bxi->dm;
|
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static __device__ __forceinline__ float vec_dot_q4_1_q8_1_mul_mat(
|
|
|
|
const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
|
|
|
|
const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
|
2023-11-18 15:11:18 +00:00
|
|
|
(void)x_qh; (void)x_sc;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
|
|
|
const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2));
|
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
int u[2*VDR_Q4_1_Q8_1_MMQ];
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
#pragma unroll
|
|
|
|
for (int l = 0; l < VDR_Q4_1_Q8_1_MMQ; ++l) {
|
2023-08-09 07:42:34 +00:00
|
|
|
u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l) % WARP_SIZE];
|
|
|
|
u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI4_1) % WARP_SIZE];
|
2023-08-02 16:04:04 +00:00
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
return vec_dot_q4_1_q8_1_impl<VDR_Q4_1_Q8_1_MMQ>
|
|
|
|
(&x_ql[i * (WARP_SIZE + 1) + k], u, x_dm[i * (WARP_SIZE/QI4_1) + i/QI4_1 + k/QI4_1],
|
2023-08-09 07:42:34 +00:00
|
|
|
y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]);
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static __device__ __forceinline__ float vec_dot_q5_0_q8_1(
|
|
|
|
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
|
|
|
|
|
2023-07-05 12:19:42 +00:00
|
|
|
const block_q5_0 * bq5_0 = (const block_q5_0 *) vbq;
|
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
int vl[VDR_Q5_0_Q8_1_MMVQ];
|
|
|
|
int vh[VDR_Q5_0_Q8_1_MMVQ];
|
|
|
|
int u[2*VDR_Q5_0_Q8_1_MMVQ];
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
#pragma unroll
|
|
|
|
for (int i = 0; i < VDR_Q5_0_Q8_1_MMVQ; ++i) {
|
|
|
|
vl[i] = get_int_from_uint8(bq5_0->qs, iqs + i);
|
|
|
|
vh[i] = get_int_from_uint8(bq5_0->qh, 0) >> (4 * (iqs + i));
|
|
|
|
u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
|
|
|
|
u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI5_0);
|
|
|
|
}
|
|
|
|
|
|
|
|
return vec_dot_q5_0_q8_1_impl<VDR_Q5_0_Q8_1_MMVQ>(vl, vh, u, bq5_0->d, bq8_1->ds);
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q5_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
|
2023-11-18 15:11:18 +00:00
|
|
|
(void)x_qh; (void)x_sc;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
__shared__ int tile_x_ql[mmq_y * (2*WARP_SIZE) + mmq_y];
|
|
|
|
__shared__ float tile_x_d[mmq_y * (WARP_SIZE/QI5_0) + mmq_y/QI5_0];
|
2023-07-29 21:04:44 +00:00
|
|
|
|
|
|
|
*x_ql = tile_x_ql;
|
2023-08-02 16:04:04 +00:00
|
|
|
*x_dm = (half2 *) tile_x_d;
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q5_0(
|
2023-07-29 21:04:44 +00:00
|
|
|
const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
|
2023-08-02 14:48:10 +00:00
|
|
|
int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
|
2023-11-18 15:11:18 +00:00
|
|
|
(void)x_qh; (void)x_sc;
|
2023-07-31 11:18:51 +00:00
|
|
|
|
2023-09-16 14:55:43 +00:00
|
|
|
GGML_CUDA_ASSUME(i_offset >= 0);
|
|
|
|
GGML_CUDA_ASSUME(i_offset < nwarps);
|
|
|
|
GGML_CUDA_ASSUME(k >= 0);
|
|
|
|
GGML_CUDA_ASSUME(k < WARP_SIZE);
|
2023-07-29 21:04:44 +00:00
|
|
|
|
|
|
|
const int kbx = k / QI5_0;
|
|
|
|
const int kqsx = k % QI5_0;
|
|
|
|
|
2023-11-18 15:11:18 +00:00
|
|
|
const block_q5_0 * bx0 = (const block_q5_0 *) vx;
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
#pragma unroll
|
2023-08-09 07:42:34 +00:00
|
|
|
for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
|
2023-08-02 14:48:10 +00:00
|
|
|
int i = i0 + i_offset;
|
|
|
|
|
|
|
|
if (need_check) {
|
|
|
|
i = min(i, i_max);
|
|
|
|
}
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
const block_q5_0 * bxi = bx0 + i*blocks_per_row + kbx;
|
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
const int ql = get_int_from_uint8(bxi->qs, kqsx);
|
|
|
|
const int qh = get_int_from_uint8(bxi->qh, 0) >> (4 * (k % QI5_0));
|
|
|
|
|
|
|
|
int qs0 = (ql >> 0) & 0x0F0F0F0F;
|
|
|
|
qs0 |= (qh << 4) & 0x00000010; // 0 -> 4
|
|
|
|
qs0 |= (qh << 11) & 0x00001000; // 1 -> 12
|
|
|
|
qs0 |= (qh << 18) & 0x00100000; // 2 -> 20
|
|
|
|
qs0 |= (qh << 25) & 0x10000000; // 3 -> 28
|
|
|
|
qs0 = __vsubss4(qs0, 0x10101010); // subtract 16
|
|
|
|
|
|
|
|
x_ql[i * (2*WARP_SIZE + 1) + 2*k+0] = qs0;
|
|
|
|
|
|
|
|
int qs1 = (ql >> 4) & 0x0F0F0F0F;
|
|
|
|
qs1 |= (qh >> 12) & 0x00000010; // 16 -> 4
|
|
|
|
qs1 |= (qh >> 5) & 0x00001000; // 17 -> 12
|
|
|
|
qs1 |= (qh << 2) & 0x00100000; // 18 -> 20
|
|
|
|
qs1 |= (qh << 9) & 0x10000000; // 19 -> 28
|
|
|
|
qs1 = __vsubss4(qs1, 0x10101010); // subtract 16
|
|
|
|
|
|
|
|
x_ql[i * (2*WARP_SIZE + 1) + 2*k+1] = qs1;
|
2023-07-31 11:18:51 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
const int blocks_per_tile_x_row = WARP_SIZE / QI5_0;
|
|
|
|
const int kbxd = k % blocks_per_tile_x_row;
|
2023-08-02 16:04:04 +00:00
|
|
|
float * x_dmf = (float *) x_dm;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-07-31 11:18:51 +00:00
|
|
|
#pragma unroll
|
2023-08-09 07:42:34 +00:00
|
|
|
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_0) {
|
2023-08-02 14:48:10 +00:00
|
|
|
int i = i0 + i_offset * QI5_0 + k / blocks_per_tile_x_row;
|
|
|
|
|
|
|
|
if (need_check) {
|
|
|
|
i = min(i, i_max);
|
|
|
|
}
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
const block_q5_0 * bxi = bx0 + i*blocks_per_row + kbxd;
|
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
x_dmf[i * (WARP_SIZE/QI5_0) + i / QI5_0 + kbxd] = bxi->d;
|
2023-07-31 11:18:51 +00:00
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static __device__ __forceinline__ float vec_dot_q5_0_q8_1_mul_mat(
|
|
|
|
const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
|
|
|
|
const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
|
2023-11-18 15:11:18 +00:00
|
|
|
(void)x_qh; (void)x_sc;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
|
|
|
const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2));
|
2023-07-31 11:18:51 +00:00
|
|
|
const int index_bx = i * (WARP_SIZE/QI5_0) + i/QI5_0 + k/QI5_0;
|
2023-08-05 16:20:44 +00:00
|
|
|
const float * x_dmf = (const float *) x_dm;
|
|
|
|
const float * y_df = (const float *) y_ds;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
int u[2*VDR_Q5_0_Q8_1_MMQ];
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
#pragma unroll
|
|
|
|
for (int l = 0; l < VDR_Q5_0_Q8_1_MMQ; ++l) {
|
2023-08-09 07:42:34 +00:00
|
|
|
u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l) % WARP_SIZE];
|
|
|
|
u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI5_0) % WARP_SIZE];
|
2023-08-02 16:04:04 +00:00
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
return vec_dot_q8_0_q8_1_impl<QR5_0*VDR_Q5_0_Q8_1_MMQ>
|
2023-08-09 07:42:34 +00:00
|
|
|
(&x_ql[i * (2*WARP_SIZE + 1) + 2 * k], u, x_dmf[index_bx], y_df[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]);
|
2023-07-05 12:19:42 +00:00
|
|
|
}
|
|
|
|
|
2023-07-14 17:44:08 +00:00
|
|
|
static __device__ __forceinline__ float vec_dot_q5_1_q8_1(
|
2023-07-29 21:04:44 +00:00
|
|
|
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
|
|
|
|
|
2023-07-05 12:19:42 +00:00
|
|
|
const block_q5_1 * bq5_1 = (const block_q5_1 *) vbq;
|
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
int vl[VDR_Q5_1_Q8_1_MMVQ];
|
|
|
|
int vh[VDR_Q5_1_Q8_1_MMVQ];
|
|
|
|
int u[2*VDR_Q5_1_Q8_1_MMVQ];
|
|
|
|
|
|
|
|
#pragma unroll
|
|
|
|
for (int i = 0; i < VDR_Q5_1_Q8_1_MMVQ; ++i) {
|
|
|
|
vl[i] = get_int_from_uint8_aligned(bq5_1->qs, iqs + i);
|
|
|
|
vh[i] = get_int_from_uint8_aligned(bq5_1->qh, 0) >> (4 * (iqs + i));
|
|
|
|
u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
|
|
|
|
u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI5_1);
|
|
|
|
}
|
2023-07-05 12:19:42 +00:00
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
return vec_dot_q5_1_q8_1_impl<VDR_Q5_1_Q8_1_MMVQ>(vl, vh, u, bq5_1->dm, bq8_1->ds);
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q5_1(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
|
2023-11-18 15:11:18 +00:00
|
|
|
(void)x_qh; (void)x_sc;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
__shared__ int tile_x_ql[mmq_y * (2*WARP_SIZE) + mmq_y];
|
|
|
|
__shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI5_1) + mmq_y/QI5_1];
|
2023-07-29 21:04:44 +00:00
|
|
|
|
|
|
|
*x_ql = tile_x_ql;
|
|
|
|
*x_dm = tile_x_dm;
|
|
|
|
}
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q5_1(
|
2023-07-29 21:04:44 +00:00
|
|
|
const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
|
2023-08-02 14:48:10 +00:00
|
|
|
int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
|
2023-11-18 15:11:18 +00:00
|
|
|
(void)x_qh; (void)x_sc;
|
2023-07-31 11:18:51 +00:00
|
|
|
|
2023-09-16 14:55:43 +00:00
|
|
|
GGML_CUDA_ASSUME(i_offset >= 0);
|
|
|
|
GGML_CUDA_ASSUME(i_offset < nwarps);
|
|
|
|
GGML_CUDA_ASSUME(k >= 0);
|
|
|
|
GGML_CUDA_ASSUME(k < WARP_SIZE);
|
2023-07-29 21:04:44 +00:00
|
|
|
|
|
|
|
const int kbx = k / QI5_1;
|
|
|
|
const int kqsx = k % QI5_1;
|
|
|
|
|
2023-11-18 15:11:18 +00:00
|
|
|
const block_q5_1 * bx0 = (const block_q5_1 *) vx;
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
#pragma unroll
|
2023-08-09 07:42:34 +00:00
|
|
|
for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
|
2023-08-02 14:48:10 +00:00
|
|
|
int i = i0 + i_offset;
|
|
|
|
|
|
|
|
if (need_check) {
|
|
|
|
i = min(i, i_max);
|
|
|
|
}
|
2023-07-05 12:19:42 +00:00
|
|
|
|
2023-07-31 11:18:51 +00:00
|
|
|
const block_q5_1 * bxi = bx0 + i*blocks_per_row + kbx;
|
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
const int ql = get_int_from_uint8_aligned(bxi->qs, kqsx);
|
|
|
|
const int qh = get_int_from_uint8_aligned(bxi->qh, 0) >> (4 * (k % QI5_1));
|
|
|
|
|
|
|
|
int qs0 = (ql >> 0) & 0x0F0F0F0F;
|
|
|
|
qs0 |= (qh << 4) & 0x00000010; // 0 -> 4
|
|
|
|
qs0 |= (qh << 11) & 0x00001000; // 1 -> 12
|
|
|
|
qs0 |= (qh << 18) & 0x00100000; // 2 -> 20
|
|
|
|
qs0 |= (qh << 25) & 0x10000000; // 3 -> 28
|
|
|
|
|
|
|
|
x_ql[i * (2*WARP_SIZE + 1) + 2*k+0] = qs0;
|
|
|
|
|
|
|
|
int qs1 = (ql >> 4) & 0x0F0F0F0F;
|
|
|
|
qs1 |= (qh >> 12) & 0x00000010; // 16 -> 4
|
|
|
|
qs1 |= (qh >> 5) & 0x00001000; // 17 -> 12
|
|
|
|
qs1 |= (qh << 2) & 0x00100000; // 18 -> 20
|
|
|
|
qs1 |= (qh << 9) & 0x10000000; // 19 -> 28
|
|
|
|
|
|
|
|
x_ql[i * (2*WARP_SIZE + 1) + 2*k+1] = qs1;
|
2023-07-31 11:18:51 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
const int blocks_per_tile_x_row = WARP_SIZE / QI5_1;
|
|
|
|
const int kbxd = k % blocks_per_tile_x_row;
|
|
|
|
|
|
|
|
#pragma unroll
|
2023-08-09 07:42:34 +00:00
|
|
|
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_1) {
|
2023-08-02 14:48:10 +00:00
|
|
|
int i = i0 + i_offset * QI5_1 + k / blocks_per_tile_x_row;
|
|
|
|
|
|
|
|
if (need_check) {
|
|
|
|
i = min(i, i_max);
|
|
|
|
}
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
const block_q5_1 * bxi = bx0 + i*blocks_per_row + kbxd;
|
|
|
|
|
|
|
|
x_dm[i * (WARP_SIZE/QI5_1) + i / QI5_1 + kbxd] = bxi->dm;
|
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static __device__ __forceinline__ float vec_dot_q5_1_q8_1_mul_mat(
|
|
|
|
const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
|
|
|
|
const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
|
2023-11-18 15:11:18 +00:00
|
|
|
(void)x_qh; (void)x_sc;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
|
|
|
const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2));
|
2023-07-31 11:18:51 +00:00
|
|
|
const int index_bx = i * (WARP_SIZE/QI5_1) + + i/QI5_1 + k/QI5_1;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
int u[2*VDR_Q5_1_Q8_1_MMQ];
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
#pragma unroll
|
|
|
|
for (int l = 0; l < VDR_Q5_1_Q8_1_MMQ; ++l) {
|
2023-08-09 07:42:34 +00:00
|
|
|
u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l) % WARP_SIZE];
|
|
|
|
u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI5_1) % WARP_SIZE];
|
2023-08-02 16:04:04 +00:00
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
return vec_dot_q8_1_q8_1_impl<QR5_1*VDR_Q5_1_Q8_1_MMQ>
|
2023-08-09 07:42:34 +00:00
|
|
|
(&x_ql[i * (2*WARP_SIZE + 1) + 2 * k], u, x_dm[index_bx], y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]);
|
2023-07-05 12:19:42 +00:00
|
|
|
}
|
|
|
|
|
2023-07-14 17:44:08 +00:00
|
|
|
static __device__ __forceinline__ float vec_dot_q8_0_q8_1(
|
2023-07-29 21:04:44 +00:00
|
|
|
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
|
|
|
|
|
2023-07-05 12:19:42 +00:00
|
|
|
const block_q8_0 * bq8_0 = (const block_q8_0 *) vbq;
|
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
int v[VDR_Q8_0_Q8_1_MMVQ];
|
|
|
|
int u[VDR_Q8_0_Q8_1_MMVQ];
|
2023-07-05 12:19:42 +00:00
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
#pragma unroll
|
2023-08-02 16:04:04 +00:00
|
|
|
for (int i = 0; i < VDR_Q8_0_Q8_1_MMVQ; ++i) {
|
|
|
|
v[i] = get_int_from_int8(bq8_0->qs, iqs + i);
|
|
|
|
u[i] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
|
|
|
|
}
|
|
|
|
|
2023-08-25 09:09:42 +00:00
|
|
|
return vec_dot_q8_0_q8_1_impl<VDR_Q8_0_Q8_1_MMVQ>(v, u, bq8_0->d, __low2half(bq8_1->ds));
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
2023-07-05 12:19:42 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q8_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
|
2023-11-18 15:11:18 +00:00
|
|
|
(void)x_qh; (void)x_sc;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
__shared__ int tile_x_qs[mmq_y * (WARP_SIZE) + mmq_y];
|
|
|
|
__shared__ float tile_x_d[mmq_y * (WARP_SIZE/QI8_0) + mmq_y/QI8_0];
|
2023-07-29 21:04:44 +00:00
|
|
|
|
|
|
|
*x_ql = tile_x_qs;
|
2023-08-02 16:04:04 +00:00
|
|
|
*x_dm = (half2 *) tile_x_d;
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q8_0(
|
2023-07-29 21:04:44 +00:00
|
|
|
const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
|
2023-08-02 14:48:10 +00:00
|
|
|
int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
|
2023-11-18 15:11:18 +00:00
|
|
|
(void)x_qh; (void)x_sc;
|
2023-07-31 11:18:51 +00:00
|
|
|
|
2023-09-16 14:55:43 +00:00
|
|
|
GGML_CUDA_ASSUME(i_offset >= 0);
|
|
|
|
GGML_CUDA_ASSUME(i_offset < nwarps);
|
|
|
|
GGML_CUDA_ASSUME(k >= 0);
|
|
|
|
GGML_CUDA_ASSUME(k < WARP_SIZE);
|
2023-07-29 21:04:44 +00:00
|
|
|
|
|
|
|
const int kbx = k / QI8_0;
|
|
|
|
const int kqsx = k % QI8_0;
|
2023-08-02 16:04:04 +00:00
|
|
|
float * x_dmf = (float *) x_dm;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-11-18 15:11:18 +00:00
|
|
|
const block_q8_0 * bx0 = (const block_q8_0 *) vx;
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
#pragma unroll
|
2023-08-09 07:42:34 +00:00
|
|
|
for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
|
2023-08-02 14:48:10 +00:00
|
|
|
int i = i0 + i_offset;
|
|
|
|
|
|
|
|
if (need_check) {
|
|
|
|
i = min(i, i_max);
|
|
|
|
}
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
const block_q8_0 * bxi = bx0 + i*blocks_per_row + kbx;
|
|
|
|
|
|
|
|
x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_int8(bxi->qs, kqsx);
|
|
|
|
}
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
const int blocks_per_tile_x_row = WARP_SIZE / QI8_0;
|
|
|
|
const int kbxd = k % blocks_per_tile_x_row;
|
2023-07-31 11:18:51 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
#pragma unroll
|
|
|
|
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI8_0) {
|
|
|
|
int i = i0 + i_offset * QI8_0 + k / blocks_per_tile_x_row;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
if (need_check) {
|
|
|
|
i = min(i, i_max);
|
|
|
|
}
|
2023-07-31 11:18:51 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
const block_q8_0 * bxi = bx0 + i*blocks_per_row + kbxd;
|
2023-07-31 11:18:51 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
x_dmf[i * (WARP_SIZE/QI8_0) + i / QI8_0 + kbxd] = bxi->d;
|
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static __device__ __forceinline__ float vec_dot_q8_0_q8_1_mul_mat(
|
|
|
|
const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
|
|
|
|
const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
|
2023-11-18 15:11:18 +00:00
|
|
|
(void)x_qh; (void)x_sc;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
const float * x_dmf = (const float *) x_dm;
|
|
|
|
const float * y_df = (const float *) y_ds;
|
2023-08-02 16:04:04 +00:00
|
|
|
|
|
|
|
return vec_dot_q8_0_q8_1_impl<VDR_Q8_0_Q8_1_MMQ>
|
|
|
|
(&x_ql[i * (WARP_SIZE + 1) + k], &y_qs[j * WARP_SIZE + k], x_dmf[i * (WARP_SIZE/QI8_0) + i/QI8_0 + k/QI8_0],
|
2023-08-05 16:20:44 +00:00
|
|
|
y_df[j * (WARP_SIZE/QI8_1) + k/QI8_1]);
|
2023-07-14 17:44:08 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static __device__ __forceinline__ float vec_dot_q2_K_q8_1(
|
2023-07-29 21:04:44 +00:00
|
|
|
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
|
2023-07-14 17:44:08 +00:00
|
|
|
|
|
|
|
const block_q2_K * bq2_K = (const block_q2_K *) vbq;
|
|
|
|
|
|
|
|
const int bq8_offset = QR2_K * (iqs / QI8_1);
|
|
|
|
const int scale_offset = iqs - iqs % QI8_1 + (iqs % QI8_1) / (QI8_1/2);
|
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
const uint8_t * scales = bq2_K->scales + scale_offset;
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
const int v = get_int_from_uint8_aligned(bq2_K->qs, iqs);
|
2023-08-05 16:20:44 +00:00
|
|
|
int u[QR2_K];
|
2023-07-29 21:04:44 +00:00
|
|
|
float d8[QR2_K];
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
#pragma unroll
|
2023-07-29 21:04:44 +00:00
|
|
|
for (int i = 0; i < QR2_K; ++ i) {
|
|
|
|
u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1);
|
2023-08-25 09:09:42 +00:00
|
|
|
d8[i] = __low2half(bq8_1[bq8_offset + i].ds);
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
return vec_dot_q2_K_q8_1_impl_mmvq(v, u, scales, bq2_K->dm, d8);
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q2_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
|
2023-11-18 15:11:18 +00:00
|
|
|
(void)x_qh;
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
__shared__ int tile_x_ql[mmq_y * (WARP_SIZE) + mmq_y];
|
|
|
|
__shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI2_K) + mmq_y/QI2_K];
|
|
|
|
__shared__ int tile_x_sc[mmq_y * (WARP_SIZE/4) + mmq_y/4];
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
*x_ql = tile_x_ql;
|
|
|
|
*x_dm = tile_x_dm;
|
|
|
|
*x_sc = tile_x_sc;
|
|
|
|
}
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q2_K(
|
2023-07-29 21:04:44 +00:00
|
|
|
const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
|
2023-08-02 14:48:10 +00:00
|
|
|
int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
|
2023-11-18 15:11:18 +00:00
|
|
|
(void)x_qh;
|
2023-07-31 11:18:51 +00:00
|
|
|
|
2023-09-16 14:55:43 +00:00
|
|
|
GGML_CUDA_ASSUME(i_offset >= 0);
|
|
|
|
GGML_CUDA_ASSUME(i_offset < nwarps);
|
|
|
|
GGML_CUDA_ASSUME(k >= 0);
|
|
|
|
GGML_CUDA_ASSUME(k < WARP_SIZE);
|
2023-07-29 21:04:44 +00:00
|
|
|
|
|
|
|
const int kbx = k / QI2_K;
|
|
|
|
const int kqsx = k % QI2_K;
|
|
|
|
|
2023-11-18 15:11:18 +00:00
|
|
|
const block_q2_K * bx0 = (const block_q2_K *) vx;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-07-31 11:18:51 +00:00
|
|
|
#pragma unroll
|
2023-08-09 07:42:34 +00:00
|
|
|
for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
|
2023-08-02 14:48:10 +00:00
|
|
|
int i = i0 + i_offset;
|
|
|
|
|
|
|
|
if (need_check) {
|
|
|
|
i = min(i, i_max);
|
|
|
|
}
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
const block_q2_K * bxi = bx0 + i*blocks_per_row + kbx;
|
|
|
|
|
|
|
|
x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx);
|
|
|
|
}
|
|
|
|
|
|
|
|
const int blocks_per_tile_x_row = WARP_SIZE / QI2_K;
|
|
|
|
const int kbxd = k % blocks_per_tile_x_row;
|
|
|
|
|
|
|
|
#pragma unroll
|
2023-08-09 07:42:34 +00:00
|
|
|
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI2_K) {
|
|
|
|
int i = (i0 + i_offset * QI2_K + k / blocks_per_tile_x_row) % mmq_y;
|
2023-08-02 14:48:10 +00:00
|
|
|
|
|
|
|
if (need_check) {
|
|
|
|
i = min(i, i_max);
|
|
|
|
}
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
const block_q2_K * bxi = bx0 + i*blocks_per_row + kbxd;
|
|
|
|
|
|
|
|
x_dm[i * (WARP_SIZE/QI2_K) + i / QI2_K + kbxd] = bxi->dm;
|
|
|
|
}
|
|
|
|
|
|
|
|
#pragma unroll
|
2023-08-09 07:42:34 +00:00
|
|
|
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 4) {
|
2023-08-02 14:48:10 +00:00
|
|
|
int i = i0 + i_offset * 4 + k / (WARP_SIZE/4);
|
|
|
|
|
|
|
|
if (need_check) {
|
|
|
|
i = min(i, i_max);
|
|
|
|
}
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
const block_q2_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/4)) / (QI2_K/4);
|
|
|
|
|
|
|
|
x_sc[i * (WARP_SIZE/4) + i / 4 + k % (WARP_SIZE/4)] = get_int_from_uint8_aligned(bxi->scales, k % (QI2_K/4));
|
|
|
|
}
|
2023-07-14 17:44:08 +00:00
|
|
|
}
|
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
static __device__ __forceinline__ float vec_dot_q2_K_q8_1_mul_mat(
|
|
|
|
const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
|
|
|
|
const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
|
2023-11-18 15:11:18 +00:00
|
|
|
(void)x_qh;
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
const int kbx = k / QI2_K;
|
|
|
|
const int ky = (k % QI2_K) * QR2_K;
|
|
|
|
const float * y_df = (const float *) y_ds;
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
int v[QR2_K*VDR_Q2_K_Q8_1_MMQ];
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
const int kqsx = i * (WARP_SIZE + 1) + kbx*QI2_K + (QI2_K/2) * (ky/(2*QI2_K)) + ky % (QI2_K/2);
|
|
|
|
const int shift = 2 * ((ky % (2*QI2_K)) / (QI2_K/2));
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
#pragma unroll
|
|
|
|
for (int l = 0; l < QR2_K*VDR_Q2_K_Q8_1_MMQ; ++l) {
|
|
|
|
v[l] = (x_ql[kqsx + l] >> shift) & 0x03030303;
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
const uint8_t * scales = ((const uint8_t *) &x_sc[i * (WARP_SIZE/4) + i/4 + kbx*4]) + ky/4;
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
const int index_y = j * WARP_SIZE + (QR2_K*k) % WARP_SIZE;
|
2023-08-05 16:20:44 +00:00
|
|
|
return vec_dot_q2_K_q8_1_impl_mmq(v, &y_qs[index_y], scales, x_dm[i * (WARP_SIZE/QI2_K) + i/QI2_K + kbx], y_df[index_y/QI8_1]);
|
2023-07-14 17:44:08 +00:00
|
|
|
}
|
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
static __device__ __forceinline__ float vec_dot_q3_K_q8_1(
|
|
|
|
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
const block_q3_K * bq3_K = (const block_q3_K *) vbq;
|
|
|
|
|
|
|
|
const int bq8_offset = QR3_K * (iqs / (QI3_K/2));
|
|
|
|
const int scale_offset = iqs - iqs % QI8_1 + (iqs % QI8_1) / (QI8_1/2);
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
const float d = bq3_K->d;
|
|
|
|
|
|
|
|
const int vl = get_int_from_uint8(bq3_K->qs, iqs);
|
|
|
|
|
|
|
|
// invert the mask with ~ so that a 0/1 results in 4/0 being subtracted
|
|
|
|
const int vh = ~get_int_from_uint8(bq3_K->hmask, iqs % (QI3_K/2)) >> bq8_offset;
|
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
int u[QR3_K];
|
2023-07-29 21:04:44 +00:00
|
|
|
float d8[QR3_K];
|
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
#pragma unroll
|
2023-07-29 21:04:44 +00:00
|
|
|
for (int i = 0; i < QR3_K; ++i) {
|
|
|
|
u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1);
|
2023-08-25 09:09:42 +00:00
|
|
|
d8[i] = __low2half(bq8_1[bq8_offset + i].ds);
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
return vec_dot_q3_K_q8_1_impl_mmvq(vl, vh, u, bq3_K->scales, scale_offset, d, d8);
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q3_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
__shared__ int tile_x_ql[mmq_y * (WARP_SIZE) + mmq_y];
|
|
|
|
__shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI3_K) + mmq_y/QI3_K];
|
|
|
|
__shared__ int tile_x_qh[mmq_y * (WARP_SIZE/2) + mmq_y/2];
|
|
|
|
__shared__ int tile_x_sc[mmq_y * (WARP_SIZE/4) + mmq_y/4];
|
2023-07-29 21:04:44 +00:00
|
|
|
|
|
|
|
*x_ql = tile_x_ql;
|
|
|
|
*x_dm = tile_x_dm;
|
|
|
|
*x_qh = tile_x_qh;
|
|
|
|
*x_sc = tile_x_sc;
|
|
|
|
}
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q3_K(
|
2023-07-29 21:04:44 +00:00
|
|
|
const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
|
2023-08-02 14:48:10 +00:00
|
|
|
int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
|
2023-07-31 11:18:51 +00:00
|
|
|
|
2023-09-16 14:55:43 +00:00
|
|
|
GGML_CUDA_ASSUME(i_offset >= 0);
|
|
|
|
GGML_CUDA_ASSUME(i_offset < nwarps);
|
|
|
|
GGML_CUDA_ASSUME(k >= 0);
|
|
|
|
GGML_CUDA_ASSUME(k < WARP_SIZE);
|
2023-07-29 21:04:44 +00:00
|
|
|
|
|
|
|
const int kbx = k / QI3_K;
|
|
|
|
const int kqsx = k % QI3_K;
|
|
|
|
|
2023-11-18 15:11:18 +00:00
|
|
|
const block_q3_K * bx0 = (const block_q3_K *) vx;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-07-31 11:18:51 +00:00
|
|
|
#pragma unroll
|
2023-08-09 07:42:34 +00:00
|
|
|
for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
|
2023-08-02 14:48:10 +00:00
|
|
|
int i = i0 + i_offset;
|
|
|
|
|
|
|
|
if (need_check) {
|
|
|
|
i = min(i, i_max);
|
|
|
|
}
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
const block_q3_K * bxi = bx0 + i*blocks_per_row + kbx;
|
|
|
|
|
|
|
|
x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8(bxi->qs, kqsx);
|
|
|
|
}
|
|
|
|
|
|
|
|
const int blocks_per_tile_x_row = WARP_SIZE / QI3_K;
|
|
|
|
const int kbxd = k % blocks_per_tile_x_row;
|
2023-08-05 16:20:44 +00:00
|
|
|
float * x_dmf = (float *) x_dm;
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
#pragma unroll
|
2023-08-09 07:42:34 +00:00
|
|
|
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI3_K) {
|
|
|
|
int i = (i0 + i_offset * QI3_K + k / blocks_per_tile_x_row) % mmq_y;
|
2023-08-02 14:48:10 +00:00
|
|
|
|
|
|
|
if (need_check) {
|
|
|
|
i = min(i, i_max);
|
|
|
|
}
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
const block_q3_K * bxi = bx0 + i*blocks_per_row + kbxd;
|
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
x_dmf[i * (WARP_SIZE/QI3_K) + i / QI3_K + kbxd] = bxi->d;
|
2023-07-31 11:18:51 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
#pragma unroll
|
2023-08-09 07:42:34 +00:00
|
|
|
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 2) {
|
2023-08-02 14:48:10 +00:00
|
|
|
int i = i0 + i_offset * 2 + k / (WARP_SIZE/2);
|
|
|
|
|
|
|
|
if (need_check) {
|
|
|
|
i = min(i, i_max);
|
|
|
|
}
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
const block_q3_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/2)) / (QI3_K/2);
|
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
// invert the mask with ~ so that a 0/1 results in 4/0 being subtracted
|
|
|
|
x_qh[i * (WARP_SIZE/2) + i / 2 + k % (WARP_SIZE/2)] = ~get_int_from_uint8(bxi->hmask, k % (QI3_K/2));
|
2023-07-31 11:18:51 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
#pragma unroll
|
2023-08-09 07:42:34 +00:00
|
|
|
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 4) {
|
2023-08-02 14:48:10 +00:00
|
|
|
int i = i0 + i_offset * 4 + k / (WARP_SIZE/4);
|
|
|
|
|
|
|
|
if (need_check) {
|
|
|
|
i = min(i, i_max);
|
|
|
|
}
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
const block_q3_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/4)) / (QI3_K/4);
|
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
const int ksc = k % (QI3_K/4);
|
|
|
|
|
|
|
|
const int ksc_low = ksc % (QI3_K/8);
|
|
|
|
const int shift_low = 4 * (ksc / (QI3_K/8));
|
|
|
|
const int sc_low = (get_int_from_uint8(bxi->scales, ksc_low) >> shift_low) & 0x0F0F0F0F;
|
|
|
|
|
|
|
|
const int ksc_high = QI3_K/8;
|
|
|
|
const int shift_high = 2 * ksc;
|
|
|
|
const int sc_high = ((get_int_from_uint8(bxi->scales, ksc_high) >> shift_high) << 4) & 0x30303030;
|
|
|
|
|
|
|
|
const int sc = __vsubss4(sc_low | sc_high, 0x20202020);
|
|
|
|
|
|
|
|
x_sc[i * (WARP_SIZE/4) + i / 4 + k % (WARP_SIZE/4)] = sc;
|
2023-07-31 11:18:51 +00:00
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static __device__ __forceinline__ float vec_dot_q3_K_q8_1_mul_mat(
|
|
|
|
const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
|
|
|
|
const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
|
|
|
|
|
|
|
|
const int kbx = k / QI3_K;
|
2023-08-05 16:20:44 +00:00
|
|
|
const int ky = (k % QI3_K) * QR3_K;
|
|
|
|
const float * x_dmf = (const float *) x_dm;
|
|
|
|
const float * y_df = (const float *) y_ds;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-11-18 15:11:18 +00:00
|
|
|
const int8_t * scales = ((const int8_t *) (x_sc + i * (WARP_SIZE/4) + i/4 + kbx*4)) + ky/4;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
int v[QR3_K*VDR_Q3_K_Q8_1_MMQ];
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
#pragma unroll
|
|
|
|
for (int l = 0; l < QR3_K*VDR_Q3_K_Q8_1_MMQ; ++l) {
|
|
|
|
const int kqsx = i * (WARP_SIZE + 1) + kbx*QI3_K + (QI3_K/2) * (ky/(2*QI3_K)) + ky % (QI3_K/2);
|
|
|
|
const int shift = 2 * ((ky % 32) / 8);
|
|
|
|
const int vll = (x_ql[kqsx + l] >> shift) & 0x03030303;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
const int vh = x_qh[i * (WARP_SIZE/2) + i/2 + kbx * (QI3_K/2) + (ky+l)%8] >> ((ky+l) / 8);
|
|
|
|
const int vlh = (vh << 2) & 0x04040404;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
v[l] = __vsubss4(vll, vlh);
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
const int index_y = j * WARP_SIZE + (k*QR3_K) % WARP_SIZE;
|
2023-08-05 16:20:44 +00:00
|
|
|
return vec_dot_q3_K_q8_1_impl_mmq(v, &y_qs[index_y], scales, x_dmf[i * (WARP_SIZE/QI3_K) + i/QI3_K + kbx], y_df[index_y/QI8_1]);
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
|
|
|
|
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
|
|
|
|
|
2023-07-25 10:48:04 +00:00
|
|
|
#ifndef GGML_QKK_64
|
2023-07-29 21:04:44 +00:00
|
|
|
const block_q4_K * bq4_K = (const block_q4_K *) vbq;
|
|
|
|
|
|
|
|
int v[2];
|
|
|
|
int u[2*QR4_K];
|
|
|
|
float d8[QR4_K];
|
2023-07-25 10:48:04 +00:00
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
// iqs is in 0,2..30. bq8_offset = iqs/4 -> bq8_offset = 0, 2, 4, 6
|
|
|
|
const int bq8_offset = QR4_K * ((iqs/2) / (QI8_1/2));
|
2023-07-25 10:48:04 +00:00
|
|
|
|
2023-07-23 21:19:47 +00:00
|
|
|
// iqs = 0....3 -> bq8_offset = 0, want q4_offset = 0, 4, 8, 12
|
|
|
|
// iqs = 4....7 -> bq8_offset = 2, want q4_offset = 32, 36, 40, 44
|
|
|
|
// iqs = 8...11 -> bq8_offset = 4, want q4_offset = 64, 68, 72, 76
|
|
|
|
// iqs = 12..15 -> bq8_offset = 6, want q4_offset = 96, 100, 104, 108
|
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
const int * q4 = (const int *)(bq4_K->qs + 16 * bq8_offset + 4 * ((iqs/2)%4));
|
2023-07-29 21:04:44 +00:00
|
|
|
v[0] = q4[0];
|
|
|
|
v[1] = q4[4];
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-07-23 05:49:20 +00:00
|
|
|
const uint16_t * scales = (const uint16_t *)bq4_K->scales;
|
|
|
|
uint16_t aux[2];
|
|
|
|
const int j = bq8_offset/2;
|
|
|
|
if (j < 2) {
|
|
|
|
aux[0] = scales[j+0] & 0x3f3f;
|
|
|
|
aux[1] = scales[j+2] & 0x3f3f;
|
|
|
|
} else {
|
|
|
|
aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2);
|
|
|
|
aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2);
|
|
|
|
}
|
|
|
|
const uint8_t * sc = (const uint8_t *)aux;
|
|
|
|
const uint8_t * m = sc + 2;
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-07-23 05:49:20 +00:00
|
|
|
for (int i = 0; i < QR4_K; ++i) {
|
2023-07-14 17:44:08 +00:00
|
|
|
const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
|
2023-08-25 09:09:42 +00:00
|
|
|
d8[i] = __low2half(bq8i->ds);
|
2023-07-23 21:19:47 +00:00
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4);
|
2023-07-29 21:04:44 +00:00
|
|
|
u[2*i+0] = q8[0];
|
|
|
|
u[2*i+1] = q8[4];
|
2023-07-14 17:44:08 +00:00
|
|
|
}
|
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
return vec_dot_q4_K_q8_1_impl_vmmq(v, u, sc, m, bq4_K->dm, d8);
|
2023-07-25 10:48:04 +00:00
|
|
|
|
|
|
|
#else
|
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
|
|
|
|
const block_q4_K * bq4_K = (const block_q4_K *) vbq;
|
|
|
|
|
|
|
|
float sumf_d = 0.0f;
|
|
|
|
float sumf_m = 0.0f;
|
|
|
|
|
2023-07-25 10:48:04 +00:00
|
|
|
uint16_t aux16[2];
|
|
|
|
const uint8_t * s = (const uint8_t *)aux16;
|
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
const uint16_t * a = (const uint16_t *)bq4_K->scales;
|
|
|
|
aux16[0] = a[0] & 0x0f0f;
|
|
|
|
aux16[1] = (a[0] >> 4) & 0x0f0f;
|
|
|
|
|
2023-08-27 12:19:59 +00:00
|
|
|
const float dall = bq4_K->dm[0];
|
|
|
|
const float dmin = bq4_K->dm[1];
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-25 09:09:42 +00:00
|
|
|
const float d8_1 = __low2float(bq8_1[0].ds);
|
|
|
|
const float d8_2 = __low2float(bq8_1[1].ds);
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2));
|
|
|
|
const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4);
|
|
|
|
const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2));
|
|
|
|
const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4);
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
const int * q4 = (const int *)bq4_K->qs + (iqs/2);
|
2023-07-29 21:04:44 +00:00
|
|
|
const int v1 = q4[0];
|
|
|
|
const int v2 = q4[4];
|
|
|
|
|
|
|
|
const int dot1 = __dp4a(ui2, v2 & 0x0f0f0f0f, __dp4a(ui1, v1 & 0x0f0f0f0f, 0));
|
|
|
|
const int dot2 = __dp4a(ui4, (v2 >> 4) & 0x0f0f0f0f, __dp4a(ui3, (v1 >> 4) & 0x0f0f0f0f, 0));
|
|
|
|
const int dot3 = __dp4a(0x01010101, ui2, __dp4a(0x01010101, ui1, 0));
|
|
|
|
const int dot4 = __dp4a(0x01010101, ui4, __dp4a(0x01010101, ui3, 0));
|
|
|
|
|
|
|
|
sumf_d += d8_1 * (dot1 * s[0]) + d8_2 * (dot2 * s[1]);
|
|
|
|
sumf_m += d8_1 * (dot3 * s[2]) + d8_2 * (dot4 * s[3]);
|
|
|
|
|
|
|
|
return dall * sumf_d - dmin * sumf_m;
|
|
|
|
|
|
|
|
#else
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2023-07-29 21:04:44 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
|
|
|
|
|
|
|
|
#endif
|
|
|
|
}
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q4_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
|
2023-11-18 15:11:18 +00:00
|
|
|
(void)x_qh;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
__shared__ int tile_x_ql[mmq_y * (WARP_SIZE) + mmq_y];
|
|
|
|
__shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI4_K) + mmq_y/QI4_K];
|
|
|
|
__shared__ int tile_x_sc[mmq_y * (WARP_SIZE/8) + mmq_y/8];
|
2023-07-29 21:04:44 +00:00
|
|
|
|
|
|
|
*x_ql = tile_x_ql;
|
|
|
|
*x_dm = tile_x_dm;
|
|
|
|
*x_sc = tile_x_sc;
|
|
|
|
}
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q4_K(
|
2023-07-29 21:04:44 +00:00
|
|
|
const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
|
2023-08-02 14:48:10 +00:00
|
|
|
int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
|
2023-11-18 15:11:18 +00:00
|
|
|
(void)x_qh;
|
2023-07-31 11:18:51 +00:00
|
|
|
|
2023-09-16 14:55:43 +00:00
|
|
|
GGML_CUDA_ASSUME(i_offset >= 0);
|
|
|
|
GGML_CUDA_ASSUME(i_offset < nwarps);
|
|
|
|
GGML_CUDA_ASSUME(k >= 0);
|
|
|
|
GGML_CUDA_ASSUME(k < WARP_SIZE);
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-07-31 11:18:51 +00:00
|
|
|
const int kbx = k / QI4_K; // == 0 if QK_K == 256
|
|
|
|
const int kqsx = k % QI4_K; // == k if QK_K == 256
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-11-18 15:11:18 +00:00
|
|
|
const block_q4_K * bx0 = (const block_q4_K *) vx;
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
#pragma unroll
|
2023-08-09 07:42:34 +00:00
|
|
|
for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
|
2023-08-02 14:48:10 +00:00
|
|
|
int i = i0 + i_offset;
|
|
|
|
|
|
|
|
if (need_check) {
|
|
|
|
i = min(i, i_max);
|
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-07-31 11:18:51 +00:00
|
|
|
const block_q4_K * bxi = bx0 + i*blocks_per_row + kbx;
|
|
|
|
|
|
|
|
x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx);
|
|
|
|
}
|
|
|
|
|
|
|
|
const int blocks_per_tile_x_row = WARP_SIZE / QI4_K; // == 1 if QK_K == 256
|
2023-08-05 16:20:44 +00:00
|
|
|
const int kbxd = k % blocks_per_tile_x_row; // == 0 if QK_K == 256
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
#pragma unroll
|
2023-08-09 07:42:34 +00:00
|
|
|
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_K) {
|
|
|
|
int i = (i0 + i_offset * QI4_K + k / blocks_per_tile_x_row) % mmq_y;
|
2023-08-02 14:48:10 +00:00
|
|
|
|
|
|
|
if (need_check) {
|
|
|
|
i = min(i, i_max);
|
|
|
|
}
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
const block_q4_K * bxi = bx0 + i*blocks_per_row + kbxd;
|
|
|
|
|
2023-08-27 12:19:59 +00:00
|
|
|
#if QK_K == 256
|
2023-07-31 11:18:51 +00:00
|
|
|
x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = bxi->dm;
|
2023-08-27 12:19:59 +00:00
|
|
|
#else
|
|
|
|
x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = {bxi->dm[0], bxi->dm[1]};
|
|
|
|
#endif
|
2023-07-31 11:18:51 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
#pragma unroll
|
2023-08-09 07:42:34 +00:00
|
|
|
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) {
|
|
|
|
int i = (i0 + i_offset * 8 + k / (WARP_SIZE/8)) % mmq_y;
|
2023-08-02 14:48:10 +00:00
|
|
|
|
|
|
|
if (need_check) {
|
|
|
|
i = min(i, i_max);
|
|
|
|
}
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
const block_q4_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / (QI4_K/8);
|
|
|
|
|
2023-11-18 15:11:18 +00:00
|
|
|
const int * scales = (const int *) bxi->scales;
|
2023-08-05 16:20:44 +00:00
|
|
|
|
|
|
|
const int ksc = k % (WARP_SIZE/8);
|
|
|
|
|
|
|
|
// scale arrangement after the following two lines: sc0,...,sc3, sc4,...,sc7, m0,...,m3, m4,...,m8
|
|
|
|
int scales8 = (scales[(ksc%2) + (ksc!=0)] >> (4 * (ksc & (ksc/2)))) & 0x0F0F0F0F; // lower 4 bits
|
|
|
|
scales8 |= (scales[ksc/2] >> (2 * (ksc % 2))) & 0x30303030; // upper 2 bits
|
|
|
|
|
|
|
|
x_sc[i * (WARP_SIZE/8) + i / 8 + ksc] = scales8;
|
2023-07-31 11:18:51 +00:00
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static __device__ __forceinline__ float vec_dot_q4_K_q8_1_mul_mat(
|
|
|
|
const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
|
|
|
|
const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
|
2023-11-18 15:11:18 +00:00
|
|
|
(void)x_qh;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
const uint8_t * sc = ((const uint8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/16]) + 2*((k % 16) / 8);
|
2023-07-25 10:48:04 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
const int index_y = j * WARP_SIZE + (QR4_K*k) % WARP_SIZE;
|
2023-08-14 08:41:22 +00:00
|
|
|
return vec_dot_q4_K_q8_1_impl_mmq(&x_ql[i * (WARP_SIZE + 1) + k], &y_qs[index_y], sc, sc+8,
|
|
|
|
x_dm[i * (WARP_SIZE/QI4_K) + i/QI4_K], &y_ds[index_y/QI8_1]);
|
2023-07-14 17:44:08 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static __device__ __forceinline__ float vec_dot_q5_K_q8_1(
|
2023-07-29 21:04:44 +00:00
|
|
|
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
#ifndef GGML_QKK_64
|
2023-07-14 17:44:08 +00:00
|
|
|
const block_q5_K * bq5_K = (const block_q5_K *) vbq;
|
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
int vl[2];
|
|
|
|
int vh[2];
|
|
|
|
int u[2*QR5_K];
|
|
|
|
float d8[QR5_K];
|
2023-07-25 10:48:04 +00:00
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
const int bq8_offset = QR5_K * ((iqs/2) / (QI8_1/2));
|
|
|
|
const int * ql = (const int *)(bq5_K->qs + 16 * bq8_offset + 4 * ((iqs/2)%4));
|
|
|
|
const int * qh = (const int *)(bq5_K->qh + 4 * ((iqs/2)%4));
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
vl[0] = ql[0];
|
|
|
|
vl[1] = ql[4];
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
vh[0] = qh[0] >> bq8_offset;
|
|
|
|
vh[1] = qh[4] >> bq8_offset;
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-07-23 21:19:47 +00:00
|
|
|
const uint16_t * scales = (const uint16_t *)bq5_K->scales;
|
|
|
|
uint16_t aux[2];
|
|
|
|
const int j = bq8_offset/2;
|
|
|
|
if (j < 2) {
|
|
|
|
aux[0] = scales[j+0] & 0x3f3f;
|
|
|
|
aux[1] = scales[j+2] & 0x3f3f;
|
|
|
|
} else {
|
|
|
|
aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2);
|
|
|
|
aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2);
|
|
|
|
}
|
|
|
|
const uint8_t * sc = (const uint8_t *)aux;
|
|
|
|
const uint8_t * m = sc + 2;
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
#pragma unroll
|
2023-07-23 21:19:47 +00:00
|
|
|
for (int i = 0; i < QR5_K; ++i) {
|
2023-07-14 17:44:08 +00:00
|
|
|
const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
|
2023-08-25 09:09:42 +00:00
|
|
|
d8[i] = __low2float(bq8i->ds);
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4);
|
2023-07-29 21:04:44 +00:00
|
|
|
u[2*i+0] = q8[0];
|
|
|
|
u[2*i+1] = q8[4];
|
2023-07-14 17:44:08 +00:00
|
|
|
}
|
|
|
|
|
2023-08-14 08:41:22 +00:00
|
|
|
return vec_dot_q5_K_q8_1_impl_vmmq(vl, vh, u, sc, m, bq5_K->dm, d8);
|
2023-07-25 10:48:04 +00:00
|
|
|
|
|
|
|
#else
|
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
|
|
|
|
const block_q5_K * bq5_K = (const block_q5_K *) vbq;
|
|
|
|
|
2023-07-25 10:48:04 +00:00
|
|
|
const int8_t * s = bq5_K->scales;
|
|
|
|
|
|
|
|
const float d = bq5_K->d;
|
|
|
|
|
2023-08-25 09:09:42 +00:00
|
|
|
const float d8_1 = __low2half(bq8_1[0].ds);
|
|
|
|
const float d8_2 = __low2half(bq8_1[1].ds);
|
2023-07-25 10:48:04 +00:00
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2));
|
|
|
|
const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4);
|
|
|
|
const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2));
|
|
|
|
const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4);
|
2023-07-25 10:48:04 +00:00
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
const int * ql = (const int *)bq5_K->qs + (iqs/2);
|
2023-07-25 10:48:04 +00:00
|
|
|
const int vl1 = ql[0];
|
|
|
|
const int vl2 = ql[4];
|
|
|
|
|
2023-08-02 16:04:04 +00:00
|
|
|
const int step = 4 * (iqs/2); // 0, 4, 8, 12
|
|
|
|
const int im = step/8; // = 0 for iqs = 0, 2, = 1 for iqs = 4, 6
|
2023-07-25 10:48:04 +00:00
|
|
|
const int in = step%8; // 0, 4, 0, 4
|
|
|
|
const int vh = (*((const int *)(bq5_K->qh + in))) >> im;
|
|
|
|
|
|
|
|
const int v1 = (((vh << 4) & 0x10101010) ^ 0x10101010) | ((vl1 >> 0) & 0x0f0f0f0f);
|
|
|
|
const int v2 = (((vh << 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 0) & 0x0f0f0f0f);
|
|
|
|
const int v3 = (((vh >> 0) & 0x10101010) ^ 0x10101010) | ((vl1 >> 4) & 0x0f0f0f0f);
|
|
|
|
const int v4 = (((vh >> 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 4) & 0x0f0f0f0f);
|
|
|
|
|
|
|
|
const float sumf_d = d8_1 * (__dp4a(ui1, v1, 0) * s[0] + __dp4a(ui2, v2, 0) * s[1])
|
|
|
|
+ d8_2 * (__dp4a(ui3, v3, 0) * s[2] + __dp4a(ui4, v4, 0) * s[3]);
|
|
|
|
|
|
|
|
return d * sumf_d;
|
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
#else
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2023-07-29 21:04:44 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
|
|
|
|
|
2023-07-25 10:48:04 +00:00
|
|
|
#endif
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q5_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
|
2023-11-18 15:11:18 +00:00
|
|
|
(void)x_qh;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
__shared__ int tile_x_ql[mmq_y * (2*WARP_SIZE) + mmq_y];
|
|
|
|
__shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI5_K) + mmq_y/QI5_K];
|
|
|
|
__shared__ int tile_x_sc[mmq_y * (WARP_SIZE/8) + mmq_y/8];
|
2023-07-29 21:04:44 +00:00
|
|
|
|
|
|
|
*x_ql = tile_x_ql;
|
|
|
|
*x_dm = tile_x_dm;
|
|
|
|
*x_sc = tile_x_sc;
|
|
|
|
}
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q5_K(
|
2023-07-29 21:04:44 +00:00
|
|
|
const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
|
2023-08-02 14:48:10 +00:00
|
|
|
int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
|
2023-11-18 15:11:18 +00:00
|
|
|
(void)x_qh;
|
2023-07-31 11:18:51 +00:00
|
|
|
|
2023-09-16 14:55:43 +00:00
|
|
|
GGML_CUDA_ASSUME(i_offset >= 0);
|
|
|
|
GGML_CUDA_ASSUME(i_offset < nwarps);
|
|
|
|
GGML_CUDA_ASSUME(k >= 0);
|
|
|
|
GGML_CUDA_ASSUME(k < WARP_SIZE);
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
const int kbx = k / QI5_K; // == 0 if QK_K == 256
|
|
|
|
const int kqsx = k % QI5_K; // == k if QK_K == 256
|
|
|
|
|
2023-11-18 15:11:18 +00:00
|
|
|
const block_q5_K * bx0 = (const block_q5_K *) vx;
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
#pragma unroll
|
2023-08-09 07:42:34 +00:00
|
|
|
for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
|
2023-08-02 14:48:10 +00:00
|
|
|
int i = i0 + i_offset;
|
|
|
|
|
|
|
|
if (need_check) {
|
|
|
|
i = min(i, i_max);
|
|
|
|
}
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
const block_q5_K * bxi = bx0 + i*blocks_per_row + kbx;
|
2023-08-05 16:20:44 +00:00
|
|
|
const int ky = QR5_K*kqsx;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
const int ql = get_int_from_uint8_aligned(bxi->qs, kqsx);
|
|
|
|
const int ql0 = (ql >> 0) & 0x0F0F0F0F;
|
|
|
|
const int ql1 = (ql >> 4) & 0x0F0F0F0F;
|
|
|
|
|
|
|
|
const int qh = get_int_from_uint8_aligned(bxi->qh, kqsx % (QI5_K/4));
|
|
|
|
const int qh0 = ((qh >> (2 * (kqsx / (QI5_K/4)) + 0)) << 4) & 0x10101010;
|
|
|
|
const int qh1 = ((qh >> (2 * (kqsx / (QI5_K/4)) + 1)) << 4) & 0x10101010;
|
|
|
|
|
|
|
|
const int kq0 = ky - ky % (QI5_K/2) + k % (QI5_K/4) + 0;
|
|
|
|
const int kq1 = ky - ky % (QI5_K/2) + k % (QI5_K/4) + (QI5_K/4);
|
|
|
|
|
|
|
|
x_ql[i * (2*WARP_SIZE + 1) + kq0] = ql0 | qh0;
|
|
|
|
x_ql[i * (2*WARP_SIZE + 1) + kq1] = ql1 | qh1;
|
2023-07-31 11:18:51 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
const int blocks_per_tile_x_row = WARP_SIZE / QI5_K; // == 1 if QK_K == 256
|
2023-08-05 16:20:44 +00:00
|
|
|
const int kbxd = k % blocks_per_tile_x_row; // == 0 if QK_K == 256
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
#pragma unroll
|
2023-08-09 07:42:34 +00:00
|
|
|
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_K) {
|
|
|
|
int i = (i0 + i_offset * QI5_K + k / blocks_per_tile_x_row) % mmq_y;
|
2023-08-02 14:48:10 +00:00
|
|
|
|
|
|
|
if (need_check) {
|
|
|
|
i = min(i, i_max);
|
|
|
|
}
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
const block_q5_K * bxi = bx0 + i*blocks_per_row + kbxd;
|
|
|
|
|
2023-08-27 12:19:59 +00:00
|
|
|
#if QK_K == 256
|
2023-07-31 11:18:51 +00:00
|
|
|
x_dm[i * (WARP_SIZE/QI5_K) + i / QI5_K + kbxd] = bxi->dm;
|
2023-08-27 12:19:59 +00:00
|
|
|
#endif
|
2023-07-31 11:18:51 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
#pragma unroll
|
2023-08-09 07:42:34 +00:00
|
|
|
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) {
|
|
|
|
int i = (i0 + i_offset * 8 + k / (WARP_SIZE/8)) % mmq_y;
|
2023-08-02 14:48:10 +00:00
|
|
|
|
|
|
|
if (need_check) {
|
|
|
|
i = min(i, i_max);
|
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-07-31 11:18:51 +00:00
|
|
|
const block_q5_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / (QI5_K/8);
|
|
|
|
|
2023-11-18 15:11:18 +00:00
|
|
|
const int * scales = (const int *) bxi->scales;
|
2023-08-05 16:20:44 +00:00
|
|
|
|
|
|
|
const int ksc = k % (WARP_SIZE/8);
|
|
|
|
|
|
|
|
// scale arrangement after the following two lines: sc0,...,sc3, sc4,...,sc7, m0,...,m3, m4,...,m8
|
|
|
|
int scales8 = (scales[(ksc%2) + (ksc!=0)] >> (4 * (ksc & (ksc/2)))) & 0x0F0F0F0F; // lower 4 bits
|
|
|
|
scales8 |= (scales[ksc/2] >> (2 * (ksc % 2))) & 0x30303030; // upper 2 bits
|
|
|
|
|
|
|
|
x_sc[i * (WARP_SIZE/8) + i / 8 + ksc] = scales8;
|
2023-07-31 11:18:51 +00:00
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static __device__ __forceinline__ float vec_dot_q5_K_q8_1_mul_mat(
|
|
|
|
const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
|
|
|
|
const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
|
2023-11-18 15:11:18 +00:00
|
|
|
(void)x_qh;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
const uint8_t * sc = ((const uint8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/16]) + 2 * ((k % 16) / 8);
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
const int index_x = i * (QR5_K*WARP_SIZE + 1) + QR5_K*k;
|
|
|
|
const int index_y = j * WARP_SIZE + (QR5_K*k) % WARP_SIZE;
|
2023-08-14 08:41:22 +00:00
|
|
|
return vec_dot_q5_K_q8_1_impl_mmq(&x_ql[index_x], &y_qs[index_y], sc, sc+8,
|
|
|
|
x_dm[i * (WARP_SIZE/QI5_K) + i/QI5_K], &y_ds[index_y/QI8_1]);
|
2023-07-14 17:44:08 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static __device__ __forceinline__ float vec_dot_q6_K_q8_1(
|
2023-07-29 21:04:44 +00:00
|
|
|
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
|
2023-07-14 17:44:08 +00:00
|
|
|
|
|
|
|
const block_q6_K * bq6_K = (const block_q6_K *) vbq;
|
|
|
|
|
|
|
|
const int bq8_offset = 2 * QR6_K * (iqs / (QI6_K/2)) + (iqs % (QI6_K/2)) / (QI6_K/4);
|
|
|
|
const int scale_offset = (QI6_K/4) * (iqs / (QI6_K/2)) + (iqs % (QI6_K/2)) / (QI6_K/8);
|
|
|
|
const int vh_shift = 2 * ((iqs % (QI6_K/2)) / (QI6_K/4));
|
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
const int vl = get_int_from_uint8(bq6_K->ql, iqs);
|
|
|
|
const int vh = get_int_from_uint8(bq6_K->qh, (QI6_K/4) * (iqs / (QI6_K/2)) + iqs % (QI6_K/4)) >> vh_shift;
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
const int8_t * scales = bq6_K->scales + scale_offset;
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
int u[QR6_K];
|
|
|
|
float d8[QR6_K];
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
#pragma unroll
|
2023-07-14 17:44:08 +00:00
|
|
|
for (int i = 0; i < QR6_K; ++i) {
|
2023-07-29 21:04:44 +00:00
|
|
|
u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + 2*i].qs, iqs % QI8_1);
|
2023-08-25 09:09:42 +00:00
|
|
|
d8[i] = __low2half(bq8_1[bq8_offset + 2*i].ds);
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
return vec_dot_q6_K_q8_1_impl_mmvq(vl, vh, u, scales, bq6_K->d, d8);
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q6_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
|
2023-11-18 15:11:18 +00:00
|
|
|
(void)x_qh;
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
__shared__ int tile_x_ql[mmq_y * (2*WARP_SIZE) + mmq_y];
|
|
|
|
__shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI6_K) + mmq_y/QI6_K];
|
|
|
|
__shared__ int tile_x_sc[mmq_y * (WARP_SIZE/8) + mmq_y/8];
|
2023-07-14 17:44:08 +00:00
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
*x_ql = tile_x_ql;
|
|
|
|
*x_dm = tile_x_dm;
|
|
|
|
*x_sc = tile_x_sc;
|
|
|
|
}
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q6_K(
|
2023-07-29 21:04:44 +00:00
|
|
|
const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
|
2023-08-02 14:48:10 +00:00
|
|
|
int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
|
2023-11-18 15:11:18 +00:00
|
|
|
(void)x_qh;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-09-16 14:55:43 +00:00
|
|
|
GGML_CUDA_ASSUME(i_offset >= 0);
|
|
|
|
GGML_CUDA_ASSUME(i_offset < nwarps);
|
|
|
|
GGML_CUDA_ASSUME(k >= 0);
|
|
|
|
GGML_CUDA_ASSUME(k < WARP_SIZE);
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-07-31 11:18:51 +00:00
|
|
|
const int kbx = k / QI6_K; // == 0 if QK_K == 256
|
|
|
|
const int kqsx = k % QI6_K; // == k if QK_K == 256
|
|
|
|
|
2023-11-18 15:11:18 +00:00
|
|
|
const block_q6_K * bx0 = (const block_q6_K *) vx;
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
#pragma unroll
|
2023-08-09 07:42:34 +00:00
|
|
|
for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
|
2023-08-02 14:48:10 +00:00
|
|
|
int i = i0 + i_offset;
|
|
|
|
|
|
|
|
if (need_check) {
|
|
|
|
i = min(i, i_max);
|
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-07-31 11:18:51 +00:00
|
|
|
const block_q6_K * bxi = bx0 + i*blocks_per_row + kbx;
|
2023-08-05 16:20:44 +00:00
|
|
|
const int ky = QR6_K*kqsx;
|
|
|
|
|
|
|
|
const int ql = get_int_from_uint8(bxi->ql, kqsx);
|
|
|
|
const int ql0 = (ql >> 0) & 0x0F0F0F0F;
|
|
|
|
const int ql1 = (ql >> 4) & 0x0F0F0F0F;
|
2023-07-31 11:18:51 +00:00
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
const int qh = get_int_from_uint8(bxi->qh, (QI6_K/4) * (kqsx / (QI6_K/2)) + kqsx % (QI6_K/4));
|
|
|
|
const int qh0 = ((qh >> (2 * ((kqsx % (QI6_K/2)) / (QI6_K/4)))) << 4) & 0x30303030;
|
|
|
|
const int qh1 = (qh >> (2 * ((kqsx % (QI6_K/2)) / (QI6_K/4)))) & 0x30303030;
|
|
|
|
|
|
|
|
const int kq0 = ky - ky % QI6_K + k % (QI6_K/2) + 0;
|
|
|
|
const int kq1 = ky - ky % QI6_K + k % (QI6_K/2) + (QI6_K/2);
|
|
|
|
|
|
|
|
x_ql[i * (2*WARP_SIZE + 1) + kq0] = __vsubss4(ql0 | qh0, 0x20202020);
|
|
|
|
x_ql[i * (2*WARP_SIZE + 1) + kq1] = __vsubss4(ql1 | qh1, 0x20202020);
|
2023-07-31 11:18:51 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
const int blocks_per_tile_x_row = WARP_SIZE / QI6_K; // == 1 if QK_K == 256
|
2023-08-05 16:20:44 +00:00
|
|
|
const int kbxd = k % blocks_per_tile_x_row; // == 0 if QK_K == 256
|
|
|
|
float * x_dmf = (float *) x_dm;
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
#pragma unroll
|
2023-08-09 07:42:34 +00:00
|
|
|
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI6_K) {
|
|
|
|
int i = (i0 + i_offset * QI6_K + k / blocks_per_tile_x_row) % mmq_y;
|
2023-08-02 14:48:10 +00:00
|
|
|
|
|
|
|
if (need_check) {
|
|
|
|
i = min(i, i_max);
|
|
|
|
}
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
const block_q6_K * bxi = bx0 + i*blocks_per_row + kbxd;
|
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
x_dmf[i * (WARP_SIZE/QI6_K) + i / QI6_K + kbxd] = bxi->d;
|
2023-07-31 11:18:51 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
#pragma unroll
|
2023-08-09 07:42:34 +00:00
|
|
|
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) {
|
|
|
|
int i = (i0 + i_offset * 8 + k / (WARP_SIZE/8)) % mmq_y;
|
2023-08-02 14:48:10 +00:00
|
|
|
|
|
|
|
if (need_check) {
|
|
|
|
i = min(i, i_max);
|
|
|
|
}
|
2023-07-31 11:18:51 +00:00
|
|
|
|
|
|
|
const block_q6_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / 4;
|
|
|
|
|
|
|
|
x_sc[i * (WARP_SIZE/8) + i / 8 + k % (WARP_SIZE/8)] = get_int_from_int8(bxi->scales, k % (QI6_K/8));
|
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static __device__ __forceinline__ float vec_dot_q6_K_q8_1_mul_mat(
|
|
|
|
const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
|
|
|
|
const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
|
2023-11-18 15:11:18 +00:00
|
|
|
(void)x_qh;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
const float * x_dmf = (const float *) x_dm;
|
|
|
|
const float * y_df = (const float *) y_ds;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-05 16:20:44 +00:00
|
|
|
const int8_t * sc = ((const int8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/8]);
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
const int index_x = i * (QR6_K*WARP_SIZE + 1) + QR6_K*k;
|
|
|
|
const int index_y = j * WARP_SIZE + (QR6_K*k) % WARP_SIZE;
|
2023-08-05 16:20:44 +00:00
|
|
|
return vec_dot_q6_K_q8_1_impl_mmq(&x_ql[index_x], &y_qs[index_y], sc, x_dmf[i * (WARP_SIZE/QI6_K) + i/QI6_K], &y_df[index_y/QI8_1]);
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
2024-01-08 15:02:32 +00:00
|
|
|
static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1(
|
|
|
|
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
|
|
|
|
#if QK_K == 256
|
|
|
|
const block_iq2_xxs * bq2 = (const block_iq2_xxs *) vbq;
|
|
|
|
|
|
|
|
#if QR2_XXS == 8
|
|
|
|
const int ib32 = iqs;
|
|
|
|
const uint16_t * q2 = bq2->qs + 4*ib32;
|
|
|
|
const uint8_t * aux8 = (const uint8_t *)q2;
|
|
|
|
const int8_t * q8 = bq8_1[ib32].qs;
|
|
|
|
uint32_t aux32 = q2[2] | (q2[3] << 16);
|
|
|
|
int sumi = 0;
|
|
|
|
for (int l = 0; l < 4; ++l) {
|
2024-01-11 19:39:39 +00:00
|
|
|
const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
|
2024-01-08 15:02:32 +00:00
|
|
|
const uint8_t signs = ksigns_iq2xs[aux32 & 127];
|
|
|
|
for (int j = 0; j < 8; ++j) {
|
|
|
|
sumi += q8[j] * grid[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
|
|
|
|
}
|
|
|
|
q8 += 8;
|
|
|
|
aux32 >>= 7;
|
|
|
|
}
|
2024-01-24 22:18:15 +00:00
|
|
|
const float d = (float)bq2->d * (0.5f + aux32) * __low2float(bq8_1[ib32].ds) * 0.25f;
|
2024-01-08 15:02:32 +00:00
|
|
|
return d * sumi;
|
|
|
|
#else
|
|
|
|
// iqs is 0...15
|
|
|
|
const int ib32 = iqs/2;
|
|
|
|
const int il = iqs%2;
|
|
|
|
const uint16_t * q2 = bq2->qs + 4*ib32;
|
|
|
|
const uint8_t * aux8 = (const uint8_t *)q2;
|
2024-01-11 19:39:39 +00:00
|
|
|
const uint8_t * grid1 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+0]);
|
|
|
|
const uint8_t * grid2 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+1]);
|
2024-01-08 15:02:32 +00:00
|
|
|
const uint32_t aux32 = q2[2] | (q2[3] << 16);
|
2024-01-24 22:18:15 +00:00
|
|
|
const float d = (float)bq2->d * (0.5f + (aux32 >> 28)) * __low2float(bq8_1[ib32].ds) * 0.25f;
|
2024-01-08 15:02:32 +00:00
|
|
|
const uint8_t signs1 = ksigns_iq2xs[(aux32 >> 14*il) & 127];
|
|
|
|
const uint8_t signs2 = ksigns_iq2xs[(aux32 >> (14*il + 7)) & 127];
|
|
|
|
const int8_t * q8 = bq8_1[ib32].qs + 16*il;
|
|
|
|
int sumi1 = 0, sumi2 = 0;
|
|
|
|
for (int j = 0; j < 8; ++j) {
|
|
|
|
sumi1 += q8[j+0] * grid1[j] * (signs1 & kmask_iq2xs[j] ? -1 : 1);
|
|
|
|
sumi2 += q8[j+8] * grid2[j] * (signs2 & kmask_iq2xs[j] ? -1 : 1);
|
|
|
|
}
|
|
|
|
return d * (sumi1 + sumi2);
|
|
|
|
#endif
|
|
|
|
#else
|
|
|
|
assert(false);
|
|
|
|
return 0.f;
|
|
|
|
#endif
|
|
|
|
}
|
|
|
|
|
2024-01-11 19:39:39 +00:00
|
|
|
static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1(
|
|
|
|
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
|
2024-01-30 13:14:12 +00:00
|
|
|
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
|
2024-01-11 19:39:39 +00:00
|
|
|
#if QK_K == 256
|
|
|
|
const block_iq2_xs * bq2 = (const block_iq2_xs *) vbq;
|
|
|
|
|
|
|
|
const int ib32 = iqs;
|
|
|
|
const uint16_t * q2 = bq2->qs + 4*ib32;
|
|
|
|
const int8_t * q8 = bq8_1[ib32].qs;
|
|
|
|
const uint8_t ls1 = bq2->scales[ib32] & 0xf;
|
|
|
|
const uint8_t ls2 = bq2->scales[ib32] >> 4;
|
|
|
|
int sumi1 = 0;
|
|
|
|
for (int l = 0; l < 2; ++l) {
|
2024-01-30 13:14:12 +00:00
|
|
|
const uint32_t * grid = (const uint32_t *)(iq2xs_grid + (q2[l] & 511));
|
|
|
|
const uint32_t * signs = (const uint32_t *)(ksigns64 + (q2[l] >> 9));
|
|
|
|
const int grid_l = __vsub4(grid[0] ^ signs[0], signs[0]);
|
|
|
|
const int grid_h = __vsub4(grid[1] ^ signs[1], signs[1]);
|
|
|
|
sumi1 = __dp4a(grid_l, *((const int *)q8 + 0), sumi1);
|
|
|
|
sumi1 = __dp4a(grid_h, *((const int *)q8 + 1), sumi1);
|
2024-01-11 19:39:39 +00:00
|
|
|
q8 += 8;
|
|
|
|
}
|
|
|
|
int sumi2 = 0;
|
|
|
|
for (int l = 2; l < 4; ++l) {
|
2024-01-30 13:14:12 +00:00
|
|
|
const uint32_t * grid = (const uint32_t *)(iq2xs_grid + (q2[l] & 511));
|
|
|
|
const uint32_t * signs = (const uint32_t *)(ksigns64 + (q2[l] >> 9));
|
|
|
|
const int grid_l = __vsub4(grid[0] ^ signs[0], signs[0]);
|
|
|
|
const int grid_h = __vsub4(grid[1] ^ signs[1], signs[1]);
|
|
|
|
sumi2 = __dp4a(grid_l, *((const int *)q8 + 0), sumi2);
|
|
|
|
sumi2 = __dp4a(grid_h, *((const int *)q8 + 1), sumi2);
|
2024-01-11 19:39:39 +00:00
|
|
|
q8 += 8;
|
|
|
|
}
|
2024-01-24 22:18:15 +00:00
|
|
|
const float d = (float)bq2->d * __low2float(bq8_1[ib32].ds) * 0.25f;
|
2024-01-11 19:39:39 +00:00
|
|
|
return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2);
|
|
|
|
#else
|
2024-02-19 12:45:41 +00:00
|
|
|
(void) ksigns64;
|
2024-01-11 19:39:39 +00:00
|
|
|
assert(false);
|
|
|
|
return 0.f;
|
|
|
|
#endif
|
2024-01-30 13:14:12 +00:00
|
|
|
#else
|
2024-02-19 12:45:41 +00:00
|
|
|
(void) ksigns64;
|
2024-01-30 13:14:12 +00:00
|
|
|
assert(false);
|
|
|
|
return 0.f;
|
|
|
|
#endif
|
|
|
|
}
|
|
|
|
|
2024-02-26 16:28:38 +00:00
|
|
|
// TODO
|
|
|
|
static __device__ __forceinline__ float vec_dot_iq2_s_q8_1(
|
|
|
|
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
|
|
|
|
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
|
|
|
|
#if QK_K == 256
|
|
|
|
const block_iq2_s * bq2 = (const block_iq2_s *) vbq;
|
|
|
|
|
|
|
|
const int ib32 = iqs;
|
|
|
|
const int8_t * q8 = bq8_1[ib32].qs;
|
|
|
|
const uint8_t * signs = bq2->qs + QK_K/8 + 4*ib32;
|
|
|
|
const uint8_t ls1 = bq2->scales[ib32] & 0xf;
|
|
|
|
const uint8_t ls2 = bq2->scales[ib32] >> 4;
|
|
|
|
int sumi1 = 0;
|
|
|
|
for (int l = 0; l < 2; ++l) {
|
|
|
|
const uint32_t * grid = (const uint32_t *)(iq2s_grid + (bq2->qs[4*ib32+l] | ((bq2->qh[ib32] << (8-2*l)) & 0x300)));
|
|
|
|
const uint32_t signs0 = __vcmpeq4(((signs[l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201);
|
|
|
|
const uint32_t signs1 = __vcmpeq4(((signs[l] >> 4) * 0x01010101) & 0x08040201, 0x08040201);
|
|
|
|
const int grid_l = __vsub4(grid[0] ^ signs0, signs0);
|
|
|
|
const int grid_h = __vsub4(grid[1] ^ signs1, signs1);
|
|
|
|
sumi1 = __dp4a(grid_l, *((const int *)q8 + 0), sumi1);
|
|
|
|
sumi1 = __dp4a(grid_h, *((const int *)q8 + 1), sumi1);
|
|
|
|
q8 += 8;
|
|
|
|
}
|
|
|
|
int sumi2 = 0;
|
|
|
|
for (int l = 2; l < 4; ++l) {
|
|
|
|
const uint32_t * grid = (const uint32_t *)(iq2s_grid + (bq2->qs[4*ib32+l] | ((bq2->qh[ib32] << (8-2*l)) & 0x300)));
|
|
|
|
const uint32_t signs0 = __vcmpeq4(((signs[l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201);
|
|
|
|
const uint32_t signs1 = __vcmpeq4(((signs[l] >> 4) * 0x01010101) & 0x08040201, 0x08040201);
|
|
|
|
const int grid_l = __vsub4(grid[0] ^ signs0, signs0);
|
|
|
|
const int grid_h = __vsub4(grid[1] ^ signs1, signs1);
|
|
|
|
sumi2 = __dp4a(grid_l, *((const int *)q8 + 0), sumi2);
|
|
|
|
sumi2 = __dp4a(grid_h, *((const int *)q8 + 1), sumi2);
|
|
|
|
q8 += 8;
|
|
|
|
}
|
|
|
|
const float d = (float)bq2->d * __low2float(bq8_1[ib32].ds) * 0.25f;
|
|
|
|
return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2);
|
|
|
|
#else
|
|
|
|
(void) ksigns64;
|
|
|
|
assert(false);
|
|
|
|
return 0.f;
|
|
|
|
#endif
|
|
|
|
#else
|
|
|
|
(void) ksigns64;
|
|
|
|
assert(false);
|
|
|
|
return 0.f;
|
|
|
|
#endif
|
|
|
|
}
|
|
|
|
|
2024-01-30 13:14:12 +00:00
|
|
|
static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1(
|
|
|
|
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
|
|
|
|
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
|
|
|
|
#if QK_K == 256
|
|
|
|
const block_iq3_xxs * bq2 = (const block_iq3_xxs *) vbq;
|
|
|
|
|
|
|
|
const int ib32 = iqs;
|
|
|
|
const uint8_t * q3 = bq2->qs + 8*ib32;
|
|
|
|
const uint16_t * gas = (const uint16_t *)(bq2->qs + QK_K/4) + 2*ib32;
|
|
|
|
const int8_t * q8 = bq8_1[ib32].qs;
|
|
|
|
uint32_t aux32 = gas[0] | (gas[1] << 16);
|
|
|
|
int sumi = 0;
|
|
|
|
for (int l = 0; l < 4; ++l) {
|
|
|
|
const uint32_t * grid1 = iq3xxs_grid + q3[2*l+0];
|
|
|
|
const uint32_t * grid2 = iq3xxs_grid + q3[2*l+1];
|
|
|
|
const uint32_t * signs = (const uint32_t *)(ksigns64 + (aux32 & 127));
|
|
|
|
const int grid_l = __vsub4(grid1[0] ^ signs[0], signs[0]);
|
|
|
|
const int grid_h = __vsub4(grid2[0] ^ signs[1], signs[1]);
|
|
|
|
sumi = __dp4a(grid_l, *((int *)q8+0), sumi);
|
|
|
|
sumi = __dp4a(grid_h, *((int *)q8+1), sumi);
|
|
|
|
q8 += 8;
|
|
|
|
aux32 >>= 7;
|
|
|
|
}
|
|
|
|
const float d = (float)bq2->d * (0.5f + aux32) * __low2float(bq8_1[ib32].ds) * 0.5f;
|
|
|
|
return d * sumi;
|
|
|
|
#else
|
|
|
|
assert(false);
|
|
|
|
return 0.f;
|
|
|
|
#endif
|
|
|
|
#else
|
|
|
|
assert(false);
|
|
|
|
return 0.f;
|
|
|
|
#endif
|
2024-01-11 19:39:39 +00:00
|
|
|
}
|
|
|
|
|
2024-02-24 14:23:52 +00:00
|
|
|
// TODO: don't use lookup table for signs
|
|
|
|
static __device__ __forceinline__ float vec_dot_iq3_s_q8_1(
|
|
|
|
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
|
|
|
|
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
|
|
|
|
#if QK_K == 256
|
|
|
|
const block_iq3_s * bq2 = (const block_iq3_s *) vbq;
|
|
|
|
|
|
|
|
const int ib32 = iqs;
|
|
|
|
const uint8_t * qs = bq2->qs + 8*ib32;
|
|
|
|
const int8_t * q8 = bq8_1[ib32].qs;
|
|
|
|
int sumi = 0;
|
|
|
|
for (int l = 0; l < 4; ++l) {
|
|
|
|
const uint32_t * grid1 = iq3xs_grid + (qs[2*l+0] | ((bq2->qh[ib32] << (8 - 2*l)) & 256));
|
|
|
|
const uint32_t * grid2 = iq3xs_grid + (qs[2*l+1] | ((bq2->qh[ib32] << (7 - 2*l)) & 256));
|
|
|
|
uint32_t signs0 = __vcmpeq4(((bq2->signs[4*ib32+l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201);
|
|
|
|
uint32_t signs1 = __vcmpeq4(((bq2->signs[4*ib32+l] >> 4) * 0x01010101) & 0x08040201, 0x08040201);
|
|
|
|
const int grid_l = __vsub4(grid1[0] ^ signs0, signs0);
|
|
|
|
const int grid_h = __vsub4(grid2[0] ^ signs1, signs1);
|
|
|
|
sumi = __dp4a(grid_l, *((int *)q8+0), sumi);
|
|
|
|
sumi = __dp4a(grid_h, *((int *)q8+1), sumi);
|
|
|
|
q8 += 8;
|
|
|
|
}
|
|
|
|
const float d = (float)bq2->d * (0.5f + ((bq2->scales[ib32/2] >> 4*(ib32%2)) & 0xf)) * __low2float(bq8_1[ib32].ds) * 0.5f;
|
|
|
|
return d * sumi;
|
|
|
|
#else
|
|
|
|
assert(false);
|
|
|
|
return 0.f;
|
|
|
|
#endif
|
|
|
|
#else
|
|
|
|
assert(false);
|
|
|
|
return 0.f;
|
|
|
|
#endif
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2024-02-18 16:16:55 +00:00
|
|
|
static __device__ __forceinline__ float vec_dot_iq1_s_q8_1(
|
|
|
|
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
|
|
|
|
#if QK_K == 256
|
|
|
|
const block_iq1_s * bq1 = (const block_iq1_s *) vbq;
|
|
|
|
|
|
|
|
const int ib32 = iqs;
|
|
|
|
int sumi1 = 0, sumi2 = 0, sumi3 = 0, sumi4 = 0;
|
|
|
|
const uint8_t h1 = bq1->scales[2*ib32+0];
|
|
|
|
const uint8_t h2 = bq1->scales[2*ib32+1];
|
|
|
|
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
|
|
|
|
const int * q8 = (const int *)bq8_1[ib32].qs;
|
|
|
|
const int * grid1 = (const int *)(iq1s_grid + (bq1->qs[4*ib32+0] | ((h1 & 0x08) << 5)));
|
|
|
|
const int * grid2 = (const int *)(iq1s_grid + (bq1->qs[4*ib32+1] | ((h1 & 0x80) << 1)));
|
|
|
|
const int * grid3 = (const int *)(iq1s_grid + (bq1->qs[4*ib32+2] | ((h2 & 0x08) << 5)));
|
|
|
|
const int * grid4 = (const int *)(iq1s_grid + (bq1->qs[4*ib32+3] | ((h2 & 0x80) << 1)));
|
|
|
|
for (int j = 0; j < 2; ++j) {
|
|
|
|
sumi1 = __dp4a(q8[j+0], grid1[j], sumi1);
|
|
|
|
sumi2 = __dp4a(q8[j+2], grid2[j], sumi2);
|
|
|
|
sumi3 = __dp4a(q8[j+4], grid3[j], sumi3);
|
|
|
|
sumi4 = __dp4a(q8[j+6], grid4[j], sumi4);
|
|
|
|
}
|
|
|
|
#else
|
|
|
|
const int8_t * q8 = bq8_1[ib32].qs;
|
|
|
|
const int8_t * grid1 = (const int8_t *)(iq1s_grid + (bq1->qs[4*ib32+0] | ((h1 & 0x08) << 5)));
|
|
|
|
const int8_t * grid2 = (const int8_t *)(iq1s_grid + (bq1->qs[4*ib32+1] | ((h1 & 0x80) << 1)));
|
|
|
|
const int8_t * grid3 = (const int8_t *)(iq1s_grid + (bq1->qs[4*ib32+2] | ((h2 & 0x08) << 5)));
|
|
|
|
const int8_t * grid4 = (const int8_t *)(iq1s_grid + (bq1->qs[4*ib32+3] | ((h2 & 0x80) << 1)));
|
|
|
|
for (int j = 0; j < 8; ++j) {
|
|
|
|
sumi1 += q8[j+ 0] * grid1[j];
|
|
|
|
sumi2 += q8[j+ 8] * grid2[j];
|
|
|
|
sumi3 += q8[j+16] * grid3[j];
|
|
|
|
sumi4 += q8[j+24] * grid4[j];
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
const float d = (float)bq1->d * __low2float(bq8_1[ib32].ds);
|
|
|
|
return d * (sumi1 * (2*(h1 & 7) + 1) + sumi2 * (2*((h1 >> 4) & 7) + 1) +
|
|
|
|
sumi3 * (2*(h2 & 7) + 1) + sumi4 * (2*((h2 >> 4) & 7) + 1));
|
|
|
|
#else
|
|
|
|
assert(false);
|
|
|
|
return 0.f;
|
|
|
|
#endif
|
|
|
|
}
|
|
|
|
|
2024-02-21 09:39:52 +00:00
|
|
|
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
|
|
|
|
static __device__ __forceinline__ void get_int_from_table_16(const uint32_t & q4, const uint8_t * values,
|
|
|
|
int & val1, int & val2) {
|
|
|
|
|
|
|
|
uint32_t aux32; const uint8_t * q8 = (const uint8_t *)&aux32;
|
|
|
|
aux32 = q4 & 0x0f0f0f0f;
|
|
|
|
uint16_t v1 = values[q8[0]] | (values[q8[1]] << 8);
|
|
|
|
uint16_t v2 = values[q8[2]] | (values[q8[3]] << 8);
|
|
|
|
val1 = v1 | (v2 << 16);
|
|
|
|
aux32 = (q4 >> 4) & 0x0f0f0f0f;
|
|
|
|
v1 = values[q8[0]] | (values[q8[1]] << 8);
|
|
|
|
v2 = values[q8[2]] | (values[q8[3]] << 8);
|
|
|
|
val2 = v1 | (v2 << 16);
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
|
|
|
static __device__ __forceinline__ float vec_dot_iq4_nl_q8_1(
|
|
|
|
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
|
|
|
|
|
|
|
|
const block_iq4_nl * bq = (const block_iq4_nl *) vbq;
|
|
|
|
|
|
|
|
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
|
|
|
|
const uint16_t * q4 = (const uint16_t *)bq->qs + 2*iqs;
|
|
|
|
const int32_t * q8 = (const int32_t *)bq8_1->qs + iqs;
|
|
|
|
|
|
|
|
const uint8_t * values = (const uint8_t *)kvalues_iq4nl;
|
|
|
|
|
|
|
|
int v1, v2;
|
|
|
|
int sumi1 = 0, sumi2 = 0;
|
|
|
|
for (int l = 0; l < VDR_Q4_0_Q8_1_MMVQ; ++l) {
|
|
|
|
const uint32_t aux = q4[2*l] | (q4[2*l+1] << 16);
|
|
|
|
get_int_from_table_16(aux, values, v1, v2);
|
|
|
|
sumi1 = __dp4a(v1, q8[l+0], sumi1);
|
|
|
|
sumi2 = __dp4a(v2, q8[l+4], sumi2);
|
|
|
|
}
|
|
|
|
|
|
|
|
#else
|
|
|
|
const uint8_t * q4 = bq->qs + 4*iqs;
|
|
|
|
const int8_t * q8 = bq8_1->qs + 4*iqs;
|
|
|
|
|
|
|
|
int sumi1 = 0, sumi2 = 0;
|
|
|
|
for (int l = 0; l < 4*VDR_Q4_0_Q8_1_MMVQ; ++l) {
|
|
|
|
sumi1 += q8[l+ 0] * kvalues_iq4nl[q4[l] & 0xf];
|
|
|
|
sumi2 += q8[l+16] * kvalues_iq4nl[q4[l] >> 4];
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
const float d = (float)bq->d * __low2float(bq8_1->ds);
|
|
|
|
return d * (sumi1 + sumi2);
|
|
|
|
}
|
|
|
|
|
2024-02-27 14:34:24 +00:00
|
|
|
static __device__ __forceinline__ float vec_dot_iq4_xs_q8_1(
|
|
|
|
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
|
|
|
|
|
|
|
|
#if QK_K == 256
|
|
|
|
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
|
|
|
|
|
|
|
|
const block_iq4_xs * bq4 = (const block_iq4_xs *) vbq;
|
|
|
|
const uint8_t * values = (const uint8_t *)kvalues_iq4nl;
|
|
|
|
|
|
|
|
//// iqs is 0...7
|
|
|
|
//const int ib64 = iqs/2;
|
|
|
|
//const int il = iqs%2;
|
|
|
|
//const int32_t * q8_1 = (const int *)bq8_1[2*ib64+0].qs + 2*il;
|
|
|
|
//const int32_t * q8_2 = (const int *)bq8_1[2*ib64+1].qs + 2*il;
|
|
|
|
//const uint32_t * q4_1 = (const uint32_t *)bq4->qs + 8*ib64 + 2*il;
|
|
|
|
//const uint32_t * q4_2 = q4_1 + 4;
|
|
|
|
//const int8_t ls1 = (bq4->scales_l[ib64] & 0xf) | (((bq4->scales_h >> (4*ib64+0)) & 3) << 4);
|
|
|
|
//const int8_t ls2 = (bq4->scales_l[ib64] >> 4) | (((bq4->scales_h >> (4*ib64+2)) & 3) << 4);
|
|
|
|
//const float d1 = (float)bq4->d * (ls1 - 32) * __low2float(bq8_1[2*ib64+0].ds);
|
|
|
|
//const float d2 = (float)bq4->d * (ls2 - 32) * __low2float(bq8_1[2*ib64+1].ds);
|
|
|
|
//int v1, v2;
|
|
|
|
//int sumi1 = 0, sumi2 = 0;
|
|
|
|
//for (int j = 0; j < 2; ++j) {
|
|
|
|
// get_int_from_table_16(q4_1[j], values, v1, v2);
|
|
|
|
// sumi1 = __dp4a(v2, q8_1[j+4], __dp4a(v1, q8_1[j+0], sumi1));
|
|
|
|
// get_int_from_table_16(q4_2[j], values, v1, v2);
|
|
|
|
// sumi2 = __dp4a(v2, q8_2[j+4], __dp4a(v1, q8_2[j+0], sumi2));
|
|
|
|
//}
|
|
|
|
//return d1 * sumi1 + d2 * sumi2;
|
|
|
|
|
|
|
|
// iqs is 0...7
|
|
|
|
const int ib32 = iqs;
|
|
|
|
const int32_t * q8 = (const int *)bq8_1[ib32].qs;
|
|
|
|
const uint32_t * q4 = (const uint32_t *)bq4->qs + 4*ib32;
|
|
|
|
const int8_t ls = ((bq4->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((bq4->scales_h >> 2*ib32) & 3) << 4);
|
|
|
|
const float d = (float)bq4->d * (ls - 32) * __low2float(bq8_1[ib32].ds);
|
|
|
|
int v1, v2;
|
|
|
|
int sumi1 = 0, sumi2 = 0;
|
|
|
|
for (int j = 0; j < 4; ++j) {
|
|
|
|
get_int_from_table_16(q4[j], values, v1, v2);
|
|
|
|
sumi1 = __dp4a(v1, q8[j+0], sumi1);
|
|
|
|
sumi2 = __dp4a(v2, q8[j+4], sumi2);
|
|
|
|
}
|
|
|
|
return d * (sumi1 + sumi2);
|
|
|
|
|
|
|
|
//// iqs is 0...15
|
|
|
|
//const int ib32 = iqs/2;
|
|
|
|
//const int il = iqs%2;
|
|
|
|
//const int32_t * q8 = (const int *)bq8_1[ib32].qs + 2*il;
|
|
|
|
//const uint32_t * q4 = (const uint32_t *)bq4->qs + 4*ib32 + 2*il;
|
|
|
|
//const int8_t ls = ((bq4->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((bq4->scales_h >> 2*ib32) & 3) << 4);
|
|
|
|
//const float d = (float)bq4->d * (ls - 32) * __low2float(bq8_1[ib32].ds);
|
|
|
|
//int v1, v2;
|
|
|
|
//int sumi1 = 0, sumi2 = 0;
|
|
|
|
//for (int j = 0; j < 2; ++j) {
|
|
|
|
// get_int_from_table_16(q4[j], values, v1, v2);
|
|
|
|
// sumi1 = __dp4a(v1, q8[j+0], sumi1);
|
|
|
|
// sumi2 = __dp4a(v2, q8[j+4], sumi2);
|
|
|
|
//}
|
|
|
|
//return d * (sumi1 + sumi2);
|
|
|
|
#else
|
|
|
|
assert(false);
|
|
|
|
return 0.f;
|
|
|
|
#endif
|
|
|
|
#else
|
2024-02-28 08:37:02 +00:00
|
|
|
return vec_dot_iq4_xs_q8_1(vbq, bq8_1, iqs);
|
2024-02-27 14:34:24 +00:00
|
|
|
#endif
|
|
|
|
}
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
template <int qk, int qr, int qi, bool need_sum, typename block_q_t, int mmq_x, int mmq_y, int nwarps,
|
2023-07-29 21:04:44 +00:00
|
|
|
allocate_tiles_cuda_t allocate_tiles, load_tiles_cuda_t load_tiles, int vdr, vec_dot_q_mul_mat_cuda_t vec_dot>
|
2023-08-12 22:24:45 +00:00
|
|
|
static __device__ __forceinline__ void mul_mat_q(
|
2023-07-29 21:04:44 +00:00
|
|
|
const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
|
|
|
|
const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
|
|
|
|
|
|
|
|
const block_q_t * x = (const block_q_t *) vx;
|
|
|
|
const block_q8_1 * y = (const block_q8_1 *) vy;
|
|
|
|
|
|
|
|
const int blocks_per_row_x = ncols_x / qk;
|
|
|
|
const int blocks_per_col_y = nrows_y / QK8_1;
|
|
|
|
const int blocks_per_warp = WARP_SIZE / qi;
|
|
|
|
|
|
|
|
const int & ncols_dst = ncols_y;
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
const int row_dst_0 = blockIdx.x*mmq_y;
|
2023-07-29 21:04:44 +00:00
|
|
|
const int & row_x_0 = row_dst_0;
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
const int col_dst_0 = blockIdx.y*mmq_x;
|
2023-07-29 21:04:44 +00:00
|
|
|
const int & col_y_0 = col_dst_0;
|
|
|
|
|
|
|
|
int * tile_x_ql = nullptr;
|
|
|
|
half2 * tile_x_dm = nullptr;
|
|
|
|
int * tile_x_qh = nullptr;
|
|
|
|
int * tile_x_sc = nullptr;
|
|
|
|
|
|
|
|
allocate_tiles(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc);
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
__shared__ int tile_y_qs[mmq_x * WARP_SIZE];
|
|
|
|
__shared__ half2 tile_y_ds[mmq_x * WARP_SIZE/QI8_1];
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-11-18 15:11:18 +00:00
|
|
|
float sum[mmq_y/WARP_SIZE][mmq_x/nwarps] = {{0.0f}};
|
2023-07-29 21:04:44 +00:00
|
|
|
|
|
|
|
for (int ib0 = 0; ib0 < blocks_per_row_x; ib0 += blocks_per_warp) {
|
|
|
|
|
2023-07-31 11:18:51 +00:00
|
|
|
load_tiles(x + row_x_0*blocks_per_row_x + ib0, tile_x_ql, tile_x_dm, tile_x_qh, tile_x_sc,
|
2023-08-09 07:42:34 +00:00
|
|
|
threadIdx.y, nrows_x-row_x_0-1, threadIdx.x, blocks_per_row_x);
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
#pragma unroll
|
2023-07-29 21:04:44 +00:00
|
|
|
for (int ir = 0; ir < qr; ++ir) {
|
2023-08-09 07:42:34 +00:00
|
|
|
const int kqs = ir*WARP_SIZE + threadIdx.x;
|
2023-07-31 11:18:51 +00:00
|
|
|
const int kbxd = kqs / QI8_1;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
#pragma unroll
|
|
|
|
for (int i = 0; i < mmq_x; i += nwarps) {
|
|
|
|
const int col_y_eff = min(col_y_0 + threadIdx.y + i, ncols_y-1); // to prevent out-of-bounds memory accesses
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-07-31 11:18:51 +00:00
|
|
|
const block_q8_1 * by0 = &y[col_y_eff*blocks_per_col_y + ib0 * (qk/QK8_1) + kbxd];
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
const int index_y = (threadIdx.y + i) * WARP_SIZE + kqs % WARP_SIZE;
|
|
|
|
tile_y_qs[index_y] = get_int_from_int8_aligned(by0->qs, threadIdx.x % QI8_1);
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
#pragma unroll
|
|
|
|
for (int ids0 = 0; ids0 < mmq_x; ids0 += nwarps * QI8_1) {
|
|
|
|
const int ids = (ids0 + threadIdx.y * QI8_1 + threadIdx.x / (WARP_SIZE/QI8_1)) % mmq_x;
|
|
|
|
const int kby = threadIdx.x % (WARP_SIZE/QI8_1);
|
|
|
|
const int col_y_eff = min(col_y_0 + ids, ncols_y-1);
|
|
|
|
|
|
|
|
// if the sum is not needed it's faster to transform the scale to f32 ahead of time
|
|
|
|
const half2 * dsi_src = &y[col_y_eff*blocks_per_col_y + ib0 * (qk/QK8_1) + ir*(WARP_SIZE/QI8_1) + kby].ds;
|
|
|
|
half2 * dsi_dst = &tile_y_ds[ids * (WARP_SIZE/QI8_1) + kby];
|
|
|
|
if (need_sum) {
|
|
|
|
*dsi_dst = *dsi_src;
|
|
|
|
} else {
|
|
|
|
float * dfi_dst = (float *) dsi_dst;
|
2023-08-25 09:09:42 +00:00
|
|
|
*dfi_dst = __low2half(*dsi_src);
|
2023-08-09 07:42:34 +00:00
|
|
|
}
|
2023-08-05 16:20:44 +00:00
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
__syncthreads();
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
// #pragma unroll // unrolling this loop causes too much register pressure
|
|
|
|
for (int k = ir*WARP_SIZE/qr; k < (ir+1)*WARP_SIZE/qr; k += vdr) {
|
2023-07-29 21:04:44 +00:00
|
|
|
#pragma unroll
|
2023-08-09 07:42:34 +00:00
|
|
|
for (int j = 0; j < mmq_x; j += nwarps) {
|
2023-07-29 21:04:44 +00:00
|
|
|
#pragma unroll
|
2023-08-09 07:42:34 +00:00
|
|
|
for (int i = 0; i < mmq_y; i += WARP_SIZE) {
|
|
|
|
sum[i/WARP_SIZE][j/nwarps] += vec_dot(
|
|
|
|
tile_x_ql, tile_x_dm, tile_x_qh, tile_x_sc, tile_y_qs, tile_y_ds,
|
|
|
|
threadIdx.x + i, threadIdx.y + j, k);
|
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
__syncthreads();
|
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
2023-08-12 22:24:45 +00:00
|
|
|
#pragma unroll
|
2023-08-09 07:42:34 +00:00
|
|
|
for (int j = 0; j < mmq_x; j += nwarps) {
|
|
|
|
const int col_dst = col_dst_0 + j + threadIdx.y;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
|
|
|
if (col_dst >= ncols_dst) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2023-08-12 22:24:45 +00:00
|
|
|
#pragma unroll
|
2023-08-09 07:42:34 +00:00
|
|
|
for (int i = 0; i < mmq_y; i += WARP_SIZE) {
|
2023-08-12 22:24:45 +00:00
|
|
|
const int row_dst = row_dst_0 + threadIdx.x + i;
|
|
|
|
|
|
|
|
if (row_dst >= nrows_dst) {
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
dst[col_dst*nrows_dst + row_dst] = sum[i/WARP_SIZE][j/nwarps];
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
}
|
2023-07-05 12:19:42 +00:00
|
|
|
}
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
#define MMQ_X_Q4_0_RDNA2 64
|
|
|
|
#define MMQ_Y_Q4_0_RDNA2 128
|
|
|
|
#define NWARPS_Q4_0_RDNA2 8
|
|
|
|
#define MMQ_X_Q4_0_RDNA1 64
|
|
|
|
#define MMQ_Y_Q4_0_RDNA1 64
|
|
|
|
#define NWARPS_Q4_0_RDNA1 8
|
2023-10-27 14:01:23 +00:00
|
|
|
#if defined(CUDA_USE_TENSOR_CORES)
|
|
|
|
#define MMQ_X_Q4_0_AMPERE 4
|
|
|
|
#define MMQ_Y_Q4_0_AMPERE 32
|
|
|
|
#define NWARPS_Q4_0_AMPERE 4
|
|
|
|
#else
|
2023-08-12 22:24:45 +00:00
|
|
|
#define MMQ_X_Q4_0_AMPERE 64
|
|
|
|
#define MMQ_Y_Q4_0_AMPERE 128
|
|
|
|
#define NWARPS_Q4_0_AMPERE 4
|
2023-10-27 14:01:23 +00:00
|
|
|
#endif
|
2023-08-12 22:24:45 +00:00
|
|
|
#define MMQ_X_Q4_0_PASCAL 64
|
|
|
|
#define MMQ_Y_Q4_0_PASCAL 64
|
|
|
|
#define NWARPS_Q4_0_PASCAL 8
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
template <bool need_check> static __global__ void
|
|
|
|
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
#if defined(RDNA3) || defined(RDNA2)
|
|
|
|
__launch_bounds__(WARP_SIZE*NWARPS_Q4_0_RDNA2, 2)
|
|
|
|
#endif // defined(RDNA3) || defined(RDNA2)
|
|
|
|
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
mul_mat_q4_0(
|
2023-08-12 22:24:45 +00:00
|
|
|
const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
|
|
|
|
const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
#if defined(RDNA3) || defined(RDNA2)
|
|
|
|
const int mmq_x = MMQ_X_Q4_0_RDNA2;
|
|
|
|
const int mmq_y = MMQ_Y_Q4_0_RDNA2;
|
|
|
|
const int nwarps = NWARPS_Q4_0_RDNA2;
|
|
|
|
#else
|
|
|
|
const int mmq_x = MMQ_X_Q4_0_RDNA1;
|
|
|
|
const int mmq_y = MMQ_Y_Q4_0_RDNA1;
|
|
|
|
const int nwarps = NWARPS_Q4_0_RDNA1;
|
|
|
|
#endif // defined(RDNA3) || defined(RDNA2)
|
|
|
|
|
|
|
|
mul_mat_q<QK4_0, QR4_0, QI4_0, true, block_q4_0, mmq_x, mmq_y, nwarps, allocate_tiles_q4_0<mmq_y>,
|
|
|
|
load_tiles_q4_0<mmq_y, nwarps, need_check>, VDR_Q4_0_Q8_1_MMQ, vec_dot_q4_0_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
|
2023-09-30 16:12:57 +00:00
|
|
|
#elif __CUDA_ARCH__ >= CC_VOLTA
|
2023-08-12 22:24:45 +00:00
|
|
|
const int mmq_x = MMQ_X_Q4_0_AMPERE;
|
|
|
|
const int mmq_y = MMQ_Y_Q4_0_AMPERE;
|
|
|
|
const int nwarps = NWARPS_Q4_0_AMPERE;
|
|
|
|
|
|
|
|
mul_mat_q<QK4_0, QR4_0, QI4_0, true, block_q4_0, mmq_x, mmq_y, nwarps, allocate_tiles_q4_0<mmq_y>,
|
|
|
|
load_tiles_q4_0<mmq_y, nwarps, need_check>, VDR_Q4_0_Q8_1_MMQ, vec_dot_q4_0_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
|
|
|
|
#elif __CUDA_ARCH__ >= MIN_CC_DP4A
|
|
|
|
const int mmq_x = MMQ_X_Q4_0_PASCAL;
|
|
|
|
const int mmq_y = MMQ_Y_Q4_0_PASCAL;
|
|
|
|
const int nwarps = NWARPS_Q4_0_PASCAL;
|
|
|
|
|
|
|
|
mul_mat_q<QK4_0, QR4_0, QI4_0, true, block_q4_0, mmq_x, mmq_y, nwarps, allocate_tiles_q4_0<mmq_y>,
|
|
|
|
load_tiles_q4_0<mmq_y, nwarps, need_check>, VDR_Q4_0_Q8_1_MMQ, vec_dot_q4_0_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
#else
|
|
|
|
(void) vec_dot_q4_0_q8_1_mul_mat;
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2023-09-30 16:12:57 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= CC_VOLTA
|
2023-08-12 22:24:45 +00:00
|
|
|
}
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
#define MMQ_X_Q4_1_RDNA2 64
|
|
|
|
#define MMQ_Y_Q4_1_RDNA2 128
|
|
|
|
#define NWARPS_Q4_1_RDNA2 8
|
|
|
|
#define MMQ_X_Q4_1_RDNA1 64
|
|
|
|
#define MMQ_Y_Q4_1_RDNA1 64
|
|
|
|
#define NWARPS_Q4_1_RDNA1 8
|
2023-10-27 14:01:23 +00:00
|
|
|
#if defined(CUDA_USE_TENSOR_CORES)
|
|
|
|
#define MMQ_X_Q4_1_AMPERE 4
|
|
|
|
#define MMQ_Y_Q4_1_AMPERE 32
|
|
|
|
#define NWARPS_Q4_1_AMPERE 4
|
|
|
|
#else
|
2023-08-12 22:24:45 +00:00
|
|
|
#define MMQ_X_Q4_1_AMPERE 64
|
|
|
|
#define MMQ_Y_Q4_1_AMPERE 128
|
|
|
|
#define NWARPS_Q4_1_AMPERE 4
|
2023-10-27 14:01:23 +00:00
|
|
|
#endif
|
2023-08-12 22:24:45 +00:00
|
|
|
#define MMQ_X_Q4_1_PASCAL 64
|
|
|
|
#define MMQ_Y_Q4_1_PASCAL 64
|
|
|
|
#define NWARPS_Q4_1_PASCAL 8
|
|
|
|
|
2023-08-14 08:41:22 +00:00
|
|
|
template <bool need_check> static __global__ void
|
2023-09-13 09:20:24 +00:00
|
|
|
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
#if defined(RDNA3) || defined(RDNA2)
|
|
|
|
__launch_bounds__(WARP_SIZE*NWARPS_Q4_1_RDNA2, 2)
|
|
|
|
#endif // defined(RDNA3) || defined(RDNA2)
|
2023-09-30 16:12:57 +00:00
|
|
|
#elif __CUDA_ARCH__ < CC_VOLTA
|
2023-08-14 08:41:22 +00:00
|
|
|
__launch_bounds__(WARP_SIZE*NWARPS_Q4_1_PASCAL, 2)
|
2023-09-30 16:12:57 +00:00
|
|
|
#endif // __CUDA_ARCH__ < CC_VOLTA
|
2023-08-14 08:41:22 +00:00
|
|
|
mul_mat_q4_1(
|
2023-08-12 22:24:45 +00:00
|
|
|
const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
|
|
|
|
const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
#if defined(RDNA3) || defined(RDNA2)
|
|
|
|
const int mmq_x = MMQ_X_Q4_1_RDNA2;
|
|
|
|
const int mmq_y = MMQ_Y_Q4_1_RDNA2;
|
|
|
|
const int nwarps = NWARPS_Q4_1_RDNA2;
|
|
|
|
#else
|
|
|
|
const int mmq_x = MMQ_X_Q4_1_RDNA1;
|
|
|
|
const int mmq_y = MMQ_Y_Q4_1_RDNA1;
|
|
|
|
const int nwarps = NWARPS_Q4_1_RDNA1;
|
|
|
|
#endif // defined(RDNA3) || defined(RDNA2)
|
|
|
|
|
|
|
|
mul_mat_q<QK4_1, QR4_1, QI4_1, true, block_q4_1, mmq_x, mmq_y, nwarps, allocate_tiles_q4_1<mmq_y>,
|
|
|
|
load_tiles_q4_1<mmq_y, nwarps, need_check>, VDR_Q4_1_Q8_1_MMQ, vec_dot_q4_1_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
|
2023-09-30 16:12:57 +00:00
|
|
|
#elif __CUDA_ARCH__ >= CC_VOLTA
|
2023-08-12 22:24:45 +00:00
|
|
|
const int mmq_x = MMQ_X_Q4_1_AMPERE;
|
|
|
|
const int mmq_y = MMQ_Y_Q4_1_AMPERE;
|
|
|
|
const int nwarps = NWARPS_Q4_1_AMPERE;
|
|
|
|
|
|
|
|
mul_mat_q<QK4_1, QR4_1, QI4_1, true, block_q4_1, mmq_x, mmq_y, nwarps, allocate_tiles_q4_1<mmq_y>,
|
|
|
|
load_tiles_q4_1<mmq_y, nwarps, need_check>, VDR_Q4_1_Q8_1_MMQ, vec_dot_q4_1_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
|
|
|
|
#elif __CUDA_ARCH__ >= MIN_CC_DP4A
|
|
|
|
const int mmq_x = MMQ_X_Q4_1_PASCAL;
|
|
|
|
const int mmq_y = MMQ_Y_Q4_1_PASCAL;
|
|
|
|
const int nwarps = NWARPS_Q4_1_PASCAL;
|
|
|
|
|
|
|
|
mul_mat_q<QK4_1, QR4_1, QI4_1, true, block_q4_1, mmq_x, mmq_y, nwarps, allocate_tiles_q4_1<mmq_y>,
|
|
|
|
load_tiles_q4_1<mmq_y, nwarps, need_check>, VDR_Q4_1_Q8_1_MMQ, vec_dot_q4_1_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
#else
|
|
|
|
(void) vec_dot_q4_1_q8_1_mul_mat;
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2023-09-30 16:12:57 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= CC_VOLTA
|
2023-08-12 22:24:45 +00:00
|
|
|
}
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
#define MMQ_X_Q5_0_RDNA2 64
|
|
|
|
#define MMQ_Y_Q5_0_RDNA2 128
|
|
|
|
#define NWARPS_Q5_0_RDNA2 8
|
|
|
|
#define MMQ_X_Q5_0_RDNA1 64
|
|
|
|
#define MMQ_Y_Q5_0_RDNA1 64
|
|
|
|
#define NWARPS_Q5_0_RDNA1 8
|
2023-10-27 14:01:23 +00:00
|
|
|
#if defined(CUDA_USE_TENSOR_CORES)
|
|
|
|
#define MMQ_X_Q5_0_AMPERE 4
|
|
|
|
#define MMQ_Y_Q5_0_AMPERE 32
|
|
|
|
#define NWARPS_Q5_0_AMPERE 4
|
|
|
|
#else
|
2023-08-12 22:24:45 +00:00
|
|
|
#define MMQ_X_Q5_0_AMPERE 128
|
|
|
|
#define MMQ_Y_Q5_0_AMPERE 64
|
|
|
|
#define NWARPS_Q5_0_AMPERE 4
|
2023-10-27 14:01:23 +00:00
|
|
|
#endif
|
2023-08-12 22:24:45 +00:00
|
|
|
#define MMQ_X_Q5_0_PASCAL 64
|
|
|
|
#define MMQ_Y_Q5_0_PASCAL 64
|
|
|
|
#define NWARPS_Q5_0_PASCAL 8
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
template <bool need_check> static __global__ void
|
|
|
|
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
#if defined(RDNA3) || defined(RDNA2)
|
|
|
|
__launch_bounds__(WARP_SIZE*NWARPS_Q5_0_RDNA2, 2)
|
|
|
|
#endif // defined(RDNA3) || defined(RDNA2)
|
|
|
|
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
mul_mat_q5_0(
|
2023-08-12 22:24:45 +00:00
|
|
|
const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
|
|
|
|
const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
#if defined(RDNA3) || defined(RDNA2)
|
|
|
|
const int mmq_x = MMQ_X_Q5_0_RDNA2;
|
|
|
|
const int mmq_y = MMQ_Y_Q5_0_RDNA2;
|
|
|
|
const int nwarps = NWARPS_Q5_0_RDNA2;
|
|
|
|
#else
|
|
|
|
const int mmq_x = MMQ_X_Q5_0_RDNA1;
|
|
|
|
const int mmq_y = MMQ_Y_Q5_0_RDNA1;
|
|
|
|
const int nwarps = NWARPS_Q5_0_RDNA1;
|
|
|
|
#endif // defined(RDNA3) || defined(RDNA2)
|
|
|
|
|
|
|
|
mul_mat_q<QK5_0, QR5_0, QI5_0, false, block_q5_0, mmq_x, mmq_y, nwarps, allocate_tiles_q5_0<mmq_y>,
|
|
|
|
load_tiles_q5_0<mmq_y, nwarps, need_check>, VDR_Q5_0_Q8_1_MMQ, vec_dot_q5_0_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
|
2023-09-30 16:12:57 +00:00
|
|
|
#elif __CUDA_ARCH__ >= CC_VOLTA
|
2023-08-12 22:24:45 +00:00
|
|
|
const int mmq_x = MMQ_X_Q5_0_AMPERE;
|
|
|
|
const int mmq_y = MMQ_Y_Q5_0_AMPERE;
|
|
|
|
const int nwarps = NWARPS_Q5_0_AMPERE;
|
|
|
|
|
|
|
|
mul_mat_q<QK5_0, QR5_0, QI5_0, false, block_q5_0, mmq_x, mmq_y, nwarps, allocate_tiles_q5_0<mmq_y>,
|
|
|
|
load_tiles_q5_0<mmq_y, nwarps, need_check>, VDR_Q5_0_Q8_1_MMQ, vec_dot_q5_0_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
|
|
|
|
#elif __CUDA_ARCH__ >= MIN_CC_DP4A
|
|
|
|
const int mmq_x = MMQ_X_Q5_0_PASCAL;
|
|
|
|
const int mmq_y = MMQ_Y_Q5_0_PASCAL;
|
|
|
|
const int nwarps = NWARPS_Q5_0_PASCAL;
|
|
|
|
|
|
|
|
mul_mat_q<QK5_0, QR5_0, QI5_0, false, block_q5_0, mmq_x, mmq_y, nwarps, allocate_tiles_q5_0<mmq_y>,
|
|
|
|
load_tiles_q5_0<mmq_y, nwarps, need_check>, VDR_Q5_0_Q8_1_MMQ, vec_dot_q5_0_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
#else
|
|
|
|
(void) vec_dot_q5_0_q8_1_mul_mat;
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2023-09-30 16:12:57 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= CC_VOLTA
|
2023-08-12 22:24:45 +00:00
|
|
|
}
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
#define MMQ_X_Q5_1_RDNA2 64
|
|
|
|
#define MMQ_Y_Q5_1_RDNA2 128
|
|
|
|
#define NWARPS_Q5_1_RDNA2 8
|
|
|
|
#define MMQ_X_Q5_1_RDNA1 64
|
|
|
|
#define MMQ_Y_Q5_1_RDNA1 64
|
|
|
|
#define NWARPS_Q5_1_RDNA1 8
|
2023-10-27 14:01:23 +00:00
|
|
|
#if defined(CUDA_USE_TENSOR_CORES)
|
|
|
|
#define MMQ_X_Q5_1_AMPERE 4
|
|
|
|
#define MMQ_Y_Q5_1_AMPERE 32
|
|
|
|
#define NWARPS_Q5_1_AMPERE 4
|
|
|
|
#else
|
2023-08-12 22:24:45 +00:00
|
|
|
#define MMQ_X_Q5_1_AMPERE 128
|
|
|
|
#define MMQ_Y_Q5_1_AMPERE 64
|
|
|
|
#define NWARPS_Q5_1_AMPERE 4
|
2023-10-27 14:01:23 +00:00
|
|
|
#endif
|
2023-08-12 22:24:45 +00:00
|
|
|
#define MMQ_X_Q5_1_PASCAL 64
|
|
|
|
#define MMQ_Y_Q5_1_PASCAL 64
|
|
|
|
#define NWARPS_Q5_1_PASCAL 8
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
template <bool need_check> static __global__ void
|
|
|
|
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
#if defined(RDNA3) || defined(RDNA2)
|
|
|
|
__launch_bounds__(WARP_SIZE*NWARPS_Q5_1_RDNA2, 2)
|
|
|
|
#endif // defined(RDNA3) || defined(RDNA2)
|
|
|
|
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
mul_mat_q5_1(
|
2023-08-12 22:24:45 +00:00
|
|
|
const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
|
|
|
|
const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
#if defined(RDNA3) || defined(RDNA2)
|
|
|
|
const int mmq_x = MMQ_X_Q5_1_RDNA2;
|
|
|
|
const int mmq_y = MMQ_Y_Q5_1_RDNA2;
|
|
|
|
const int nwarps = NWARPS_Q5_1_RDNA2;
|
|
|
|
#else
|
|
|
|
const int mmq_x = MMQ_X_Q5_1_RDNA1;
|
|
|
|
const int mmq_y = MMQ_Y_Q5_1_RDNA1;
|
|
|
|
const int nwarps = NWARPS_Q5_1_RDNA1;
|
|
|
|
#endif // defined(RDNA3) || defined(RDNA2)
|
|
|
|
|
|
|
|
mul_mat_q<QK5_1, QR5_1, QI5_1, true, block_q5_1, mmq_x, mmq_y, nwarps, allocate_tiles_q5_1<mmq_y>,
|
|
|
|
load_tiles_q5_1<mmq_y, nwarps, need_check>, VDR_Q5_1_Q8_1_MMQ, vec_dot_q5_1_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
|
2023-09-30 16:12:57 +00:00
|
|
|
#elif __CUDA_ARCH__ >= CC_VOLTA
|
2023-08-12 22:24:45 +00:00
|
|
|
const int mmq_x = MMQ_X_Q5_1_AMPERE;
|
|
|
|
const int mmq_y = MMQ_Y_Q5_1_AMPERE;
|
|
|
|
const int nwarps = NWARPS_Q5_1_AMPERE;
|
|
|
|
|
|
|
|
mul_mat_q<QK5_1, QR5_1, QI5_1, true, block_q5_1, mmq_x, mmq_y, nwarps, allocate_tiles_q5_1<mmq_y>,
|
|
|
|
load_tiles_q5_1<mmq_y, nwarps, need_check>, VDR_Q5_1_Q8_1_MMQ, vec_dot_q5_1_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
|
|
|
|
#elif __CUDA_ARCH__ >= MIN_CC_DP4A
|
|
|
|
const int mmq_x = MMQ_X_Q5_1_PASCAL;
|
|
|
|
const int mmq_y = MMQ_Y_Q5_1_PASCAL;
|
|
|
|
const int nwarps = NWARPS_Q5_1_PASCAL;
|
|
|
|
|
|
|
|
mul_mat_q<QK5_1, QR5_1, QI5_1, true, block_q5_1, mmq_x, mmq_y, nwarps, allocate_tiles_q5_1<mmq_y>,
|
|
|
|
load_tiles_q5_1<mmq_y, nwarps, need_check>, VDR_Q5_1_Q8_1_MMQ, vec_dot_q5_1_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
#else
|
|
|
|
(void) vec_dot_q5_1_q8_1_mul_mat;
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2023-09-30 16:12:57 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= CC_VOLTA
|
2023-08-12 22:24:45 +00:00
|
|
|
}
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
#define MMQ_X_Q8_0_RDNA2 64
|
|
|
|
#define MMQ_Y_Q8_0_RDNA2 128
|
|
|
|
#define NWARPS_Q8_0_RDNA2 8
|
|
|
|
#define MMQ_X_Q8_0_RDNA1 64
|
|
|
|
#define MMQ_Y_Q8_0_RDNA1 64
|
|
|
|
#define NWARPS_Q8_0_RDNA1 8
|
2023-10-27 14:01:23 +00:00
|
|
|
#if defined(CUDA_USE_TENSOR_CORES)
|
|
|
|
#define MMQ_X_Q8_0_AMPERE 4
|
|
|
|
#define MMQ_Y_Q8_0_AMPERE 32
|
|
|
|
#define NWARPS_Q8_0_AMPERE 4
|
|
|
|
#else
|
2023-08-12 22:24:45 +00:00
|
|
|
#define MMQ_X_Q8_0_AMPERE 128
|
|
|
|
#define MMQ_Y_Q8_0_AMPERE 64
|
|
|
|
#define NWARPS_Q8_0_AMPERE 4
|
2023-10-27 14:01:23 +00:00
|
|
|
#endif
|
2023-08-12 22:24:45 +00:00
|
|
|
#define MMQ_X_Q8_0_PASCAL 64
|
|
|
|
#define MMQ_Y_Q8_0_PASCAL 64
|
|
|
|
#define NWARPS_Q8_0_PASCAL 8
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
template <bool need_check> static __global__ void
|
|
|
|
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
#if defined(RDNA3) || defined(RDNA2)
|
|
|
|
__launch_bounds__(WARP_SIZE*NWARPS_Q8_0_RDNA2, 2)
|
|
|
|
#endif // defined(RDNA3) || defined(RDNA2)
|
|
|
|
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
mul_mat_q8_0(
|
2023-08-12 22:24:45 +00:00
|
|
|
const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
|
|
|
|
const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
#if defined(RDNA3) || defined(RDNA2)
|
|
|
|
const int mmq_x = MMQ_X_Q8_0_RDNA2;
|
|
|
|
const int mmq_y = MMQ_Y_Q8_0_RDNA2;
|
|
|
|
const int nwarps = NWARPS_Q8_0_RDNA2;
|
|
|
|
#else
|
|
|
|
const int mmq_x = MMQ_X_Q8_0_RDNA1;
|
|
|
|
const int mmq_y = MMQ_Y_Q8_0_RDNA1;
|
|
|
|
const int nwarps = NWARPS_Q8_0_RDNA1;
|
|
|
|
#endif // defined(RDNA3) || defined(RDNA2)
|
|
|
|
|
|
|
|
mul_mat_q<QK8_0, QR8_0, QI8_0, false, block_q8_0, mmq_x, mmq_y, nwarps, allocate_tiles_q8_0<mmq_y>,
|
|
|
|
load_tiles_q8_0<mmq_y, nwarps, need_check>, VDR_Q8_0_Q8_1_MMQ, vec_dot_q8_0_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
|
2023-09-30 16:12:57 +00:00
|
|
|
#elif __CUDA_ARCH__ >= CC_VOLTA
|
2023-08-12 22:24:45 +00:00
|
|
|
const int mmq_x = MMQ_X_Q8_0_AMPERE;
|
|
|
|
const int mmq_y = MMQ_Y_Q8_0_AMPERE;
|
|
|
|
const int nwarps = NWARPS_Q8_0_AMPERE;
|
|
|
|
|
|
|
|
mul_mat_q<QK8_0, QR8_0, QI8_0, false, block_q8_0, mmq_x, mmq_y, nwarps, allocate_tiles_q8_0<mmq_y>,
|
|
|
|
load_tiles_q8_0<mmq_y, nwarps, need_check>, VDR_Q8_0_Q8_1_MMQ, vec_dot_q8_0_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
|
|
|
|
#elif __CUDA_ARCH__ >= MIN_CC_DP4A
|
|
|
|
const int mmq_x = MMQ_X_Q8_0_PASCAL;
|
|
|
|
const int mmq_y = MMQ_Y_Q8_0_PASCAL;
|
|
|
|
const int nwarps = NWARPS_Q8_0_PASCAL;
|
|
|
|
|
|
|
|
mul_mat_q<QK8_0, QR8_0, QI8_0, false, block_q8_0, mmq_x, mmq_y, nwarps, allocate_tiles_q8_0<mmq_y>,
|
|
|
|
load_tiles_q8_0<mmq_y, nwarps, need_check>, VDR_Q8_0_Q8_1_MMQ, vec_dot_q8_0_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
#else
|
|
|
|
(void) vec_dot_q8_0_q8_1_mul_mat;
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2023-09-30 16:12:57 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= CC_VOLTA
|
2023-08-12 22:24:45 +00:00
|
|
|
}
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
#define MMQ_X_Q2_K_RDNA2 64
|
|
|
|
#define MMQ_Y_Q2_K_RDNA2 128
|
|
|
|
#define NWARPS_Q2_K_RDNA2 8
|
|
|
|
#define MMQ_X_Q2_K_RDNA1 128
|
|
|
|
#define MMQ_Y_Q2_K_RDNA1 32
|
|
|
|
#define NWARPS_Q2_K_RDNA1 8
|
2023-10-27 14:01:23 +00:00
|
|
|
#if defined(CUDA_USE_TENSOR_CORES)
|
|
|
|
#define MMQ_X_Q2_K_AMPERE 4
|
|
|
|
#define MMQ_Y_Q2_K_AMPERE 32
|
|
|
|
#define NWARPS_Q2_K_AMPERE 4
|
|
|
|
#else
|
2023-08-12 22:24:45 +00:00
|
|
|
#define MMQ_X_Q2_K_AMPERE 64
|
|
|
|
#define MMQ_Y_Q2_K_AMPERE 128
|
|
|
|
#define NWARPS_Q2_K_AMPERE 4
|
2023-10-27 14:01:23 +00:00
|
|
|
#endif
|
2023-08-12 22:24:45 +00:00
|
|
|
#define MMQ_X_Q2_K_PASCAL 64
|
|
|
|
#define MMQ_Y_Q2_K_PASCAL 64
|
|
|
|
#define NWARPS_Q2_K_PASCAL 8
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
template <bool need_check> static __global__ void
|
|
|
|
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
#if defined(RDNA3) || defined(RDNA2)
|
|
|
|
__launch_bounds__(WARP_SIZE*NWARPS_Q2_K_RDNA2, 2)
|
|
|
|
#endif // defined(RDNA3) || defined(RDNA2)
|
|
|
|
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
mul_mat_q2_K(
|
2023-08-12 22:24:45 +00:00
|
|
|
const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
|
|
|
|
const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
#if defined(RDNA3) || defined(RDNA2)
|
|
|
|
const int mmq_x = MMQ_X_Q2_K_RDNA2;
|
|
|
|
const int mmq_y = MMQ_Y_Q2_K_RDNA2;
|
|
|
|
const int nwarps = NWARPS_Q2_K_RDNA2;
|
|
|
|
#else
|
|
|
|
const int mmq_x = MMQ_X_Q2_K_RDNA1;
|
|
|
|
const int mmq_y = MMQ_Y_Q2_K_RDNA1;
|
|
|
|
const int nwarps = NWARPS_Q2_K_RDNA1;
|
|
|
|
#endif // defined(RDNA3) || defined(RDNA2)
|
|
|
|
|
|
|
|
mul_mat_q<QK_K, QR2_K, QI2_K, false, block_q2_K, mmq_x, mmq_y, nwarps, allocate_tiles_q2_K<mmq_y>,
|
|
|
|
load_tiles_q2_K<mmq_y, nwarps, need_check>, VDR_Q2_K_Q8_1_MMQ, vec_dot_q2_K_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
|
2023-09-30 16:12:57 +00:00
|
|
|
#elif __CUDA_ARCH__ >= CC_VOLTA
|
2023-08-12 22:24:45 +00:00
|
|
|
const int mmq_x = MMQ_X_Q2_K_AMPERE;
|
|
|
|
const int mmq_y = MMQ_Y_Q2_K_AMPERE;
|
|
|
|
const int nwarps = NWARPS_Q2_K_AMPERE;
|
|
|
|
|
|
|
|
mul_mat_q<QK_K, QR2_K, QI2_K, false, block_q2_K, mmq_x, mmq_y, nwarps, allocate_tiles_q2_K<mmq_y>,
|
|
|
|
load_tiles_q2_K<mmq_y, nwarps, need_check>, VDR_Q2_K_Q8_1_MMQ, vec_dot_q2_K_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
|
|
|
|
#elif __CUDA_ARCH__ >= MIN_CC_DP4A
|
|
|
|
const int mmq_x = MMQ_X_Q2_K_PASCAL;
|
|
|
|
const int mmq_y = MMQ_Y_Q2_K_PASCAL;
|
|
|
|
const int nwarps = NWARPS_Q2_K_PASCAL;
|
|
|
|
|
|
|
|
mul_mat_q<QK_K, QR2_K, QI2_K, false, block_q2_K, mmq_x, mmq_y, nwarps, allocate_tiles_q2_K<mmq_y>,
|
|
|
|
load_tiles_q2_K<mmq_y, nwarps, need_check>, VDR_Q2_K_Q8_1_MMQ, vec_dot_q2_K_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
#else
|
|
|
|
(void) vec_dot_q2_K_q8_1_mul_mat;
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2023-09-30 16:12:57 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= CC_VOLTA
|
2023-08-12 22:24:45 +00:00
|
|
|
}
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
#define MMQ_X_Q3_K_RDNA2 128
|
|
|
|
#define MMQ_Y_Q3_K_RDNA2 64
|
|
|
|
#define NWARPS_Q3_K_RDNA2 8
|
|
|
|
#define MMQ_X_Q3_K_RDNA1 32
|
|
|
|
#define MMQ_Y_Q3_K_RDNA1 128
|
|
|
|
#define NWARPS_Q3_K_RDNA1 8
|
2023-10-27 14:01:23 +00:00
|
|
|
#if defined(CUDA_USE_TENSOR_CORES)
|
|
|
|
#define MMQ_X_Q3_K_AMPERE 4
|
|
|
|
#define MMQ_Y_Q3_K_AMPERE 32
|
|
|
|
#define NWARPS_Q3_K_AMPERE 4
|
|
|
|
#else
|
2023-08-12 22:24:45 +00:00
|
|
|
#define MMQ_X_Q3_K_AMPERE 128
|
|
|
|
#define MMQ_Y_Q3_K_AMPERE 128
|
|
|
|
#define NWARPS_Q3_K_AMPERE 4
|
2023-10-27 14:01:23 +00:00
|
|
|
#endif
|
2023-08-12 22:24:45 +00:00
|
|
|
#define MMQ_X_Q3_K_PASCAL 64
|
|
|
|
#define MMQ_Y_Q3_K_PASCAL 64
|
|
|
|
#define NWARPS_Q3_K_PASCAL 8
|
|
|
|
|
2023-08-14 08:41:22 +00:00
|
|
|
template <bool need_check> static __global__ void
|
2023-09-13 09:20:24 +00:00
|
|
|
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
#if defined(RDNA3) || defined(RDNA2)
|
|
|
|
__launch_bounds__(WARP_SIZE*NWARPS_Q3_K_RDNA2, 2)
|
|
|
|
#endif // defined(RDNA3) || defined(RDNA2)
|
2023-09-30 16:12:57 +00:00
|
|
|
#elif __CUDA_ARCH__ < CC_VOLTA
|
2023-08-14 08:41:22 +00:00
|
|
|
__launch_bounds__(WARP_SIZE*NWARPS_Q3_K_PASCAL, 2)
|
2023-09-30 16:12:57 +00:00
|
|
|
#endif // __CUDA_ARCH__ < CC_VOLTA
|
2023-08-14 08:41:22 +00:00
|
|
|
mul_mat_q3_K(
|
2023-08-12 22:24:45 +00:00
|
|
|
const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
|
|
|
|
const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
#if defined(RDNA3) || defined(RDNA2)
|
|
|
|
const int mmq_x = MMQ_X_Q3_K_RDNA2;
|
|
|
|
const int mmq_y = MMQ_Y_Q3_K_RDNA2;
|
|
|
|
const int nwarps = NWARPS_Q3_K_RDNA2;
|
|
|
|
#else
|
|
|
|
const int mmq_x = MMQ_X_Q3_K_RDNA1;
|
|
|
|
const int mmq_y = MMQ_Y_Q3_K_RDNA1;
|
|
|
|
const int nwarps = NWARPS_Q3_K_RDNA1;
|
|
|
|
#endif // defined(RDNA3) || defined(RDNA2)
|
|
|
|
|
|
|
|
mul_mat_q<QK_K, QR3_K, QI3_K, false, block_q3_K, mmq_x, mmq_y, nwarps, allocate_tiles_q3_K<mmq_y>,
|
|
|
|
load_tiles_q3_K<mmq_y, nwarps, need_check>, VDR_Q3_K_Q8_1_MMQ, vec_dot_q3_K_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
|
2023-09-30 16:12:57 +00:00
|
|
|
#elif __CUDA_ARCH__ >= CC_VOLTA
|
2023-08-12 22:24:45 +00:00
|
|
|
const int mmq_x = MMQ_X_Q3_K_AMPERE;
|
|
|
|
const int mmq_y = MMQ_Y_Q3_K_AMPERE;
|
|
|
|
const int nwarps = NWARPS_Q3_K_AMPERE;
|
|
|
|
|
|
|
|
mul_mat_q<QK_K, QR3_K, QI3_K, false, block_q3_K, mmq_x, mmq_y, nwarps, allocate_tiles_q3_K<mmq_y>,
|
|
|
|
load_tiles_q3_K<mmq_y, nwarps, need_check>, VDR_Q3_K_Q8_1_MMQ, vec_dot_q3_K_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
|
|
|
|
#elif __CUDA_ARCH__ >= MIN_CC_DP4A
|
|
|
|
const int mmq_x = MMQ_X_Q3_K_PASCAL;
|
|
|
|
const int mmq_y = MMQ_Y_Q3_K_PASCAL;
|
|
|
|
const int nwarps = NWARPS_Q3_K_PASCAL;
|
|
|
|
|
|
|
|
mul_mat_q<QK_K, QR3_K, QI3_K, false, block_q3_K, mmq_x, mmq_y, nwarps, allocate_tiles_q3_K<mmq_y>,
|
|
|
|
load_tiles_q3_K<mmq_y, nwarps, need_check>, VDR_Q3_K_Q8_1_MMQ, vec_dot_q3_K_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
#else
|
|
|
|
(void) vec_dot_q3_K_q8_1_mul_mat;
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2023-09-30 16:12:57 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= CC_VOLTA
|
2023-08-12 22:24:45 +00:00
|
|
|
}
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
#define MMQ_X_Q4_K_RDNA2 64
|
|
|
|
#define MMQ_Y_Q4_K_RDNA2 128
|
|
|
|
#define NWARPS_Q4_K_RDNA2 8
|
|
|
|
#define MMQ_X_Q4_K_RDNA1 32
|
|
|
|
#define MMQ_Y_Q4_K_RDNA1 64
|
|
|
|
#define NWARPS_Q4_K_RDNA1 8
|
2023-10-27 14:01:23 +00:00
|
|
|
#if defined(CUDA_USE_TENSOR_CORES)
|
|
|
|
#define MMQ_X_Q4_K_AMPERE 4
|
|
|
|
#define MMQ_Y_Q4_K_AMPERE 32
|
|
|
|
#define NWARPS_Q4_K_AMPERE 4
|
|
|
|
#else
|
2023-08-12 22:24:45 +00:00
|
|
|
#define MMQ_X_Q4_K_AMPERE 64
|
|
|
|
#define MMQ_Y_Q4_K_AMPERE 128
|
|
|
|
#define NWARPS_Q4_K_AMPERE 4
|
2023-10-27 14:01:23 +00:00
|
|
|
#endif
|
2023-08-14 08:41:22 +00:00
|
|
|
#define MMQ_X_Q4_K_PASCAL 64
|
2023-08-12 22:24:45 +00:00
|
|
|
#define MMQ_Y_Q4_K_PASCAL 64
|
|
|
|
#define NWARPS_Q4_K_PASCAL 8
|
|
|
|
|
2023-08-14 08:41:22 +00:00
|
|
|
template <bool need_check> static __global__ void
|
2023-09-13 09:20:24 +00:00
|
|
|
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
#if defined(RDNA3) || defined(RDNA2)
|
|
|
|
__launch_bounds__(WARP_SIZE*NWARPS_Q4_K_RDNA2, 2)
|
|
|
|
#endif // defined(RDNA3) || defined(RDNA2)
|
2023-09-30 16:12:57 +00:00
|
|
|
#elif __CUDA_ARCH__ < CC_VOLTA
|
2023-08-14 08:41:22 +00:00
|
|
|
__launch_bounds__(WARP_SIZE*NWARPS_Q4_K_PASCAL, 2)
|
2023-09-30 16:12:57 +00:00
|
|
|
#endif // __CUDA_ARCH__ < CC_VOLTA
|
2023-08-14 08:41:22 +00:00
|
|
|
mul_mat_q4_K(
|
2023-08-12 22:24:45 +00:00
|
|
|
const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
|
|
|
|
const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
#if defined(RDNA3) || defined(RDNA2)
|
|
|
|
const int mmq_x = MMQ_X_Q4_K_RDNA2;
|
|
|
|
const int mmq_y = MMQ_Y_Q4_K_RDNA2;
|
|
|
|
const int nwarps = NWARPS_Q4_K_RDNA2;
|
|
|
|
#else
|
|
|
|
const int mmq_x = MMQ_X_Q4_K_RDNA1;
|
|
|
|
const int mmq_y = MMQ_Y_Q4_K_RDNA1;
|
|
|
|
const int nwarps = NWARPS_Q4_K_RDNA1;
|
|
|
|
#endif // defined(RDNA3) || defined(RDNA2)
|
|
|
|
|
|
|
|
mul_mat_q<QK_K, QR4_K, QI4_K, true, block_q4_K, mmq_x, mmq_y, nwarps, allocate_tiles_q4_K<mmq_y>,
|
|
|
|
load_tiles_q4_K<mmq_y, nwarps, need_check>, VDR_Q4_K_Q8_1_MMQ, vec_dot_q4_K_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
|
2023-09-30 16:12:57 +00:00
|
|
|
#elif __CUDA_ARCH__ >= CC_VOLTA
|
2023-08-12 22:24:45 +00:00
|
|
|
const int mmq_x = MMQ_X_Q4_K_AMPERE;
|
|
|
|
const int mmq_y = MMQ_Y_Q4_K_AMPERE;
|
|
|
|
const int nwarps = NWARPS_Q4_K_AMPERE;
|
|
|
|
|
|
|
|
mul_mat_q<QK_K, QR4_K, QI4_K, true, block_q4_K, mmq_x, mmq_y, nwarps, allocate_tiles_q4_K<mmq_y>,
|
|
|
|
load_tiles_q4_K<mmq_y, nwarps, need_check>, VDR_Q4_K_Q8_1_MMQ, vec_dot_q4_K_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
|
|
|
|
#elif __CUDA_ARCH__ >= MIN_CC_DP4A
|
|
|
|
const int mmq_x = MMQ_X_Q4_K_PASCAL;
|
|
|
|
const int mmq_y = MMQ_Y_Q4_K_PASCAL;
|
|
|
|
const int nwarps = NWARPS_Q4_K_PASCAL;
|
|
|
|
|
|
|
|
mul_mat_q<QK_K, QR4_K, QI4_K, true, block_q4_K, mmq_x, mmq_y, nwarps, allocate_tiles_q4_K<mmq_y>,
|
|
|
|
load_tiles_q4_K<mmq_y, nwarps, need_check>, VDR_Q4_K_Q8_1_MMQ, vec_dot_q4_K_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
#else
|
|
|
|
(void) vec_dot_q4_K_q8_1_mul_mat;
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2023-09-30 16:12:57 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= CC_VOLTA
|
2023-08-12 22:24:45 +00:00
|
|
|
}
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
#define MMQ_X_Q5_K_RDNA2 64
|
|
|
|
#define MMQ_Y_Q5_K_RDNA2 128
|
|
|
|
#define NWARPS_Q5_K_RDNA2 8
|
|
|
|
#define MMQ_X_Q5_K_RDNA1 32
|
|
|
|
#define MMQ_Y_Q5_K_RDNA1 64
|
|
|
|
#define NWARPS_Q5_K_RDNA1 8
|
2023-10-27 14:01:23 +00:00
|
|
|
#if defined(CUDA_USE_TENSOR_CORES)
|
|
|
|
#define MMQ_X_Q5_K_AMPERE 4
|
|
|
|
#define MMQ_Y_Q5_K_AMPERE 32
|
|
|
|
#define NWARPS_Q5_K_AMPERE 4
|
|
|
|
#else
|
2023-08-12 22:24:45 +00:00
|
|
|
#define MMQ_X_Q5_K_AMPERE 64
|
|
|
|
#define MMQ_Y_Q5_K_AMPERE 128
|
|
|
|
#define NWARPS_Q5_K_AMPERE 4
|
2023-10-27 14:01:23 +00:00
|
|
|
#endif
|
2023-08-12 22:24:45 +00:00
|
|
|
#define MMQ_X_Q5_K_PASCAL 64
|
|
|
|
#define MMQ_Y_Q5_K_PASCAL 64
|
|
|
|
#define NWARPS_Q5_K_PASCAL 8
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
template <bool need_check> static __global__ void
|
|
|
|
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
#if defined(RDNA3) || defined(RDNA2)
|
|
|
|
__launch_bounds__(WARP_SIZE*NWARPS_Q5_K_RDNA2, 2)
|
|
|
|
#endif // defined(RDNA3) || defined(RDNA2)
|
|
|
|
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
mul_mat_q5_K(
|
2023-08-12 22:24:45 +00:00
|
|
|
const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
|
|
|
|
const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
#if defined(RDNA3) || defined(RDNA2)
|
|
|
|
const int mmq_x = MMQ_X_Q5_K_RDNA2;
|
|
|
|
const int mmq_y = MMQ_Y_Q5_K_RDNA2;
|
|
|
|
const int nwarps = NWARPS_Q5_K_RDNA2;
|
|
|
|
#else
|
|
|
|
const int mmq_x = MMQ_X_Q5_K_RDNA1;
|
|
|
|
const int mmq_y = MMQ_Y_Q5_K_RDNA1;
|
|
|
|
const int nwarps = NWARPS_Q5_K_RDNA1;
|
|
|
|
#endif // defined(RDNA3) || defined(RDNA2)
|
|
|
|
|
|
|
|
mul_mat_q<QK_K, QR5_K, QI5_K, true, block_q5_K, mmq_x, mmq_y, nwarps, allocate_tiles_q5_K<mmq_y>,
|
|
|
|
load_tiles_q5_K<mmq_y, nwarps, need_check>, VDR_Q5_K_Q8_1_MMQ, vec_dot_q5_K_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
|
2023-09-30 16:12:57 +00:00
|
|
|
#elif __CUDA_ARCH__ >= CC_VOLTA
|
2023-08-12 22:24:45 +00:00
|
|
|
const int mmq_x = MMQ_X_Q5_K_AMPERE;
|
|
|
|
const int mmq_y = MMQ_Y_Q5_K_AMPERE;
|
|
|
|
const int nwarps = NWARPS_Q5_K_AMPERE;
|
|
|
|
|
|
|
|
mul_mat_q<QK_K, QR5_K, QI5_K, true, block_q5_K, mmq_x, mmq_y, nwarps, allocate_tiles_q5_K<mmq_y>,
|
|
|
|
load_tiles_q5_K<mmq_y, nwarps, need_check>, VDR_Q5_K_Q8_1_MMQ, vec_dot_q5_K_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
|
|
|
|
#elif __CUDA_ARCH__ >= MIN_CC_DP4A
|
|
|
|
const int mmq_x = MMQ_X_Q5_K_PASCAL;
|
|
|
|
const int mmq_y = MMQ_Y_Q5_K_PASCAL;
|
|
|
|
const int nwarps = NWARPS_Q5_K_PASCAL;
|
|
|
|
|
|
|
|
mul_mat_q<QK_K, QR5_K, QI5_K, true, block_q5_K, mmq_x, mmq_y, nwarps, allocate_tiles_q5_K<mmq_y>,
|
|
|
|
load_tiles_q5_K<mmq_y, nwarps, need_check>, VDR_Q5_K_Q8_1_MMQ, vec_dot_q5_K_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
#else
|
|
|
|
(void) vec_dot_q5_K_q8_1_mul_mat;
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2023-09-30 16:12:57 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= CC_VOLTA
|
2023-08-12 22:24:45 +00:00
|
|
|
}
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
#define MMQ_X_Q6_K_RDNA2 64
|
|
|
|
#define MMQ_Y_Q6_K_RDNA2 128
|
|
|
|
#define NWARPS_Q6_K_RDNA2 8
|
|
|
|
#define MMQ_X_Q6_K_RDNA1 32
|
|
|
|
#define MMQ_Y_Q6_K_RDNA1 64
|
|
|
|
#define NWARPS_Q6_K_RDNA1 8
|
2023-10-27 14:01:23 +00:00
|
|
|
#if defined(CUDA_USE_TENSOR_CORES)
|
|
|
|
#define MMQ_X_Q6_K_AMPERE 4
|
|
|
|
#define MMQ_Y_Q6_K_AMPERE 32
|
|
|
|
#define NWARPS_Q6_K_AMPERE 4
|
|
|
|
#else
|
2023-08-12 22:24:45 +00:00
|
|
|
#define MMQ_X_Q6_K_AMPERE 64
|
|
|
|
#define MMQ_Y_Q6_K_AMPERE 64
|
|
|
|
#define NWARPS_Q6_K_AMPERE 4
|
2023-10-27 14:01:23 +00:00
|
|
|
#endif
|
2023-08-14 08:41:22 +00:00
|
|
|
#define MMQ_X_Q6_K_PASCAL 64
|
2023-08-12 22:24:45 +00:00
|
|
|
#define MMQ_Y_Q6_K_PASCAL 64
|
|
|
|
#define NWARPS_Q6_K_PASCAL 8
|
|
|
|
|
2023-08-14 08:41:22 +00:00
|
|
|
template <bool need_check> static __global__ void
|
2023-09-13 09:20:24 +00:00
|
|
|
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
#if defined(RDNA3) || defined(RDNA2)
|
|
|
|
__launch_bounds__(WARP_SIZE*NWARPS_Q6_K_RDNA2, 2)
|
|
|
|
#endif // defined(RDNA3) || defined(RDNA2)
|
2023-09-30 16:12:57 +00:00
|
|
|
#elif __CUDA_ARCH__ < CC_VOLTA
|
2023-08-14 08:41:22 +00:00
|
|
|
__launch_bounds__(WARP_SIZE*NWARPS_Q6_K_PASCAL, 2)
|
2023-09-30 16:12:57 +00:00
|
|
|
#endif // __CUDA_ARCH__ < CC_VOLTA
|
2023-08-14 08:41:22 +00:00
|
|
|
mul_mat_q6_K(
|
2023-08-12 22:24:45 +00:00
|
|
|
const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
|
|
|
|
const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
#if defined(RDNA3) || defined(RDNA2)
|
|
|
|
const int mmq_x = MMQ_X_Q6_K_RDNA2;
|
|
|
|
const int mmq_y = MMQ_Y_Q6_K_RDNA2;
|
|
|
|
const int nwarps = NWARPS_Q6_K_RDNA2;
|
|
|
|
#else
|
|
|
|
const int mmq_x = MMQ_X_Q6_K_RDNA1;
|
|
|
|
const int mmq_y = MMQ_Y_Q6_K_RDNA1;
|
|
|
|
const int nwarps = NWARPS_Q6_K_RDNA1;
|
|
|
|
#endif // defined(RDNA3) || defined(RDNA2)
|
|
|
|
|
|
|
|
mul_mat_q<QK_K, QR6_K, QI6_K, false, block_q6_K, mmq_x, mmq_y, nwarps, allocate_tiles_q6_K<mmq_y>,
|
|
|
|
load_tiles_q6_K<mmq_y, nwarps, need_check>, VDR_Q6_K_Q8_1_MMQ, vec_dot_q6_K_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
|
2023-09-30 16:12:57 +00:00
|
|
|
#elif __CUDA_ARCH__ >= CC_VOLTA
|
2023-08-12 22:24:45 +00:00
|
|
|
const int mmq_x = MMQ_X_Q6_K_AMPERE;
|
|
|
|
const int mmq_y = MMQ_Y_Q6_K_AMPERE;
|
|
|
|
const int nwarps = NWARPS_Q6_K_AMPERE;
|
|
|
|
|
|
|
|
mul_mat_q<QK_K, QR6_K, QI6_K, false, block_q6_K, mmq_x, mmq_y, nwarps, allocate_tiles_q6_K<mmq_y>,
|
|
|
|
load_tiles_q6_K<mmq_y, nwarps, need_check>, VDR_Q6_K_Q8_1_MMQ, vec_dot_q6_K_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
|
|
|
|
#elif __CUDA_ARCH__ >= MIN_CC_DP4A
|
|
|
|
const int mmq_x = MMQ_X_Q6_K_PASCAL;
|
|
|
|
const int mmq_y = MMQ_Y_Q6_K_PASCAL;
|
|
|
|
const int nwarps = NWARPS_Q6_K_PASCAL;
|
|
|
|
|
|
|
|
mul_mat_q<QK_K, QR6_K, QI6_K, false, block_q6_K, mmq_x, mmq_y, nwarps, allocate_tiles_q6_K<mmq_y>,
|
|
|
|
load_tiles_q6_K<mmq_y, nwarps, need_check>, VDR_Q6_K_Q8_1_MMQ, vec_dot_q6_K_q8_1_mul_mat>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
#else
|
|
|
|
(void) vec_dot_q6_K_q8_1_mul_mat;
|
2024-01-23 12:31:56 +00:00
|
|
|
NO_DEVICE_CODE;
|
2023-09-30 16:12:57 +00:00
|
|
|
#endif // __CUDA_ARCH__ >= CC_VOLTA
|
2023-08-12 22:24:45 +00:00
|
|
|
}
|
|
|
|
|
2024-02-11 18:08:39 +00:00
|
|
|
template <int ncols_y, int qk, int qi, typename block_q_t, int vdr, vec_dot_q_cuda_t vec_dot_q_cuda>
|
2024-02-08 20:56:40 +00:00
|
|
|
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
|
2024-02-11 18:08:39 +00:00
|
|
|
// tell the compiler to use as many registers as it wants, see nwarps definition below
|
|
|
|
__launch_bounds__((ncols_y <= 4 ? 4 : 2)*WARP_SIZE, 1)
|
2024-02-08 20:56:40 +00:00
|
|
|
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
|
2024-02-06 13:44:06 +00:00
|
|
|
static __global__ void mul_mat_vec_q(
|
|
|
|
const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
|
2024-02-11 18:08:39 +00:00
|
|
|
const int ncols_x, const int nrows_x, const int nrows_y, const int nrows_dst) {
|
2024-02-06 13:44:06 +00:00
|
|
|
|
2024-02-11 18:08:39 +00:00
|
|
|
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3))
|
|
|
|
constexpr int nwarps = 1;
|
|
|
|
constexpr int rows_per_cuda_block = 1;
|
|
|
|
#else
|
|
|
|
constexpr int nwarps = ncols_y <= 4 ? 4 : 2;
|
|
|
|
constexpr int rows_per_cuda_block = ncols_y == 1 ? 1 : 2;
|
|
|
|
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3)
|
2024-02-06 13:44:06 +00:00
|
|
|
|
2024-02-11 18:08:39 +00:00
|
|
|
const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
|
|
|
|
const int row0 = rows_per_cuda_block*blockIdx.x;
|
|
|
|
const int blocks_per_row_x = ncols_x / qk;
|
|
|
|
const int blocks_per_col_y = nrows_y / QK8_1;
|
|
|
|
constexpr int blocks_per_iter = vdr * nwarps*WARP_SIZE / qi;
|
2023-07-05 12:19:42 +00:00
|
|
|
|
|
|
|
// partial sum for each thread
|
2024-02-11 18:08:39 +00:00
|
|
|
float tmp[ncols_y][rows_per_cuda_block] = {0.0f};
|
2023-07-05 12:19:42 +00:00
|
|
|
|
|
|
|
const block_q_t * x = (const block_q_t *) vx;
|
|
|
|
const block_q8_1 * y = (const block_q8_1 *) vy;
|
|
|
|
|
2024-02-11 18:08:39 +00:00
|
|
|
for (int kbx = tid / (qi/vdr); kbx < blocks_per_row_x; kbx += blocks_per_iter) {
|
|
|
|
const int kby = kbx * (qk/QK8_1); // y block index that aligns with kbx
|
2023-07-05 12:19:42 +00:00
|
|
|
|
2024-02-11 18:08:39 +00:00
|
|
|
// x block quant index when casting the quants to int
|
|
|
|
const int kqs = vdr * (tid % (qi/vdr));
|
2023-07-05 12:19:42 +00:00
|
|
|
|
2024-02-06 13:44:06 +00:00
|
|
|
#pragma unroll
|
|
|
|
for (int j = 0; j < ncols_y; ++j) {
|
2024-02-11 18:08:39 +00:00
|
|
|
#pragma unroll
|
|
|
|
for (int i = 0; i < rows_per_cuda_block; ++i) {
|
|
|
|
tmp[j][i] += vec_dot_q_cuda(
|
|
|
|
&x[kbx + (row0 + i)*blocks_per_row_x], &y[j*blocks_per_col_y + kby], kqs);
|
|
|
|
}
|
2024-02-06 13:44:06 +00:00
|
|
|
}
|
2023-07-05 12:19:42 +00:00
|
|
|
}
|
|
|
|
|
2024-02-11 18:08:39 +00:00
|
|
|
__shared__ float tmp_shared[nwarps-1 > 0 ? nwarps-1 : 1][ncols_y][rows_per_cuda_block][WARP_SIZE];
|
2024-02-08 20:56:40 +00:00
|
|
|
if (threadIdx.y > 0) {
|
|
|
|
#pragma unroll
|
|
|
|
for (int j = 0; j < ncols_y; ++j) {
|
2024-02-11 18:08:39 +00:00
|
|
|
#pragma unroll
|
|
|
|
for (int i = 0; i < rows_per_cuda_block; ++i) {
|
|
|
|
tmp_shared[threadIdx.y-1][j][i][threadIdx.x] = tmp[j][i];
|
|
|
|
}
|
2024-02-08 20:56:40 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
__syncthreads();
|
|
|
|
if (threadIdx.y > 0) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2023-07-05 12:19:42 +00:00
|
|
|
// sum up partial sums and write back result
|
|
|
|
#pragma unroll
|
2024-02-06 13:44:06 +00:00
|
|
|
for (int j = 0; j < ncols_y; ++j) {
|
2024-02-08 20:56:40 +00:00
|
|
|
#pragma unroll
|
2024-02-11 18:08:39 +00:00
|
|
|
for (int i = 0; i < rows_per_cuda_block; ++i) {
|
|
|
|
#pragma unroll
|
|
|
|
for (int l = 0; l < nwarps-1; ++l) {
|
|
|
|
tmp[j][i] += tmp_shared[l][j][i][threadIdx.x];
|
|
|
|
}
|
|
|
|
tmp[j][i] = warp_reduce_sum(tmp[j][i]);
|
2024-02-08 20:56:40 +00:00
|
|
|
}
|
2023-07-05 12:19:42 +00:00
|
|
|
|
2024-02-11 18:08:39 +00:00
|
|
|
if (threadIdx.x < rows_per_cuda_block) {
|
|
|
|
dst[j*nrows_dst + row0 + threadIdx.x] = tmp[j][threadIdx.x];
|
2024-02-06 13:44:06 +00:00
|
|
|
}
|
2023-07-05 12:19:42 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-05-25 21:07:29 +00:00
|
|
|
template <int qk, int qr, dequantize_kernel_t dequantize_kernel>
|
2023-07-07 22:25:15 +00:00
|
|
|
static __global__ void dequantize_mul_mat_vec(const void * __restrict__ vx, const dfloat * __restrict__ y, float * __restrict__ dst, const int ncols, const int nrows) {
|
2023-05-25 21:07:29 +00:00
|
|
|
// qk = quantized weights per x block
|
|
|
|
// qr = number of quantized weights per data value in x block
|
2023-11-03 11:13:09 +00:00
|
|
|
const int row = blockIdx.x*blockDim.y + threadIdx.y;
|
2023-06-14 17:47:19 +00:00
|
|
|
|
|
|
|
if (row >= nrows) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2023-05-13 13:38:36 +00:00
|
|
|
const int tid = threadIdx.x;
|
|
|
|
|
2023-05-25 21:07:29 +00:00
|
|
|
const int iter_stride = 2*GGML_CUDA_DMMV_X;
|
|
|
|
const int vals_per_iter = iter_stride / WARP_SIZE; // num quantized vals per thread and i iter
|
2023-05-13 13:38:36 +00:00
|
|
|
const int y_offset = qr == 1 ? 1 : qk/2;
|
|
|
|
|
2023-06-19 08:23:56 +00:00
|
|
|
// partial sum for each thread
|
2023-07-29 21:04:44 +00:00
|
|
|
#ifdef GGML_CUDA_F16
|
2023-06-19 08:23:56 +00:00
|
|
|
half2 tmp = {0.0f, 0.0f}; // two sums for f16 to take advantage of half2 intrinsics
|
|
|
|
#else
|
|
|
|
float tmp = 0.0f;
|
2023-07-29 21:04:44 +00:00
|
|
|
#endif // GGML_CUDA_F16
|
2023-05-13 13:38:36 +00:00
|
|
|
|
2023-05-25 21:07:29 +00:00
|
|
|
for (int i = 0; i < ncols; i += iter_stride) {
|
|
|
|
const int col = i + vals_per_iter*tid;
|
|
|
|
const int ib = (row*ncols + col)/qk; // x block index
|
|
|
|
const int iqs = (col%qk)/qr; // x quant index
|
2023-05-13 13:38:36 +00:00
|
|
|
const int iybs = col - col%qk; // y block start index
|
|
|
|
|
2023-05-25 21:07:29 +00:00
|
|
|
// processing >2 values per i iter is faster for fast GPUs
|
|
|
|
#pragma unroll
|
|
|
|
for (int j = 0; j < vals_per_iter; j += 2) {
|
|
|
|
// process 2 vals per j iter
|
2023-05-13 13:38:36 +00:00
|
|
|
|
2023-05-25 21:07:29 +00:00
|
|
|
// dequantize
|
|
|
|
// for qr = 2 the iqs needs to increase by 1 per j iter because 2 weights per data val
|
2023-06-19 08:23:56 +00:00
|
|
|
dfloat2 v;
|
|
|
|
dequantize_kernel(vx, ib, iqs + j/qr, v);
|
2023-05-25 21:07:29 +00:00
|
|
|
|
|
|
|
// matrix multiplication
|
|
|
|
// for qr = 2 the y index needs to increase by 1 per j iter because of y_offset = qk/2
|
2023-07-29 21:04:44 +00:00
|
|
|
#ifdef GGML_CUDA_F16
|
2023-06-19 08:23:56 +00:00
|
|
|
tmp += __hmul2(v, {
|
|
|
|
y[iybs + iqs + j/qr + 0],
|
|
|
|
y[iybs + iqs + j/qr + y_offset]
|
|
|
|
});
|
|
|
|
#else
|
|
|
|
tmp += v.x * y[iybs + iqs + j/qr + 0];
|
|
|
|
tmp += v.y * y[iybs + iqs + j/qr + y_offset];
|
2023-07-29 21:04:44 +00:00
|
|
|
#endif // GGML_CUDA_F16
|
2023-05-25 21:07:29 +00:00
|
|
|
}
|
2023-05-13 13:38:36 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// sum up partial sums and write back result
|
2024-02-27 13:22:45 +00:00
|
|
|
tmp = warp_reduce_sum(tmp);
|
2023-05-25 21:07:29 +00:00
|
|
|
|
2023-05-13 13:38:36 +00:00
|
|
|
if (tid == 0) {
|
2023-07-29 21:04:44 +00:00
|
|
|
#ifdef GGML_CUDA_F16
|
2023-06-19 08:23:56 +00:00
|
|
|
dst[row] = tmp.x + tmp.y;
|
|
|
|
#else
|
2023-05-25 21:07:29 +00:00
|
|
|
dst[row] = tmp;
|
2023-07-29 21:04:44 +00:00
|
|
|
#endif // GGML_CUDA_F16
|
2023-05-13 13:38:36 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-07-23 12:09:47 +00:00
|
|
|
static __global__ void mul_mat_p021_f16_f32(
|
|
|
|
const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst,
|
|
|
|
const int ncols_x, const int nrows_x, const int nchannels_x, const int nchannels_y) {
|
|
|
|
|
2023-06-28 17:26:26 +00:00
|
|
|
const half * x = (const half *) vx;
|
2023-06-14 17:47:19 +00:00
|
|
|
|
|
|
|
const int row_x = blockDim.y*blockIdx.y + threadIdx.y;
|
|
|
|
const int channel = blockDim.z*blockIdx.z + threadIdx.z;
|
2023-07-23 12:09:47 +00:00
|
|
|
const int channel_x = channel / (nchannels_y / nchannels_x);
|
2023-06-14 17:47:19 +00:00
|
|
|
|
|
|
|
const int nrows_y = ncols_x;
|
|
|
|
const int nrows_dst = nrows_x;
|
|
|
|
const int row_dst = row_x;
|
|
|
|
|
|
|
|
float tmp = 0.0f;
|
|
|
|
|
|
|
|
for (int col_x0 = 0; col_x0 < ncols_x; col_x0 += blockDim.x) {
|
|
|
|
const int col_x = col_x0 + threadIdx.x;
|
|
|
|
|
|
|
|
if (col_x >= ncols_x) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
// x is transposed and permuted
|
2023-07-23 12:09:47 +00:00
|
|
|
const int ix = row_x*nchannels_x*ncols_x + channel_x*ncols_x + col_x;
|
2023-06-14 17:47:19 +00:00
|
|
|
const float xi = __half2float(x[ix]);
|
|
|
|
|
|
|
|
const int row_y = col_x;
|
|
|
|
|
|
|
|
// y is not transposed but permuted
|
|
|
|
const int iy = channel*nrows_y + row_y;
|
|
|
|
|
|
|
|
tmp += xi * y[iy];
|
|
|
|
}
|
|
|
|
|
|
|
|
// dst is not transposed and not permuted
|
|
|
|
const int idst = channel*nrows_dst + row_dst;
|
|
|
|
|
|
|
|
// sum up partial sums and write back result
|
2024-02-27 13:22:45 +00:00
|
|
|
tmp = warp_reduce_sum(tmp);
|
2023-06-14 17:47:19 +00:00
|
|
|
|
|
|
|
if (threadIdx.x == 0) {
|
|
|
|
dst[idst] = tmp;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static __global__ void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
|
2023-07-07 22:25:15 +00:00
|
|
|
const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst, const int ncols_x, const int nrows_x,
|
2023-07-23 12:09:47 +00:00
|
|
|
const int row_stride_x, const int channel_stride_x, const int channel_x_divisor) {
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2023-06-28 17:26:26 +00:00
|
|
|
const half * x = (const half *) vx;
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2023-10-24 13:48:37 +00:00
|
|
|
const int row_x = blockDim.y*blockIdx.y + threadIdx.y;
|
|
|
|
const int channel = blockDim.z*blockIdx.z + threadIdx.z;
|
2023-07-23 12:09:47 +00:00
|
|
|
const int channel_x = channel / channel_x_divisor;
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2023-10-24 13:48:37 +00:00
|
|
|
const int nrows_y = ncols_x;
|
2023-06-14 17:47:19 +00:00
|
|
|
const int nrows_dst = nrows_x;
|
2023-10-24 13:48:37 +00:00
|
|
|
const int row_dst = row_x;
|
2023-06-14 17:47:19 +00:00
|
|
|
|
|
|
|
const int idst = channel*nrows_dst + row_dst;
|
|
|
|
|
|
|
|
float tmp = 0.0f;
|
|
|
|
|
|
|
|
for (int col_x0 = 0; col_x0 < ncols_x; col_x0 += blockDim.x) {
|
|
|
|
const int col_x = col_x0 + threadIdx.x;
|
|
|
|
|
|
|
|
if (col_x >= ncols_x) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
const int row_y = col_x;
|
|
|
|
|
2023-10-24 13:48:37 +00:00
|
|
|
const int ix = channel_x*channel_stride_x + row_x*row_stride_x + col_x;
|
2023-06-14 17:47:19 +00:00
|
|
|
const int iy = channel*nrows_y + row_y;
|
|
|
|
|
2023-10-24 13:48:37 +00:00
|
|
|
const float xi = __half2float(x[ix]);
|
|
|
|
|
2023-06-14 17:47:19 +00:00
|
|
|
tmp += xi * y[iy];
|
|
|
|
}
|
|
|
|
|
|
|
|
// sum up partial sums and write back result
|
2024-02-27 13:22:45 +00:00
|
|
|
tmp = warp_reduce_sum(tmp);
|
2023-06-14 17:47:19 +00:00
|
|
|
|
|
|
|
if (threadIdx.x == 0) {
|
|
|
|
dst[idst] = tmp;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static __device__ void cpy_1_f32_f32(const char * cxi, char * cdsti) {
|
2023-06-28 17:26:26 +00:00
|
|
|
const float * xi = (const float *) cxi;
|
2023-06-14 17:47:19 +00:00
|
|
|
float * dsti = (float *) cdsti;
|
|
|
|
|
|
|
|
*dsti = *xi;
|
|
|
|
}
|
|
|
|
|
|
|
|
static __device__ void cpy_1_f32_f16(const char * cxi, char * cdsti) {
|
2023-06-28 17:26:26 +00:00
|
|
|
const float * xi = (const float *) cxi;
|
2023-06-14 17:47:19 +00:00
|
|
|
half * dsti = (half *) cdsti;
|
|
|
|
|
|
|
|
*dsti = __float2half(*xi);
|
|
|
|
}
|
|
|
|
|
2023-11-13 14:55:52 +00:00
|
|
|
static __device__ void cpy_1_f16_f16(const char * cxi, char * cdsti) {
|
|
|
|
const half * xi = (const half *) cxi;
|
|
|
|
half * dsti = (half *) cdsti;
|
|
|
|
|
|
|
|
*dsti = *xi;
|
|
|
|
}
|
|
|
|
|
2024-01-29 12:37:33 +00:00
|
|
|
static __device__ void cpy_1_f16_f32(const char * cxi, char * cdsti) {
|
|
|
|
const half * xi = (const half *) cxi;
|
|
|
|
float * dsti = (float *) cdsti;
|
|
|
|
|
|
|
|
*dsti = *xi;
|
|
|
|
}
|
|
|
|
|
2023-06-14 17:47:19 +00:00
|
|
|
template <cpy_kernel_t cpy_1>
|
|
|
|
static __global__ void cpy_f32_f16(const char * cx, char * cdst, const int ne,
|
2024-01-29 12:37:33 +00:00
|
|
|
const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
|
|
|
|
const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
|
|
|
|
const int nb12, const int nb13) {
|
2023-06-14 17:47:19 +00:00
|
|
|
const int i = blockDim.x*blockIdx.x + threadIdx.x;
|
|
|
|
|
|
|
|
if (i >= ne) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2024-01-29 12:37:33 +00:00
|
|
|
// determine indices i03/i13, i02/i12, i01/i11, i00/i10 as a function of index i of flattened tensor
|
2023-06-14 17:47:19 +00:00
|
|
|
// then combine those indices with the corresponding byte offsets to get the total offsets
|
2024-01-29 12:37:33 +00:00
|
|
|
const int i03 = i/(ne00 * ne01 * ne02);
|
|
|
|
const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01);
|
|
|
|
const int i01 = (i - i03*ne00*ne01*ne02 - i02*ne01*ne00) / ne00;
|
|
|
|
const int i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00;
|
|
|
|
const int x_offset = i00*nb00 + i01*nb01 + i02*nb02 + i03 * nb03;
|
|
|
|
|
|
|
|
const int i13 = i/(ne10 * ne11 * ne12);
|
|
|
|
const int i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11);
|
|
|
|
const int i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10;
|
|
|
|
const int i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10;
|
|
|
|
const int dst_offset = i10*nb10 + i11*nb11 + i12*nb12 + i13 * nb13;
|
2023-06-14 17:47:19 +00:00
|
|
|
|
|
|
|
cpy_1(cx + x_offset, cdst + dst_offset);
|
|
|
|
}
|
|
|
|
|
2023-12-07 11:03:17 +00:00
|
|
|
static __device__ void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) {
|
|
|
|
const float * xi = (const float *) cxi;
|
|
|
|
block_q8_0 * dsti = (block_q8_0 *) cdsti;
|
|
|
|
|
|
|
|
float amax = 0.0f; // absolute max
|
|
|
|
|
|
|
|
for (int j = 0; j < QK8_0; j++) {
|
|
|
|
const float v = xi[j];
|
|
|
|
amax = fmaxf(amax, fabsf(v));
|
|
|
|
}
|
|
|
|
|
|
|
|
const float d = amax / ((1 << 7) - 1);
|
|
|
|
const float id = d ? 1.0f/d : 0.0f;
|
|
|
|
|
|
|
|
dsti->d = d;
|
|
|
|
|
|
|
|
for (int j = 0; j < QK8_0; ++j) {
|
|
|
|
const float x0 = xi[j]*id;
|
|
|
|
|
|
|
|
dsti->qs[j] = roundf(x0);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static __device__ void cpy_blck_f32_q4_0(const char * cxi, char * cdsti) {
|
|
|
|
const float * xi = (const float *) cxi;
|
|
|
|
block_q4_0 * dsti = (block_q4_0 *) cdsti;
|
|
|
|
|
|
|
|
float amax = 0.0f;
|
|
|
|
float vmax = 0.0f;
|
|
|
|
|
|
|
|
for (int j = 0; j < QK4_0; ++j) {
|
|
|
|
const float v = xi[j];
|
|
|
|
if (amax < fabsf(v)) {
|
|
|
|
amax = fabsf(v);
|
|
|
|
vmax = v;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
const float d = vmax / -8;
|
|
|
|
const float id = d ? 1.0f/d : 0.0f;
|
|
|
|
|
|
|
|
dsti->d = d;
|
|
|
|
|
|
|
|
for (int j = 0; j < QK4_0/2; ++j) {
|
|
|
|
const float x0 = xi[0 + j]*id;
|
|
|
|
const float x1 = xi[QK4_0/2 + j]*id;
|
|
|
|
|
|
|
|
const uint8_t xi0 = min(15, (int8_t)(x0 + 8.5f));
|
|
|
|
const uint8_t xi1 = min(15, (int8_t)(x1 + 8.5f));
|
|
|
|
|
|
|
|
dsti->qs[j] = xi0;
|
|
|
|
dsti->qs[j] |= xi1 << 4;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static __device__ void cpy_blck_f32_q4_1(const char * cxi, char * cdsti) {
|
|
|
|
const float * xi = (const float *) cxi;
|
|
|
|
block_q4_1 * dsti = (block_q4_1 *) cdsti;
|
|
|
|
|
|
|
|
float vmin = FLT_MAX;
|
|
|
|
float vmax = -FLT_MAX;
|
|
|
|
|
|
|
|
for (int j = 0; j < QK4_1; ++j) {
|
|
|
|
const float v = xi[j];
|
|
|
|
|
|
|
|
if (v < vmin) vmin = v;
|
|
|
|
if (v > vmax) vmax = v;
|
|
|
|
}
|
|
|
|
|
|
|
|
const float d = (vmax - vmin) / ((1 << 4) - 1);
|
|
|
|
const float id = d ? 1.0f/d : 0.0f;
|
|
|
|
|
|
|
|
dsti->dm.x = d;
|
|
|
|
dsti->dm.y = vmin;
|
|
|
|
|
|
|
|
for (int j = 0; j < QK4_1/2; ++j) {
|
|
|
|
const float x0 = (xi[0 + j] - vmin)*id;
|
|
|
|
const float x1 = (xi[QK4_1/2 + j] - vmin)*id;
|
|
|
|
|
|
|
|
const uint8_t xi0 = min(15, (int8_t)(x0 + 0.5f));
|
|
|
|
const uint8_t xi1 = min(15, (int8_t)(x1 + 0.5f));
|
|
|
|
|
|
|
|
dsti->qs[j] = xi0;
|
|
|
|
dsti->qs[j] |= xi1 << 4;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
template <cpy_kernel_t cpy_blck, int qk>
|
|
|
|
static __global__ void cpy_f32_q(const char * cx, char * cdst, const int ne,
|
2024-01-29 12:37:33 +00:00
|
|
|
const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
|
|
|
|
const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
|
|
|
|
const int nb12, const int nb13) {
|
2023-12-07 11:03:17 +00:00
|
|
|
const int i = (blockDim.x*blockIdx.x + threadIdx.x)*qk;
|
|
|
|
|
|
|
|
if (i >= ne) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2024-01-29 12:37:33 +00:00
|
|
|
const int i03 = i/(ne00 * ne01 * ne02);
|
|
|
|
const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01);
|
|
|
|
const int i01 = (i - i03*ne00*ne01*ne02 - i02*ne01*ne00) / ne00;
|
|
|
|
const int i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00;
|
|
|
|
const int x_offset = i00*nb00 + i01*nb01 + i02*nb02 + i03 * nb03;
|
2023-12-07 11:03:17 +00:00
|
|
|
|
2024-01-29 12:37:33 +00:00
|
|
|
const int i13 = i/(ne10 * ne11 * ne12);
|
|
|
|
const int i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11);
|
|
|
|
const int i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10;
|
|
|
|
const int i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10;
|
|
|
|
const int dst_offset = (i10/qk)*nb10 + i11*nb11 + i12*nb12 + i13*nb13;
|
2023-12-07 11:03:17 +00:00
|
|
|
|
|
|
|
cpy_blck(cx + x_offset, cdst + dst_offset);
|
|
|
|
}
|
|
|
|
|
2023-11-01 22:04:33 +00:00
|
|
|
static __device__ float rope_yarn_ramp(const float low, const float high, const int i0) {
|
|
|
|
const float y = (i0 / 2 - low) / max(0.001f, high - low);
|
|
|
|
return 1.0f - min(1.0f, max(0.0f, y));
|
|
|
|
}
|
|
|
|
|
|
|
|
struct rope_corr_dims {
|
|
|
|
float v[4];
|
|
|
|
};
|
|
|
|
|
|
|
|
// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
|
|
|
|
// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
|
|
|
|
static __device__ void rope_yarn(
|
|
|
|
float theta_extrap, float freq_scale, rope_corr_dims corr_dims, int64_t i0, float ext_factor, float mscale,
|
|
|
|
float * cos_theta, float * sin_theta
|
|
|
|
) {
|
|
|
|
// Get n-d rotational scaling corrected for extrapolation
|
|
|
|
float theta_interp = freq_scale * theta_extrap;
|
|
|
|
float theta = theta_interp;
|
|
|
|
if (ext_factor != 0.0f) {
|
|
|
|
float ramp_mix = rope_yarn_ramp(corr_dims.v[0], corr_dims.v[1], i0) * ext_factor;
|
|
|
|
theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
|
2023-09-28 16:04:36 +00:00
|
|
|
|
2023-11-01 22:04:33 +00:00
|
|
|
// Get n-d magnitude scaling corrected for interpolation
|
|
|
|
mscale *= 1.0f + 0.1f * logf(1.0f / freq_scale);
|
|
|
|
}
|
|
|
|
*cos_theta = cosf(theta) * mscale;
|
|
|
|
*sin_theta = sinf(theta) * mscale;
|
|
|
|
}
|
|
|
|
|
|
|
|
// rope == RoPE == rotary positional embedding
|
2023-09-28 16:04:36 +00:00
|
|
|
template<typename T, bool has_pos>
|
2023-11-01 22:04:33 +00:00
|
|
|
static __global__ void rope(
|
|
|
|
const T * x, T * dst, int ncols, const int32_t * pos, float freq_scale, int p_delta_rows, float freq_base,
|
|
|
|
float ext_factor, float attn_factor, rope_corr_dims corr_dims
|
|
|
|
) {
|
2023-08-22 13:25:19 +00:00
|
|
|
const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y);
|
2023-06-06 19:33:23 +00:00
|
|
|
|
|
|
|
if (col >= ncols) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2023-08-22 13:25:19 +00:00
|
|
|
const int row = blockDim.x*blockIdx.x + threadIdx.x;
|
2023-06-06 19:33:23 +00:00
|
|
|
const int i = row*ncols + col;
|
2023-09-28 16:04:36 +00:00
|
|
|
const int i2 = row/p_delta_rows;
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2023-09-28 16:04:36 +00:00
|
|
|
const int p = has_pos ? pos[i2] : 0;
|
2023-11-02 05:49:44 +00:00
|
|
|
const float theta_base = p*powf(freq_base, -float(col)/ncols);
|
2023-11-01 22:04:33 +00:00
|
|
|
|
|
|
|
float cos_theta, sin_theta;
|
|
|
|
rope_yarn(theta_base, freq_scale, corr_dims, col, ext_factor, attn_factor, &cos_theta, &sin_theta);
|
2023-06-06 19:33:23 +00:00
|
|
|
|
|
|
|
const float x0 = x[i + 0];
|
|
|
|
const float x1 = x[i + 1];
|
|
|
|
|
|
|
|
dst[i + 0] = x0*cos_theta - x1*sin_theta;
|
|
|
|
dst[i + 1] = x0*sin_theta + x1*cos_theta;
|
|
|
|
}
|
|
|
|
|
2023-09-28 16:04:36 +00:00
|
|
|
template<typename T, bool has_pos>
|
2023-11-01 22:04:33 +00:00
|
|
|
static __global__ void rope_neox(
|
2023-11-24 17:04:31 +00:00
|
|
|
const T * x, T * dst, int ncols, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows,
|
|
|
|
float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, float inv_ndims
|
2023-11-01 22:04:33 +00:00
|
|
|
) {
|
2023-08-25 08:55:59 +00:00
|
|
|
const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y);
|
|
|
|
|
|
|
|
if (col >= ncols) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
const int row = blockDim.x*blockIdx.x + threadIdx.x;
|
2023-11-24 17:04:31 +00:00
|
|
|
const int ib = col / n_dims;
|
|
|
|
const int ic = col % n_dims;
|
|
|
|
|
2023-12-18 17:27:47 +00:00
|
|
|
if (ib > 0) {
|
|
|
|
const int i = row*ncols + ib*n_dims + ic;
|
|
|
|
|
|
|
|
dst[i + 0] = x[i + 0];
|
|
|
|
dst[i + 1] = x[i + 1];
|
|
|
|
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
const int i = row*ncols + ib*n_dims + ic/2;
|
2023-09-28 16:04:36 +00:00
|
|
|
const int i2 = row/p_delta_rows;
|
2023-08-25 08:55:59 +00:00
|
|
|
|
2023-11-24 17:04:31 +00:00
|
|
|
float cur_rot = inv_ndims * ic - ib;
|
2023-11-01 22:04:33 +00:00
|
|
|
|
2023-09-28 16:04:36 +00:00
|
|
|
const int p = has_pos ? pos[i2] : 0;
|
2023-11-24 17:04:31 +00:00
|
|
|
const float theta_base = p*freq_scale*powf(theta_scale, col/2.0f);
|
2023-11-01 22:04:33 +00:00
|
|
|
|
|
|
|
float cos_theta, sin_theta;
|
|
|
|
rope_yarn(theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta);
|
2023-08-25 08:55:59 +00:00
|
|
|
|
|
|
|
const float x0 = x[i + 0];
|
2023-11-24 17:04:31 +00:00
|
|
|
const float x1 = x[i + n_dims/2];
|
2023-08-25 08:55:59 +00:00
|
|
|
|
2023-11-24 17:04:31 +00:00
|
|
|
dst[i + 0] = x0*cos_theta - x1*sin_theta;
|
|
|
|
dst[i + n_dims/2] = x0*sin_theta + x1*cos_theta;
|
2023-08-25 08:55:59 +00:00
|
|
|
}
|
2023-08-23 20:08:04 +00:00
|
|
|
|
2023-11-01 22:04:33 +00:00
|
|
|
static __global__ void rope_glm_f32(
|
|
|
|
const float * x, float * dst, int ncols, const int32_t * pos, float freq_scale, int p_delta_rows, float freq_base,
|
|
|
|
int n_ctx
|
|
|
|
) {
|
2023-07-14 13:36:41 +00:00
|
|
|
const int col = blockDim.x*blockIdx.x + threadIdx.x;
|
|
|
|
const int half_n_dims = ncols/4;
|
|
|
|
|
|
|
|
if (col >= half_n_dims) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
const int row = blockDim.y*blockIdx.y + threadIdx.y;
|
|
|
|
const int i = row*ncols + col;
|
2023-09-28 16:04:36 +00:00
|
|
|
const int i2 = row/p_delta_rows;
|
2023-07-14 13:36:41 +00:00
|
|
|
|
2023-11-01 22:04:33 +00:00
|
|
|
const float col_theta_scale = powf(freq_base, -2.0f*col/ncols);
|
2023-09-28 16:04:36 +00:00
|
|
|
// FIXME: this is likely wrong
|
|
|
|
const int p = pos != nullptr ? pos[i2] : 0;
|
2023-07-14 13:36:41 +00:00
|
|
|
|
2023-09-28 16:04:36 +00:00
|
|
|
const float theta = min(p, n_ctx - 2)*freq_scale*col_theta_scale;
|
2023-07-14 13:36:41 +00:00
|
|
|
const float sin_theta = sinf(theta);
|
|
|
|
const float cos_theta = cosf(theta);
|
|
|
|
|
|
|
|
const float x0 = x[i + 0];
|
|
|
|
const float x1 = x[i + half_n_dims];
|
|
|
|
|
|
|
|
dst[i + 0] = x0*cos_theta - x1*sin_theta;
|
|
|
|
dst[i + half_n_dims] = x0*sin_theta + x1*cos_theta;
|
|
|
|
|
2023-09-28 16:04:36 +00:00
|
|
|
const float block_theta = ((float)max(p - n_ctx - 2, 0))*col_theta_scale;
|
2023-07-14 13:36:41 +00:00
|
|
|
const float sin_block_theta = sinf(block_theta);
|
|
|
|
const float cos_block_theta = cosf(block_theta);
|
|
|
|
|
|
|
|
const float x2 = x[i + half_n_dims * 2];
|
|
|
|
const float x3 = x[i + half_n_dims * 3];
|
|
|
|
|
|
|
|
dst[i + half_n_dims * 2] = x2*cos_block_theta - x3*sin_block_theta;
|
|
|
|
dst[i + half_n_dims * 3] = x2*sin_block_theta + x3*cos_block_theta;
|
|
|
|
}
|
|
|
|
|
2023-08-22 11:22:08 +00:00
|
|
|
static __global__ void alibi_f32(const float * x, float * dst, const int ncols, const int k_rows,
|
|
|
|
const int n_heads_log2_floor, const float m0, const float m1) {
|
|
|
|
const int col = blockDim.x*blockIdx.x + threadIdx.x;
|
|
|
|
|
|
|
|
if (col >= ncols) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
const int row = blockDim.y*blockIdx.y + threadIdx.y;
|
|
|
|
const int i = row*ncols + col;
|
|
|
|
|
|
|
|
const int k = row/k_rows;
|
|
|
|
|
|
|
|
float m_k;
|
|
|
|
if (k < n_heads_log2_floor) {
|
|
|
|
m_k = powf(m0, k + 1);
|
|
|
|
} else {
|
|
|
|
m_k = powf(m1, 2 * (k - n_heads_log2_floor) + 1);
|
|
|
|
}
|
|
|
|
|
|
|
|
dst[i] = col * m_k + x[i];
|
|
|
|
}
|
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
static __global__ void k_sum_rows_f32(const float * x, float * dst, const int ncols) {
|
2024-01-31 13:10:15 +00:00
|
|
|
const int row = blockIdx.x;
|
2023-12-07 20:26:54 +00:00
|
|
|
const int col = threadIdx.x;
|
|
|
|
|
|
|
|
float sum = 0.0f;
|
|
|
|
for (int i = col; i < ncols; i += blockDim.x) {
|
|
|
|
sum += x[row * ncols + i];
|
|
|
|
}
|
|
|
|
|
|
|
|
sum = warp_reduce_sum(sum);
|
|
|
|
|
|
|
|
if (col == 0) {
|
|
|
|
dst[row] = sum;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
template<typename T>
|
|
|
|
static inline __device__ void swap(T & a, T & b) {
|
|
|
|
T tmp = a;
|
|
|
|
a = b;
|
|
|
|
b = tmp;
|
|
|
|
}
|
|
|
|
|
|
|
|
template<ggml_sort_order order>
|
|
|
|
static __global__ void k_argsort_f32_i32(const float * x, int * dst, const int ncols) {
|
|
|
|
// bitonic sort
|
|
|
|
int col = threadIdx.x;
|
|
|
|
int row = blockIdx.y;
|
|
|
|
|
|
|
|
if (col >= ncols) return;
|
|
|
|
|
|
|
|
const float * x_row = x + row * ncols;
|
|
|
|
int * dst_row = dst + row * ncols;
|
|
|
|
|
|
|
|
// initialize indices
|
|
|
|
if (col < ncols) {
|
|
|
|
dst_row[col] = col;
|
|
|
|
}
|
|
|
|
__syncthreads();
|
|
|
|
|
|
|
|
for (int k = 2; k <= ncols; k *= 2) {
|
|
|
|
for (int j = k / 2; j > 0; j /= 2) {
|
|
|
|
int ixj = col ^ j;
|
|
|
|
if (ixj > col) {
|
|
|
|
if ((col & k) == 0) {
|
2024-02-25 10:09:09 +00:00
|
|
|
if (order == GGML_SORT_ORDER_ASC ? x_row[dst_row[col]] > x_row[dst_row[ixj]] : x_row[dst_row[col]] < x_row[dst_row[ixj]]) {
|
2023-12-07 20:26:54 +00:00
|
|
|
swap(dst_row[col], dst_row[ixj]);
|
|
|
|
}
|
|
|
|
} else {
|
2024-02-25 10:09:09 +00:00
|
|
|
if (order == GGML_SORT_ORDER_ASC ? x_row[dst_row[col]] < x_row[dst_row[ixj]] : x_row[dst_row[col]] > x_row[dst_row[ixj]]) {
|
2023-12-07 20:26:54 +00:00
|
|
|
swap(dst_row[col], dst_row[ixj]);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
__syncthreads();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-06-14 17:47:19 +00:00
|
|
|
static __global__ void diag_mask_inf_f32(const float * x, float * dst, const int ncols, const int rows_per_channel, const int n_past) {
|
2023-08-22 13:25:19 +00:00
|
|
|
const int col = blockDim.y*blockIdx.y + threadIdx.y;
|
|
|
|
const int row = blockDim.x*blockIdx.x + threadIdx.x;
|
2023-06-14 17:47:19 +00:00
|
|
|
|
|
|
|
if (col >= ncols) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
const int i = row*ncols + col;
|
2023-12-07 20:26:54 +00:00
|
|
|
//dst[i] = col > (n_past + row % rows_per_channel) ? -INFINITY : x[i];
|
|
|
|
//dst[i] = x[i] - (col > n_past + row % rows_per_channel) * INT_MAX; // equivalent within rounding error but slightly faster on GPU
|
|
|
|
dst[i] = x[i] - (col > n_past + row % rows_per_channel) * FLT_MAX;
|
2023-06-14 17:47:19 +00:00
|
|
|
}
|
|
|
|
|
2024-02-17 21:04:16 +00:00
|
|
|
template <bool vals_smem, int ncols_template, int block_size_template>
|
|
|
|
static __global__ void soft_max_f32(const float * x, const float * mask, const float * pos, float * dst, const int ncols_par, const int nrows_y, const float scale, const float max_bias, const float m0, const float m1, uint32_t n_head_log2) {
|
|
|
|
const int ncols = ncols_template == 0 ? ncols_par : ncols_template;
|
2024-01-09 07:58:55 +00:00
|
|
|
|
|
|
|
const int tid = threadIdx.x;
|
|
|
|
const int rowx = blockIdx.x;
|
2024-02-17 21:04:16 +00:00
|
|
|
const int rowy = rowx % nrows_y; // broadcast the mask in the row dimension
|
2024-01-09 07:58:55 +00:00
|
|
|
|
|
|
|
const int block_size = block_size_template == 0 ? blockDim.x : block_size_template;
|
|
|
|
|
|
|
|
const int warp_id = threadIdx.x / WARP_SIZE;
|
|
|
|
const int lane_id = threadIdx.x % WARP_SIZE;
|
|
|
|
|
2024-02-17 21:04:16 +00:00
|
|
|
float slope = 0.0f;
|
2024-01-09 07:58:55 +00:00
|
|
|
|
2024-02-17 21:04:16 +00:00
|
|
|
// ALiBi
|
|
|
|
if (max_bias > 0.0f) {
|
|
|
|
const int h = rowx/nrows_y; // head index
|
2024-01-09 07:58:55 +00:00
|
|
|
|
2024-02-17 21:04:16 +00:00
|
|
|
const float base = h < n_head_log2 ? m0 : m1;
|
|
|
|
const int exp = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
|
2024-01-09 07:58:55 +00:00
|
|
|
|
2024-02-17 21:04:16 +00:00
|
|
|
slope = powf(base, exp);
|
2024-01-09 07:58:55 +00:00
|
|
|
}
|
2023-12-01 08:51:24 +00:00
|
|
|
|
2024-01-09 07:58:55 +00:00
|
|
|
extern __shared__ float data_soft_max_f32[];
|
|
|
|
float * buf_iw = data_soft_max_f32; // shared memory buffer for inter-warp communication
|
|
|
|
// shared memory buffer to cache values between iterations:
|
|
|
|
float * vals = vals_smem ? buf_iw + WARP_SIZE : dst + rowx*ncols;
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2023-08-22 18:27:06 +00:00
|
|
|
float max_val = -INFINITY;
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2024-01-09 07:58:55 +00:00
|
|
|
#pragma unroll
|
|
|
|
for (int col0 = 0; col0 < ncols; col0 += block_size) {
|
|
|
|
const int col = col0 + tid;
|
|
|
|
|
|
|
|
if (ncols_template == 0 && col >= ncols) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
2023-12-01 08:51:24 +00:00
|
|
|
const int ix = rowx*ncols + col;
|
|
|
|
const int iy = rowy*ncols + col;
|
2024-01-09 07:58:55 +00:00
|
|
|
|
2024-02-19 08:04:45 +00:00
|
|
|
const float val = x[ix]*scale + (mask ? mask[iy] : 0.0f) + (pos ? slope*pos[col] : 0.0f);
|
2024-02-17 21:04:16 +00:00
|
|
|
|
2024-01-09 07:58:55 +00:00
|
|
|
vals[col] = val;
|
|
|
|
max_val = max(max_val, val);
|
2023-08-22 18:27:06 +00:00
|
|
|
}
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2023-08-22 18:27:06 +00:00
|
|
|
// find the max value in the block
|
2023-12-01 08:51:24 +00:00
|
|
|
max_val = warp_reduce_max(max_val);
|
|
|
|
if (block_size > WARP_SIZE) {
|
|
|
|
if (warp_id == 0) {
|
2024-01-09 07:58:55 +00:00
|
|
|
buf_iw[lane_id] = -INFINITY;
|
2023-12-01 08:51:24 +00:00
|
|
|
}
|
|
|
|
__syncthreads();
|
|
|
|
|
|
|
|
if (lane_id == 0) {
|
2024-01-09 07:58:55 +00:00
|
|
|
buf_iw[warp_id] = max_val;
|
2023-12-01 08:51:24 +00:00
|
|
|
}
|
|
|
|
__syncthreads();
|
|
|
|
|
2024-01-09 07:58:55 +00:00
|
|
|
max_val = buf_iw[lane_id];
|
2023-12-01 08:51:24 +00:00
|
|
|
max_val = warp_reduce_max(max_val);
|
2023-08-22 18:27:06 +00:00
|
|
|
}
|
|
|
|
|
2024-01-09 07:58:55 +00:00
|
|
|
float tmp = 0.0f; // partial sum
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2024-01-09 07:58:55 +00:00
|
|
|
#pragma unroll
|
|
|
|
for (int col0 = 0; col0 < ncols; col0 += block_size) {
|
|
|
|
const int col = col0 + tid;
|
|
|
|
|
|
|
|
if (ncols_template == 0 && col >= ncols) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
const float val = expf(vals[col] - max_val);
|
2023-06-14 17:47:19 +00:00
|
|
|
tmp += val;
|
2024-01-09 07:58:55 +00:00
|
|
|
vals[col] = val;
|
2023-06-14 17:47:19 +00:00
|
|
|
}
|
|
|
|
|
2023-12-01 08:51:24 +00:00
|
|
|
// find the sum of exps in the block
|
|
|
|
tmp = warp_reduce_sum(tmp);
|
|
|
|
if (block_size > WARP_SIZE) {
|
|
|
|
if (warp_id == 0) {
|
2024-01-09 07:58:55 +00:00
|
|
|
buf_iw[lane_id] = 0.0f;
|
2023-12-01 08:51:24 +00:00
|
|
|
}
|
|
|
|
__syncthreads();
|
|
|
|
|
|
|
|
if (lane_id == 0) {
|
2024-01-09 07:58:55 +00:00
|
|
|
buf_iw[warp_id] = tmp;
|
2023-12-01 08:51:24 +00:00
|
|
|
}
|
|
|
|
__syncthreads();
|
|
|
|
|
2024-01-09 07:58:55 +00:00
|
|
|
tmp = buf_iw[lane_id];
|
2023-12-01 08:51:24 +00:00
|
|
|
tmp = warp_reduce_sum(tmp);
|
2023-06-14 17:47:19 +00:00
|
|
|
}
|
|
|
|
|
2024-01-09 07:58:55 +00:00
|
|
|
const float inv_sum = 1.0f / tmp;
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2024-01-09 07:58:55 +00:00
|
|
|
#pragma unroll
|
|
|
|
for (int col0 = 0; col0 < ncols; col0 += block_size) {
|
|
|
|
const int col = col0 + tid;
|
|
|
|
|
|
|
|
if (ncols_template == 0 && col >= ncols) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
const int idst = rowx*ncols + col;
|
|
|
|
dst[idst] = vals[col] * inv_sum;
|
2023-06-14 17:47:19 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static __global__ void scale_f32(const float * x, float * dst, const float scale, const int k) {
|
|
|
|
const int i = blockDim.x*blockIdx.x + threadIdx.x;
|
|
|
|
|
|
|
|
if (i >= k) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
dst[i] = scale * x[i];
|
|
|
|
}
|
|
|
|
|
2023-10-10 07:50:23 +00:00
|
|
|
static __global__ void clamp_f32(const float * x, float * dst, const float min, const float max, const int k) {
|
|
|
|
const int i = blockDim.x*blockIdx.x + threadIdx.x;
|
|
|
|
|
|
|
|
if (i >= k) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
dst[i] = x[i] < min ? min : (x[i] > max ? max : x[i]);
|
|
|
|
}
|
2023-10-08 17:19:14 +00:00
|
|
|
|
2024-01-31 13:10:15 +00:00
|
|
|
template <typename T>
|
|
|
|
static __global__ void im2col_kernel(
|
|
|
|
const float * x, T * dst, int batch_offset,
|
|
|
|
int offset_delta, int IC, int IW, int IH, int OH, int OW, int KW, int KH, int pelements, int CHW,
|
2023-11-13 14:55:52 +00:00
|
|
|
int s0, int s1, int p0, int p1, int d0, int d1) {
|
2023-12-13 19:54:54 +00:00
|
|
|
const int i = threadIdx.x + blockIdx.x * blockDim.x;
|
|
|
|
if (i >= pelements) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
const int ksize = OW * (KH > 1 ? KW : 1);
|
|
|
|
const int kx = i / ksize;
|
|
|
|
const int kd = kx * ksize;
|
|
|
|
const int ky = (i - kd) / OW;
|
|
|
|
const int ix = i % OW;
|
|
|
|
|
2024-01-31 13:10:15 +00:00
|
|
|
const int oh = blockIdx.y;
|
|
|
|
const int batch = blockIdx.z / IC;
|
|
|
|
const int ic = blockIdx.z % IC;
|
|
|
|
|
2023-12-22 15:53:43 +00:00
|
|
|
const int64_t iiw = ix * s0 + kx * d0 - p0;
|
2024-01-31 13:10:15 +00:00
|
|
|
const int64_t iih = oh * s1 + ky * d1 - p1;
|
2023-11-13 14:55:52 +00:00
|
|
|
|
2023-12-22 15:53:43 +00:00
|
|
|
const int64_t offset_dst =
|
2024-01-31 13:10:15 +00:00
|
|
|
((batch * OH + oh) * OW + ix) * CHW +
|
|
|
|
(ic * (KW * KH) + ky * KW + kx);
|
2023-11-13 14:55:52 +00:00
|
|
|
|
|
|
|
if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
|
2024-01-31 13:10:15 +00:00
|
|
|
dst[offset_dst] = 0.0f;
|
2023-11-13 14:55:52 +00:00
|
|
|
} else {
|
2024-01-31 13:10:15 +00:00
|
|
|
const int64_t offset_src = ic * offset_delta + batch * batch_offset;
|
|
|
|
dst[offset_dst] = x[offset_src + iih * IW + iiw];
|
2023-11-13 14:55:52 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2024-01-31 13:10:15 +00:00
|
|
|
template <typename Ti, typename To>
|
|
|
|
static __global__ void pool2d_nchw_kernel(
|
|
|
|
const int ih, const int iw, const int oh, const int ow,
|
|
|
|
const int kh, const int kw, const int sh, const int sw,
|
|
|
|
const int ph, const int pw, const int parallel_elements,
|
|
|
|
const Ti* src, To* dst, const enum ggml_op_pool op) {
|
|
|
|
int idx = threadIdx.x + blockIdx.x * blockDim.x;
|
|
|
|
if (idx >= parallel_elements) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
const int I_HW = ih * iw;
|
|
|
|
const int O_HW = oh * ow;
|
|
|
|
const int nc = idx / O_HW;
|
|
|
|
const int cur_oh = idx % O_HW / ow;
|
|
|
|
const int cur_ow = idx % O_HW % ow;
|
|
|
|
const Ti* i_ptr = src + nc * I_HW;
|
|
|
|
To* o_ptr = dst + nc * O_HW;
|
|
|
|
const int start_h = cur_oh * sh - ph;
|
|
|
|
const int bh = max(0, start_h);
|
|
|
|
const int eh = min(ih, start_h + kh);
|
|
|
|
const int start_w = cur_ow * sw - pw;
|
|
|
|
const int bw = max(0, start_w);
|
|
|
|
const int ew = min(iw, start_w + kw);
|
|
|
|
const To scale = 1. / (kh * kw);
|
|
|
|
To res = 0;
|
|
|
|
|
|
|
|
switch (op) {
|
|
|
|
case GGML_OP_POOL_AVG: res = 0; break;
|
|
|
|
case GGML_OP_POOL_MAX: res = -FLT_MAX; break;
|
|
|
|
}
|
|
|
|
|
|
|
|
for (int i = bh; i < eh; i += 1) {
|
|
|
|
for (int j = bw; j < ew; j += 1) {
|
|
|
|
#if __CUDA_ARCH__ >= 350
|
|
|
|
Ti cur = __ldg(i_ptr + i * iw + j);
|
|
|
|
#else
|
|
|
|
Ti cur = i_ptr[i * iw + j];
|
|
|
|
#endif
|
|
|
|
switch (op) {
|
|
|
|
case GGML_OP_POOL_AVG: res += cur * scale; break;
|
|
|
|
case GGML_OP_POOL_MAX: res = max(res, (To)cur); break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
o_ptr[cur_oh * ow + cur_ow] = res;
|
|
|
|
}
|
|
|
|
|
2023-10-08 17:19:14 +00:00
|
|
|
template<int qk, int qr, dequantize_kernel_t dq>
|
2023-12-13 12:04:25 +00:00
|
|
|
static void get_rows_cuda(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
|
|
|
const void * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) {
|
|
|
|
|
|
|
|
GGML_TENSOR_BINARY_OP_LOCALS
|
|
|
|
|
|
|
|
const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
|
|
|
|
const int block_num_x = (ne00 + 2*CUDA_GET_ROWS_BLOCK_SIZE - 1) / (2*CUDA_GET_ROWS_BLOCK_SIZE);
|
|
|
|
const dim3 block_nums(block_num_x, ne10, ne11*ne12);
|
|
|
|
|
|
|
|
// strides in elements
|
|
|
|
//const size_t s0 = nb0 / ggml_element_size(dst);
|
|
|
|
const size_t s1 = nb1 / ggml_element_size(dst);
|
|
|
|
const size_t s2 = nb2 / ggml_element_size(dst);
|
|
|
|
const size_t s3 = nb3 / ggml_element_size(dst);
|
|
|
|
|
|
|
|
const size_t s10 = nb10 / ggml_element_size(src1);
|
|
|
|
const size_t s11 = nb11 / ggml_element_size(src1);
|
|
|
|
const size_t s12 = nb12 / ggml_element_size(src1);
|
|
|
|
//const size_t s13 = nb13 / ggml_element_size(src1);
|
|
|
|
|
|
|
|
GGML_ASSERT(ne00 % 2 == 0);
|
|
|
|
|
|
|
|
k_get_rows<qk, qr, dq><<<block_nums, block_dims, 0, stream>>>(
|
|
|
|
src0_dd, src1_dd, dst_dd,
|
|
|
|
ne00, /*ne01, ne02, ne03,*/
|
|
|
|
/*ne10, ne11,*/ ne12, /*ne13,*/
|
|
|
|
/* s0,*/ s1, s2, s3,
|
|
|
|
/* nb00,*/ nb01, nb02, nb03,
|
|
|
|
s10, s11, s12/*, s13*/);
|
|
|
|
|
|
|
|
(void) dst;
|
|
|
|
}
|
|
|
|
|
|
|
|
template<typename src0_t>
|
|
|
|
static void get_rows_cuda_float(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
|
|
|
const src0_t * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) {
|
|
|
|
|
|
|
|
GGML_TENSOR_BINARY_OP_LOCALS
|
|
|
|
|
2023-10-08 17:19:14 +00:00
|
|
|
const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
|
2023-12-13 12:04:25 +00:00
|
|
|
const int block_num_x = (ne00 + CUDA_GET_ROWS_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BLOCK_SIZE;
|
|
|
|
const dim3 block_nums(block_num_x, ne10, ne11*ne12);
|
|
|
|
|
|
|
|
// strides in elements
|
|
|
|
//const size_t s0 = nb0 / ggml_element_size(dst);
|
|
|
|
const size_t s1 = nb1 / ggml_element_size(dst);
|
|
|
|
const size_t s2 = nb2 / ggml_element_size(dst);
|
|
|
|
const size_t s3 = nb3 / ggml_element_size(dst);
|
|
|
|
|
|
|
|
const size_t s10 = nb10 / ggml_element_size(src1);
|
|
|
|
const size_t s11 = nb11 / ggml_element_size(src1);
|
|
|
|
const size_t s12 = nb12 / ggml_element_size(src1);
|
|
|
|
//const size_t s13 = nb13 / ggml_element_size(src1);
|
|
|
|
|
|
|
|
k_get_rows_float<<<block_nums, block_dims, 0, stream>>>(
|
|
|
|
src0_dd, src1_dd, dst_dd,
|
|
|
|
ne00, /*ne01, ne02, ne03,*/
|
|
|
|
/*ne10, ne11,*/ ne12, /*ne13,*/
|
|
|
|
/* s0,*/ s1, s2, s3,
|
|
|
|
/* nb00,*/ nb01, nb02, nb03,
|
|
|
|
s10, s11, s12/*, s13*/);
|
|
|
|
|
|
|
|
(void) dst;
|
2023-10-08 17:19:14 +00:00
|
|
|
}
|
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
template<float (*bin_op)(const float, const float)>
|
|
|
|
struct bin_bcast_cuda {
|
|
|
|
template<typename src0_t, typename src1_t, typename dst_t>
|
|
|
|
void operator()(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst,
|
|
|
|
const src0_t * src0_dd, const src1_t * src1_dd, dst_t * dst_dd,
|
|
|
|
cudaStream_t stream) {
|
|
|
|
|
|
|
|
GGML_TENSOR_BINARY_OP_LOCALS
|
|
|
|
|
|
|
|
int nr0 = ne10/ne0;
|
|
|
|
int nr1 = ne11/ne1;
|
|
|
|
int nr2 = ne12/ne2;
|
|
|
|
int nr3 = ne13/ne3;
|
|
|
|
|
|
|
|
int nr[4] = { nr0, nr1, nr2, nr3 };
|
|
|
|
|
|
|
|
// collapse dimensions until first broadcast dimension
|
|
|
|
int64_t cne0[] = {ne0, ne1, ne2, ne3};
|
|
|
|
int64_t cne1[] = {ne10, ne11, ne12, ne13};
|
|
|
|
size_t cnb0[] = {nb0, nb1, nb2, nb3};
|
|
|
|
size_t cnb1[] = {nb10, nb11, nb12, nb13};
|
|
|
|
auto collapse = [](int64_t cne[]) {
|
|
|
|
cne[0] *= cne[1];
|
|
|
|
cne[1] = cne[2];
|
|
|
|
cne[2] = cne[3];
|
|
|
|
cne[3] = 1;
|
|
|
|
};
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
auto collapse_nb = [](size_t cnb[], const int64_t cne[]) {
|
2023-12-07 20:26:54 +00:00
|
|
|
cnb[1] *= cne[1];
|
|
|
|
cnb[2] *= cne[2];
|
|
|
|
cnb[3] *= cne[3];
|
|
|
|
};
|
|
|
|
|
|
|
|
for (int i = 0; i < 4; i++) {
|
|
|
|
if (nr[i] != 1) {
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
if (i > 0) {
|
|
|
|
collapse_nb(cnb0, cne0);
|
|
|
|
collapse_nb(cnb1, cne1);
|
|
|
|
collapse(cne0);
|
|
|
|
collapse(cne1);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
{
|
|
|
|
int64_t ne0 = cne0[0];
|
|
|
|
int64_t ne1 = cne0[1];
|
|
|
|
int64_t ne2 = cne0[2];
|
|
|
|
int64_t ne3 = cne0[3];
|
|
|
|
|
|
|
|
int64_t ne10 = cne1[0];
|
|
|
|
int64_t ne11 = cne1[1];
|
|
|
|
int64_t ne12 = cne1[2];
|
|
|
|
int64_t ne13 = cne1[3];
|
|
|
|
|
2023-12-13 12:04:25 +00:00
|
|
|
size_t nb0 = cnb0[0];
|
2023-12-07 20:26:54 +00:00
|
|
|
size_t nb1 = cnb0[1];
|
|
|
|
size_t nb2 = cnb0[2];
|
|
|
|
size_t nb3 = cnb0[3];
|
|
|
|
|
2023-12-13 12:04:25 +00:00
|
|
|
size_t nb10 = cnb1[0];
|
2023-12-07 20:26:54 +00:00
|
|
|
size_t nb11 = cnb1[1];
|
|
|
|
size_t nb12 = cnb1[2];
|
|
|
|
size_t nb13 = cnb1[3];
|
|
|
|
|
2023-12-13 19:54:54 +00:00
|
|
|
size_t s0 = nb0 / sizeof(dst_t);
|
|
|
|
size_t s1 = nb1 / sizeof(dst_t);
|
|
|
|
size_t s2 = nb2 / sizeof(dst_t);
|
|
|
|
size_t s3 = nb3 / sizeof(dst_t);
|
2023-12-07 20:26:54 +00:00
|
|
|
|
2023-12-13 12:04:25 +00:00
|
|
|
size_t s10 = nb10 / sizeof(src1_t);
|
2023-12-07 20:26:54 +00:00
|
|
|
size_t s11 = nb11 / sizeof(src1_t);
|
|
|
|
size_t s12 = nb12 / sizeof(src1_t);
|
|
|
|
size_t s13 = nb13 / sizeof(src1_t);
|
|
|
|
|
2023-12-13 12:04:25 +00:00
|
|
|
GGML_ASSERT(s0 == 1);
|
|
|
|
GGML_ASSERT(s10 == 1);
|
2023-12-07 20:26:54 +00:00
|
|
|
|
|
|
|
const int block_size = 128;
|
|
|
|
|
|
|
|
int64_t hne0 = std::max(ne0/2LL, 1LL);
|
|
|
|
|
|
|
|
dim3 block_dims;
|
|
|
|
block_dims.x = std::min<unsigned int>(hne0, block_size);
|
|
|
|
block_dims.y = std::min<unsigned int>(ne1, block_size / block_dims.x);
|
|
|
|
block_dims.z = std::min(std::min<unsigned int>(ne2*ne3, block_size / block_dims.x / block_dims.y), 64U);
|
|
|
|
|
|
|
|
dim3 block_nums(
|
|
|
|
(hne0 + block_dims.x - 1) / block_dims.x,
|
|
|
|
(ne1 + block_dims.y - 1) / block_dims.y,
|
|
|
|
(ne2*ne3 + block_dims.z - 1) / block_dims.z
|
|
|
|
);
|
2023-11-01 11:49:04 +00:00
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
if (block_nums.z > 65535) {
|
|
|
|
// this is the maximum number of blocks in z direction, fallback to 1D grid kernel
|
|
|
|
int block_num = (ne0*ne1*ne2*ne3 + block_size - 1) / block_size;
|
|
|
|
k_bin_bcast_unravel<bin_op><<<block_num, block_size, 0, stream>>>(
|
|
|
|
src0_dd, src1_dd, dst_dd,
|
|
|
|
ne0, ne1, ne2, ne3,
|
|
|
|
ne10, ne11, ne12, ne13,
|
|
|
|
/* s0, */ s1, s2, s3,
|
|
|
|
/* s10, */ s11, s12, s13);
|
|
|
|
} else {
|
|
|
|
k_bin_bcast<bin_op><<<block_nums, block_dims, 0, stream>>>(
|
|
|
|
src0_dd, src1_dd, dst_dd,
|
|
|
|
ne0, ne1, ne2, ne3,
|
|
|
|
ne10, ne11, ne12, ne13,
|
|
|
|
/* s0, */ s1, s2, s3,
|
|
|
|
/* s10, */ s11, s12, s13);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
};
|
cuda : loading models directly into VRAM, norm calculation on GPU, broadcasting for ggml_mul (#1483)
* Broadcasting for ggml_mul
* CUDA kernel for ggml_mul, norms in VRAM
* GPU weights not in RAM, direct loading with cuFile
* fixup! GPU weights not in RAM, direct loading with cuFile
* fixup! GPU weights not in RAM, direct loading with cuFile
* define default model path once, sync path with readme (#1366)
* ~7% faster Q5_1 AVX2 code (#1477)
* convert.py: Support models which are stored in a single pytorch_model.bin (#1469)
* Support models in a single pytorch_model.bin
* Remove spurious line with typo
* benchmark-matmul: Print the average of the test results (#1490)
* Remove unused n_parts parameter (#1509)
* Fixes #1511 lambda issue for w64devkit (mingw) (#1513)
* Fix for w64devkit and mingw
* make kv_f16 the default for api users (#1517)
* minor : fix compile warnings
* readme : adds WizardLM to the list of supported models (#1485)
* main : make reverse prompt option act as a stop token in non-interactive mode (#1032)
* Make reverse prompt option act as a stop token in non-interactive scenarios
* Making requested review changes
* Update gpt_params_parse and fix a merge error
* Revert "Update gpt_params_parse and fix a merge error"
This reverts commit 2bb2ff1748513591ad45b175a75ed1d8089d84c8.
* Update gpt_params_parse and fix a merge error take 2
* examples : add persistent chat (#1495)
* examples : add persistent chat
* examples : fix whitespace
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* tests : add missing header
* ggml : use F16 instead of F32 in Q4_0, Q4_1, Q8_0 (#1508)
* ggml : use F16 instead of F32 in Q4_0, Q4_1 and Q8_0
* llama : bump LLAMA_FILE_VERSION to 3
* cuda : update Q4 and Q8 dequantize kernels
* ggml : fix AVX dot products
* readme : update performance table + hot topics
* ggml : fix scalar implementation of Q4_1 dot
* llama : fix compile warnings in llama_set_state_data()
* llama : fix name shadowing and C4146 (#1526)
* Fix name shadowing and C4146
* Fix if macros not using defined when required
* Update llama-util.h
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
* Update llama-util.h
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
* Code style
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Fix for mingw (#1462)
* llama : add llama_init_backend() API (close #1527)
* feature : add blis and other BLAS implementation support (#1502)
* feature: add blis support
* feature: allow all BLA_VENDOR to be assigned in cmake arguments. align with whisper.cpp pr 927
* fix: version detection for BLA_SIZEOF_INTEGER, recover min version of cmake
* Fix typo in INTEGER
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Revert "feature : add blis and other BLAS implementation support (#1502)"
This reverts commit 07e9ace0f9da424d82e75df969642522880feb92.
* GPU weights not in RAM, direct loading with cuFile
* llama : code style fixes + progress print fix
* ggml : ggml_mul better broadcast support
* cmake : workarounds for cufile when CMake version < 3.25
* gg rebase fixup
* Loop in llama.cpp, fixed progress callback
* Attempt clang-tidy fix
* llama : fix vram size computation
* Add forgotten fclose()
---------
Co-authored-by: András Salamon <ott2@users.noreply.github.com>
Co-authored-by: Ilya Kurdyukov <59548320+ilyakurdyukov@users.noreply.github.com>
Co-authored-by: Tom Jobbins <784313+TheBloke@users.noreply.github.com>
Co-authored-by: rankaiyx <rankaiyx@rankaiyx.com>
Co-authored-by: Stephan Walter <stephan@walter.name>
Co-authored-by: DannyDaemonic <DannyDaemonic@gmail.com>
Co-authored-by: Erik Scholz <Green-Sky@users.noreply.github.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: David Kennedy <dakennedyd@gmail.com>
Co-authored-by: Jason McCartney <jmac@theroot.org>
Co-authored-by: Evan Jones <evan.q.jones@gmail.com>
Co-authored-by: Maxime <672982+maximegmd@users.noreply.github.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Zenix <zenixls2@gmail.com>
2023-05-20 12:19:28 +00:00
|
|
|
|
2023-12-13 19:54:54 +00:00
|
|
|
static void acc_f32_cuda(const float * x, const float * y, float * dst, const int n_elements,
|
|
|
|
const int ne10, const int ne11, const int ne12,
|
|
|
|
const int nb1, const int nb2, const int offset, cudaStream_t stream) {
|
|
|
|
int num_blocks = (n_elements + CUDA_ACC_BLOCK_SIZE - 1) / CUDA_ACC_BLOCK_SIZE;
|
|
|
|
acc_f32<<<num_blocks, CUDA_ACC_BLOCK_SIZE, 0, stream>>>(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset);
|
|
|
|
}
|
|
|
|
|
2023-07-12 17:26:18 +00:00
|
|
|
static void gelu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
|
|
|
|
const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE;
|
|
|
|
gelu_f32<<<num_blocks, CUDA_GELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
|
|
|
|
}
|
|
|
|
|
2023-06-06 19:33:23 +00:00
|
|
|
static void silu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
|
|
|
|
const int num_blocks = (k + CUDA_SILU_BLOCK_SIZE - 1) / CUDA_SILU_BLOCK_SIZE;
|
|
|
|
silu_f32<<<num_blocks, CUDA_SILU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
|
|
|
|
}
|
|
|
|
|
2023-12-13 19:54:54 +00:00
|
|
|
static void gelu_quick_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
|
|
|
|
const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE;
|
|
|
|
gelu_quick_f32<<<num_blocks, CUDA_GELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void tanh_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
|
|
|
|
const int num_blocks = (k + CUDA_TANH_BLOCK_SIZE - 1) / CUDA_TANH_BLOCK_SIZE;
|
|
|
|
tanh_f32<<<num_blocks, CUDA_TANH_BLOCK_SIZE, 0, stream>>>(x, dst, k);
|
|
|
|
}
|
|
|
|
|
2023-11-13 08:58:15 +00:00
|
|
|
static void relu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
|
|
|
|
const int num_blocks = (k + CUDA_RELU_BLOCK_SIZE - 1) / CUDA_RELU_BLOCK_SIZE;
|
|
|
|
relu_f32<<<num_blocks, CUDA_RELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
|
|
|
|
}
|
|
|
|
|
2024-01-31 13:10:15 +00:00
|
|
|
static void hardsigmoid_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
|
|
|
|
const int num_blocks = (k + CUDA_HARDSIGMOID_BLOCK_SIZE - 1) / CUDA_HARDSIGMOID_BLOCK_SIZE;
|
|
|
|
hardsigmoid_f32<<<num_blocks, CUDA_HARDSIGMOID_BLOCK_SIZE, 0, stream>>>(x, dst, k);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void hardswish_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
|
|
|
|
const int num_blocks = (k + CUDA_HARDSWISH_BLOCK_SIZE - 1) / CUDA_HARDSWISH_BLOCK_SIZE;
|
|
|
|
hardswish_f32<<<num_blocks, CUDA_HARDSWISH_BLOCK_SIZE, 0, stream>>>(x, dst, k);
|
|
|
|
}
|
|
|
|
|
2023-12-13 19:54:54 +00:00
|
|
|
static void leaky_relu_f32_cuda(const float * x, float * dst, const int k, const float negative_slope, cudaStream_t stream) {
|
|
|
|
const int num_blocks = (k + CUDA_RELU_BLOCK_SIZE - 1) / CUDA_RELU_BLOCK_SIZE;
|
|
|
|
leaky_relu_f32<<<num_blocks, CUDA_RELU_BLOCK_SIZE, 0, stream>>>(x, dst, k, negative_slope);
|
|
|
|
}
|
|
|
|
|
2023-11-13 08:58:15 +00:00
|
|
|
static void sqr_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
|
|
|
|
const int num_blocks = (k + CUDA_SQR_BLOCK_SIZE - 1) / CUDA_SQR_BLOCK_SIZE;
|
|
|
|
sqr_f32<<<num_blocks, CUDA_SQR_BLOCK_SIZE, 0, stream>>>(x, dst, k);
|
|
|
|
}
|
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
static void norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float eps, cudaStream_t stream) {
|
2023-07-11 19:53:34 +00:00
|
|
|
GGML_ASSERT(ncols % WARP_SIZE == 0);
|
2023-09-04 06:53:30 +00:00
|
|
|
if (ncols < 1024) {
|
|
|
|
const dim3 block_dims(WARP_SIZE, 1, 1);
|
2023-12-07 20:26:54 +00:00
|
|
|
norm_f32<WARP_SIZE><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
|
2023-09-04 06:53:30 +00:00
|
|
|
} else {
|
|
|
|
const dim3 block_dims(1024, 1, 1);
|
2023-12-07 20:26:54 +00:00
|
|
|
norm_f32<1024><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
|
2023-09-04 06:53:30 +00:00
|
|
|
}
|
2023-07-11 19:53:34 +00:00
|
|
|
}
|
|
|
|
|
2023-12-13 19:54:54 +00:00
|
|
|
static void group_norm_f32_cuda(const float * x, float * dst, const int num_groups, const int group_size, const int ne_elements, cudaStream_t stream) {
|
|
|
|
static const float eps = 1e-6f;
|
|
|
|
if (group_size < 1024) {
|
|
|
|
const dim3 block_dims(WARP_SIZE, 1, 1);
|
|
|
|
group_norm_f32<WARP_SIZE><<<num_groups, block_dims, 0, stream>>>(x, dst, group_size, ne_elements, eps);
|
|
|
|
} else {
|
|
|
|
const dim3 block_dims(1024, 1, 1);
|
|
|
|
group_norm_f32<1024><<<num_groups, block_dims, 0, stream>>>(x, dst, group_size, ne_elements, eps);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static void concat_f32_cuda(const float * x, const float * y, float * dst, const int ne0, int ne1, int ne2, int ne02, cudaStream_t stream) {
|
|
|
|
int num_blocks = (ne0 + CUDA_CONCAT_BLOCK_SIZE - 1) / CUDA_CONCAT_BLOCK_SIZE;
|
|
|
|
dim3 gridDim(num_blocks, ne1, ne2);
|
|
|
|
concat_f32<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne02);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void upscale_f32_cuda(const float * x, float * dst, const int ne00, const int ne01, const int ne02, const int scale_factor, cudaStream_t stream) {
|
|
|
|
int ne0 = (ne00 * scale_factor);
|
|
|
|
int num_blocks = (ne0 + CUDA_UPSCALE_BLOCK_SIZE - 1) / CUDA_UPSCALE_BLOCK_SIZE;
|
|
|
|
dim3 gridDim(num_blocks, (ne01 * scale_factor), ne02);
|
|
|
|
upscale_f32<<<gridDim, CUDA_UPSCALE_BLOCK_SIZE, 0, stream>>>(x, dst, ne00, ne00 * ne01, scale_factor);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void pad_f32_cuda(const float * x, float * dst,
|
|
|
|
const int ne00, const int ne01, const int ne02,
|
|
|
|
const int ne0, const int ne1, const int ne2, cudaStream_t stream) {
|
|
|
|
int num_blocks = (ne0 + CUDA_PAD_BLOCK_SIZE - 1) / CUDA_PAD_BLOCK_SIZE;
|
|
|
|
dim3 gridDim(num_blocks, ne1, ne2);
|
|
|
|
pad_f32<<<gridDim, CUDA_PAD_BLOCK_SIZE, 0, stream>>>(x, dst, ne0, ne00, ne01, ne02);
|
|
|
|
}
|
|
|
|
|
2023-07-24 15:57:12 +00:00
|
|
|
static void rms_norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float eps, cudaStream_t stream) {
|
2023-06-06 19:33:23 +00:00
|
|
|
GGML_ASSERT(ncols % WARP_SIZE == 0);
|
2023-09-04 06:53:30 +00:00
|
|
|
if (ncols < 1024) {
|
|
|
|
const dim3 block_dims(WARP_SIZE, 1, 1);
|
|
|
|
rms_norm_f32<WARP_SIZE><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
|
|
|
|
} else {
|
|
|
|
const dim3 block_dims(1024, 1, 1);
|
|
|
|
rms_norm_f32<1024><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
|
|
|
|
}
|
2023-06-06 19:33:23 +00:00
|
|
|
}
|
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
static void quantize_row_q8_1_cuda(const float * x, void * vy, const int kx, const int ky, const int kx_padded, cudaStream_t stream) {
|
|
|
|
const int block_num_x = (kx_padded + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE;
|
|
|
|
const dim3 num_blocks(block_num_x, ky, 1);
|
|
|
|
const dim3 block_size(CUDA_DEQUANTIZE_BLOCK_SIZE, 1, 1);
|
|
|
|
quantize_q8_1<<<num_blocks, block_size, 0, stream>>>(x, vy, kx, kx_padded);
|
2023-07-05 12:19:42 +00:00
|
|
|
}
|
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
|
|
|
|
static void dequantize_block_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int k, cudaStream_t stream) {
|
2024-01-07 16:24:08 +00:00
|
|
|
const int num_blocks = (k + 2*CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / (2*CUDA_DEQUANTIZE_BLOCK_SIZE);
|
2023-12-07 20:26:54 +00:00
|
|
|
dequantize_block<qk, qr, dequantize_kernel><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
|
2023-04-25 20:40:51 +00:00
|
|
|
}
|
|
|
|
|
2024-01-12 19:38:54 +00:00
|
|
|
static void dequantize_block_q8_0_f16_cuda(const void * __restrict__ vx, half * __restrict__ y, const int k, cudaStream_t stream) {
|
|
|
|
const int num_blocks = (k + CUDA_Q8_0_NE_ALIGN - 1) / CUDA_Q8_0_NE_ALIGN;
|
|
|
|
if (k % CUDA_Q8_0_NE_ALIGN == 0) {
|
|
|
|
const bool need_check = false;
|
|
|
|
dequantize_block_q8_0_f16<need_check><<<num_blocks, WARP_SIZE, 0, stream>>>(vx, y, k);
|
|
|
|
} else {
|
|
|
|
const bool need_check = true;
|
|
|
|
dequantize_block_q8_0_f16<need_check><<<num_blocks, WARP_SIZE, 0, stream>>>(vx, y, k);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-09-30 16:12:57 +00:00
|
|
|
template<typename dst_t>
|
|
|
|
static void dequantize_row_q2_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
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 = k / 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
|
|
|
#if QK_K == 256
|
2023-06-07 07:59:52 +00:00
|
|
|
dequantize_block_q2_K<<<nb, 64, 0, stream>>>(vx, y);
|
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
|
|
|
|
dequantize_block_q2_K<<<nb, 32, 0, stream>>>(vx, y);
|
|
|
|
#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
|
|
|
}
|
|
|
|
|
2023-09-30 16:12:57 +00:00
|
|
|
template<typename dst_t>
|
|
|
|
static void dequantize_row_q3_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
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 = k / 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
|
|
|
#if QK_K == 256
|
2023-06-07 07:59:52 +00:00
|
|
|
dequantize_block_q3_K<<<nb, 64, 0, stream>>>(vx, y);
|
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
|
|
|
|
dequantize_block_q3_K<<<nb, 32, 0, stream>>>(vx, y);
|
|
|
|
#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
|
|
|
}
|
|
|
|
|
2024-01-15 05:48:06 +00:00
|
|
|
template<typename dst_t>
|
2024-01-15 11:27:00 +00:00
|
|
|
static void dequantize_row_q4_0_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
2024-01-15 05:48:06 +00:00
|
|
|
const int nb32 = k / 32;
|
|
|
|
const int nb = (k + 255) / 256;
|
|
|
|
dequantize_block_q4_0<<<nb, 32, 0, stream>>>(vx, y, nb32);
|
|
|
|
}
|
|
|
|
|
|
|
|
template<typename dst_t>
|
2024-01-15 11:27:00 +00:00
|
|
|
static void dequantize_row_q4_1_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
2024-01-15 05:48:06 +00:00
|
|
|
const int nb32 = k / 32;
|
|
|
|
const int nb = (k + 255) / 256;
|
|
|
|
dequantize_block_q4_1<<<nb, 32, 0, stream>>>(vx, y, nb32);
|
|
|
|
}
|
|
|
|
|
2023-09-30 16:12:57 +00:00
|
|
|
template<typename dst_t>
|
|
|
|
static void dequantize_row_q4_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
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 = k / QK_K;
|
2023-06-07 07:59:52 +00:00
|
|
|
dequantize_block_q4_K<<<nb, 32, 0, stream>>>(vx, y);
|
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
|
|
|
}
|
|
|
|
|
2023-09-30 16:12:57 +00:00
|
|
|
template<typename dst_t>
|
|
|
|
static void dequantize_row_q5_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
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 = k / 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
|
|
|
#if QK_K == 256
|
2023-06-07 07:59:52 +00:00
|
|
|
dequantize_block_q5_K<<<nb, 64, 0, stream>>>(vx, y);
|
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
|
|
|
|
dequantize_block_q5_K<<<nb, 32, 0, stream>>>(vx, y);
|
|
|
|
#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
|
|
|
}
|
|
|
|
|
2023-09-30 16:12:57 +00:00
|
|
|
template<typename dst_t>
|
|
|
|
static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
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 = k / 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
|
|
|
#if QK_K == 256
|
2023-06-07 07:59:52 +00:00
|
|
|
dequantize_block_q6_K<<<nb, 64, 0, stream>>>(vx, y);
|
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
|
|
|
|
dequantize_block_q6_K<<<nb, 32, 0, stream>>>(vx, y);
|
|
|
|
#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
|
|
|
}
|
|
|
|
|
2024-01-08 15:02:32 +00:00
|
|
|
template<typename dst_t>
|
|
|
|
static void dequantize_row_iq2_xxs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
|
|
|
const int nb = k / QK_K;
|
|
|
|
dequantize_block_iq2_xxs<<<nb, 32, 0, stream>>>(vx, y);
|
|
|
|
}
|
|
|
|
|
2024-01-11 19:39:39 +00:00
|
|
|
template<typename dst_t>
|
|
|
|
static void dequantize_row_iq2_xs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
|
|
|
const int nb = k / QK_K;
|
|
|
|
dequantize_block_iq2_xs<<<nb, 32, 0, stream>>>(vx, y);
|
|
|
|
}
|
|
|
|
|
2024-02-26 16:28:38 +00:00
|
|
|
template<typename dst_t>
|
|
|
|
static void dequantize_row_iq2_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
|
|
|
const int nb = k / QK_K;
|
|
|
|
dequantize_block_iq2_s<<<nb, 32, 0, stream>>>(vx, y);
|
|
|
|
}
|
|
|
|
|
2024-01-30 13:14:12 +00:00
|
|
|
template<typename dst_t>
|
|
|
|
static void dequantize_row_iq3_xxs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
|
|
|
const int nb = k / QK_K;
|
|
|
|
dequantize_block_iq3_xxs<<<nb, 32, 0, stream>>>(vx, y);
|
|
|
|
}
|
|
|
|
|
2024-02-24 14:23:52 +00:00
|
|
|
template<typename dst_t>
|
|
|
|
static void dequantize_row_iq3_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
|
|
|
const int nb = k / QK_K;
|
|
|
|
dequantize_block_iq3_s<<<nb, 32, 0, stream>>>(vx, y);
|
|
|
|
}
|
|
|
|
|
2024-02-18 16:16:55 +00:00
|
|
|
template<typename dst_t>
|
|
|
|
static void dequantize_row_iq1_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
|
|
|
const int nb = k / QK_K;
|
|
|
|
dequantize_block_iq1_s<<<nb, 32, 0, stream>>>(vx, y);
|
|
|
|
}
|
|
|
|
|
2024-02-21 09:39:52 +00:00
|
|
|
template<typename dst_t>
|
|
|
|
static void dequantize_row_iq4_nl_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
|
|
|
const int nb = (k + QK_K - 1) / QK_K;
|
|
|
|
dequantize_block_iq4_nl<<<nb, 32, 0, stream>>>(vx, y);
|
|
|
|
}
|
|
|
|
|
2024-02-27 14:34:24 +00:00
|
|
|
template<typename dst_t>
|
|
|
|
static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
|
|
|
const int nb = (k + QK_K - 1) / QK_K;
|
2024-02-28 08:37:02 +00:00
|
|
|
#if QK_K == 64
|
|
|
|
dequantize_block_iq4_nl<<<nb, 32, 0, stream>>>(vx, y);
|
|
|
|
#else
|
2024-02-27 14:34:24 +00:00
|
|
|
dequantize_block_iq4_xs<<<nb, 32, 0, stream>>>(vx, y);
|
2024-02-28 08:37:02 +00:00
|
|
|
#endif
|
2024-02-27 14:34:24 +00:00
|
|
|
}
|
|
|
|
|
2024-01-07 16:24:08 +00:00
|
|
|
template <typename src_t, typename dst_t>
|
|
|
|
static void convert_unary_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int k, cudaStream_t stream) {
|
|
|
|
const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
|
|
|
|
convert_unary<src_t><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
|
|
|
|
}
|
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
static to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
|
2024-01-12 19:38:54 +00:00
|
|
|
int id;
|
2023-12-07 20:26:54 +00:00
|
|
|
switch (type) {
|
|
|
|
case GGML_TYPE_Q4_0:
|
2024-01-15 11:27:00 +00:00
|
|
|
return dequantize_row_q4_0_cuda;
|
2023-12-07 20:26:54 +00:00
|
|
|
case GGML_TYPE_Q4_1:
|
2024-01-15 11:27:00 +00:00
|
|
|
return dequantize_row_q4_1_cuda;
|
2023-12-07 20:26:54 +00:00
|
|
|
case GGML_TYPE_Q5_0:
|
|
|
|
return dequantize_block_cuda<QK5_0, QR5_0, dequantize_q5_0>;
|
|
|
|
case GGML_TYPE_Q5_1:
|
|
|
|
return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
|
|
|
|
case GGML_TYPE_Q8_0:
|
2024-01-12 19:38:54 +00:00
|
|
|
CUDA_CHECK(cudaGetDevice(&id));
|
|
|
|
if (g_device_caps[id].cc >= CC_PASCAL) {
|
|
|
|
return dequantize_block_q8_0_f16_cuda;
|
|
|
|
}
|
2023-12-07 20:26:54 +00:00
|
|
|
return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
|
|
|
|
case GGML_TYPE_Q2_K:
|
|
|
|
return dequantize_row_q2_K_cuda;
|
|
|
|
case GGML_TYPE_Q3_K:
|
|
|
|
return dequantize_row_q3_K_cuda;
|
|
|
|
case GGML_TYPE_Q4_K:
|
|
|
|
return dequantize_row_q4_K_cuda;
|
|
|
|
case GGML_TYPE_Q5_K:
|
|
|
|
return dequantize_row_q5_K_cuda;
|
|
|
|
case GGML_TYPE_Q6_K:
|
|
|
|
return dequantize_row_q6_K_cuda;
|
2024-01-08 15:02:32 +00:00
|
|
|
case GGML_TYPE_IQ2_XXS:
|
|
|
|
return dequantize_row_iq2_xxs_cuda;
|
2024-01-11 19:39:39 +00:00
|
|
|
case GGML_TYPE_IQ2_XS:
|
|
|
|
return dequantize_row_iq2_xs_cuda;
|
2024-02-26 16:28:38 +00:00
|
|
|
case GGML_TYPE_IQ2_S:
|
|
|
|
return dequantize_row_iq2_s_cuda;
|
2024-01-30 13:14:12 +00:00
|
|
|
case GGML_TYPE_IQ3_XXS:
|
|
|
|
return dequantize_row_iq3_xxs_cuda;
|
2024-02-18 16:16:55 +00:00
|
|
|
case GGML_TYPE_IQ1_S:
|
|
|
|
return dequantize_row_iq1_s_cuda;
|
2024-02-21 09:39:52 +00:00
|
|
|
case GGML_TYPE_IQ4_NL:
|
|
|
|
return dequantize_row_iq4_nl_cuda;
|
2024-02-27 14:34:24 +00:00
|
|
|
case GGML_TYPE_IQ4_XS:
|
|
|
|
return dequantize_row_iq4_xs_cuda;
|
2024-02-24 14:23:52 +00:00
|
|
|
case GGML_TYPE_IQ3_S:
|
|
|
|
return dequantize_row_iq3_s_cuda;
|
2023-12-07 20:26:54 +00:00
|
|
|
case GGML_TYPE_F32:
|
2024-01-07 16:24:08 +00:00
|
|
|
return convert_unary_cuda<float>;
|
2023-12-07 20:26:54 +00:00
|
|
|
default:
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
|
|
|
|
switch (type) {
|
|
|
|
case GGML_TYPE_Q4_0:
|
2024-01-15 11:27:00 +00:00
|
|
|
return dequantize_row_q4_0_cuda;
|
2023-12-07 20:26:54 +00:00
|
|
|
case GGML_TYPE_Q4_1:
|
2024-01-15 11:27:00 +00:00
|
|
|
return dequantize_row_q4_1_cuda;
|
2023-12-07 20:26:54 +00:00
|
|
|
case GGML_TYPE_Q5_0:
|
|
|
|
return dequantize_block_cuda<QK5_0, QR5_0, dequantize_q5_0>;
|
|
|
|
case GGML_TYPE_Q5_1:
|
|
|
|
return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
|
|
|
|
case GGML_TYPE_Q8_0:
|
|
|
|
return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
|
|
|
|
case GGML_TYPE_Q2_K:
|
|
|
|
return dequantize_row_q2_K_cuda;
|
|
|
|
case GGML_TYPE_Q3_K:
|
|
|
|
return dequantize_row_q3_K_cuda;
|
|
|
|
case GGML_TYPE_Q4_K:
|
|
|
|
return dequantize_row_q4_K_cuda;
|
|
|
|
case GGML_TYPE_Q5_K:
|
|
|
|
return dequantize_row_q5_K_cuda;
|
|
|
|
case GGML_TYPE_Q6_K:
|
|
|
|
return dequantize_row_q6_K_cuda;
|
2024-01-08 15:02:32 +00:00
|
|
|
case GGML_TYPE_IQ2_XXS:
|
|
|
|
return dequantize_row_iq2_xxs_cuda;
|
2024-01-11 19:39:39 +00:00
|
|
|
case GGML_TYPE_IQ2_XS:
|
|
|
|
return dequantize_row_iq2_xs_cuda;
|
2024-02-26 16:28:38 +00:00
|
|
|
case GGML_TYPE_IQ2_S:
|
|
|
|
return dequantize_row_iq2_s_cuda;
|
2024-01-30 13:14:12 +00:00
|
|
|
case GGML_TYPE_IQ3_XXS:
|
|
|
|
return dequantize_row_iq3_xxs_cuda;
|
2024-02-18 16:16:55 +00:00
|
|
|
case GGML_TYPE_IQ1_S:
|
|
|
|
return dequantize_row_iq1_s_cuda;
|
2024-02-21 09:39:52 +00:00
|
|
|
case GGML_TYPE_IQ4_NL:
|
|
|
|
return dequantize_row_iq4_nl_cuda;
|
2024-02-27 14:34:24 +00:00
|
|
|
case GGML_TYPE_IQ4_XS:
|
|
|
|
return dequantize_row_iq4_xs_cuda;
|
2024-02-24 14:23:52 +00:00
|
|
|
case GGML_TYPE_IQ3_S:
|
|
|
|
return dequantize_row_iq3_s_cuda;
|
2023-12-07 20:26:54 +00:00
|
|
|
case GGML_TYPE_F16:
|
2024-01-07 16:24:08 +00:00
|
|
|
return convert_unary_cuda<half>;
|
2023-12-07 20:26:54 +00:00
|
|
|
default:
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-06-19 08:23:56 +00:00
|
|
|
static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
|
2023-05-25 21:07:29 +00:00
|
|
|
GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
|
2023-07-05 12:19:42 +00:00
|
|
|
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
|
2023-11-03 11:13:09 +00:00
|
|
|
// the number of rows may exceed maximum grid size in the y or z dimensions, use the x dimension instead
|
|
|
|
const dim3 block_nums(block_num_y, 1, 1);
|
2023-07-05 12:19:42 +00:00
|
|
|
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
|
2023-05-25 21:07:29 +00:00
|
|
|
dequantize_mul_mat_vec<QK4_0, QR4_0, dequantize_q4_0>
|
2023-06-14 17:47:19 +00:00
|
|
|
<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
|
2023-05-13 13:38:36 +00:00
|
|
|
}
|
|
|
|
|
2023-06-19 08:23:56 +00:00
|
|
|
static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
|
2023-05-25 21:07:29 +00:00
|
|
|
GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
|
2023-07-05 12:19:42 +00:00
|
|
|
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
|
2023-11-03 11:13:09 +00:00
|
|
|
const dim3 block_nums(block_num_y, 1, 1);
|
2023-07-05 12:19:42 +00:00
|
|
|
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
|
2023-05-25 21:07:29 +00:00
|
|
|
dequantize_mul_mat_vec<QK4_1, QR4_1, dequantize_q4_1>
|
2023-06-14 17:47:19 +00:00
|
|
|
<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
|
2023-05-13 13:38:36 +00:00
|
|
|
}
|
|
|
|
|
2023-06-19 08:23:56 +00:00
|
|
|
static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
|
2023-05-25 21:07:29 +00:00
|
|
|
GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
|
2023-07-05 12:19:42 +00:00
|
|
|
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
|
2023-11-03 11:13:09 +00:00
|
|
|
const dim3 block_nums(block_num_y, 1, 1);
|
2023-07-05 12:19:42 +00:00
|
|
|
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
|
2023-05-25 21:07:29 +00:00
|
|
|
dequantize_mul_mat_vec<QK5_0, QR5_0, dequantize_q5_0>
|
2023-06-14 17:47:19 +00:00
|
|
|
<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
|
2023-05-13 13:38:36 +00:00
|
|
|
}
|
|
|
|
|
2023-06-19 08:23:56 +00:00
|
|
|
static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
|
2023-05-25 21:07:29 +00:00
|
|
|
GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
|
2023-07-05 12:19:42 +00:00
|
|
|
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
|
2023-11-03 11:13:09 +00:00
|
|
|
const dim3 block_nums(block_num_y, 1, 1);
|
2023-07-05 12:19:42 +00:00
|
|
|
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
|
2023-05-25 21:07:29 +00:00
|
|
|
dequantize_mul_mat_vec<QK5_1, QR5_1, dequantize_q5_1>
|
2023-06-14 17:47:19 +00:00
|
|
|
<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
|
2023-05-13 13:38:36 +00:00
|
|
|
}
|
|
|
|
|
2023-06-19 08:23:56 +00:00
|
|
|
static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
|
2023-05-25 21:07:29 +00:00
|
|
|
GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
|
2023-07-05 12:19:42 +00:00
|
|
|
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
|
2023-11-03 11:13:09 +00:00
|
|
|
const dim3 block_nums(block_num_y, 1, 1);
|
2023-07-05 12:19:42 +00:00
|
|
|
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
|
2023-05-25 21:07:29 +00:00
|
|
|
dequantize_mul_mat_vec<QK8_0, QR8_0, dequantize_q8_0>
|
2023-06-14 17:47:19 +00:00
|
|
|
<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
|
2023-05-13 13:38:36 +00:00
|
|
|
}
|
|
|
|
|
2023-06-07 07:59:52 +00:00
|
|
|
static void dequantize_mul_mat_vec_q2_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
|
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_ASSERT(ncols % QK_K == 0);
|
2023-06-19 15:14:09 +00:00
|
|
|
const int ny = 2; // very slightly faster than 1 even when K_QUANTS_PER_ITERATION = 2
|
2023-06-14 17:47:19 +00:00
|
|
|
const int block_num_y = (nrows + ny - 1) / ny;
|
2023-11-03 11:13:09 +00:00
|
|
|
const dim3 block_nums(block_num_y, 1, 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
|
|
|
const dim3 block_dims(32, ny, 1);
|
2023-06-16 17:08:44 +00:00
|
|
|
dequantize_mul_mat_vec_q2_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
|
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
|
|
|
}
|
|
|
|
|
2023-06-07 07:59:52 +00:00
|
|
|
static void dequantize_mul_mat_vec_q3_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
|
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_ASSERT(ncols % QK_K == 0);
|
2023-06-19 15:14:09 +00:00
|
|
|
const int ny = 2 / K_QUANTS_PER_ITERATION;
|
|
|
|
const int block_num_y = (nrows + ny - 1) / ny;
|
2023-11-03 11:13:09 +00:00
|
|
|
const dim3 block_nums(block_num_y, 1, 1);
|
2023-06-19 15:14:09 +00:00
|
|
|
const dim3 block_dims(32, ny, 1);
|
|
|
|
dequantize_mul_mat_vec_q3_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
|
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
|
|
|
}
|
|
|
|
|
2023-06-07 07:59:52 +00:00
|
|
|
static void dequantize_mul_mat_vec_q4_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
|
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_ASSERT(ncols % QK_K == 0);
|
2023-06-19 15:14:09 +00:00
|
|
|
const int ny = 2 / K_QUANTS_PER_ITERATION;
|
|
|
|
const int block_num_y = (nrows + ny - 1) / ny;
|
2023-11-03 11:13:09 +00:00
|
|
|
const dim3 block_nums(block_num_y, 1, 1);
|
2023-06-19 15:14:09 +00:00
|
|
|
const dim3 block_dims(32, ny, 1);
|
|
|
|
dequantize_mul_mat_vec_q4_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
|
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
|
|
|
}
|
|
|
|
|
2023-06-07 07:59:52 +00:00
|
|
|
static void dequantize_mul_mat_vec_q5_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
|
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_ASSERT(ncols % QK_K == 0);
|
2023-06-16 17:08:44 +00:00
|
|
|
const dim3 block_dims(32, 1, 1);
|
|
|
|
dequantize_mul_mat_vec_q5_k<<<nrows, block_dims, 0, stream>>>(vx, y, dst, ncols);
|
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
|
|
|
}
|
|
|
|
|
2023-06-07 07:59:52 +00:00
|
|
|
static void dequantize_mul_mat_vec_q6_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
|
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_ASSERT(ncols % QK_K == 0);
|
2023-06-16 17:08:44 +00:00
|
|
|
const int ny = 2 / K_QUANTS_PER_ITERATION;
|
2023-06-14 17:47:19 +00:00
|
|
|
const int block_num_y = (nrows + ny - 1) / ny;
|
2023-11-03 11:13:09 +00:00
|
|
|
const dim3 block_nums(block_num_y, 1, 1);
|
2023-06-14 17:47:19 +00:00
|
|
|
const dim3 block_dims(32, ny, 1);
|
2023-06-16 17:08:44 +00:00
|
|
|
dequantize_mul_mat_vec_q6_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
|
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
|
|
|
}
|
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
static void convert_mul_mat_vec_f16_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
|
|
|
|
GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
|
|
|
|
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
|
|
|
|
const dim3 block_nums(block_num_y, 1, 1);
|
|
|
|
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
|
|
|
|
dequantize_mul_mat_vec<1, 1, convert_f16>
|
|
|
|
<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
|
|
|
|
}
|
|
|
|
|
2024-02-06 13:44:06 +00:00
|
|
|
template <int qk, int qi, typename block_q_t, int vdr, vec_dot_q_cuda_t vec_dot>
|
|
|
|
static void mul_mat_vec_q_cuda(
|
|
|
|
const void * vx, const void * vy, float * dst,
|
2024-02-07 11:40:26 +00:00
|
|
|
const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
|
2024-01-08 15:02:32 +00:00
|
|
|
|
2024-02-06 13:44:06 +00:00
|
|
|
GGML_ASSERT(ncols_x % qk == 0);
|
2024-02-11 18:08:39 +00:00
|
|
|
GGML_ASSERT(ncols_y <= MMVQ_MAX_BATCH_SIZE);
|
2024-01-11 19:39:39 +00:00
|
|
|
|
2024-02-08 20:56:40 +00:00
|
|
|
int id;
|
|
|
|
CUDA_CHECK(cudaGetDevice(&id));
|
|
|
|
|
2024-02-11 18:08:39 +00:00
|
|
|
int64_t nwarps = 1;
|
|
|
|
int64_t rows_per_cuda_block = 1;
|
2024-02-08 20:56:40 +00:00
|
|
|
|
2024-02-11 18:08:39 +00:00
|
|
|
if (g_device_caps[id].cc < CC_RDNA2) { // NVIDIA and AMD older than RDNA2
|
|
|
|
switch(ncols_y) {
|
2024-02-08 20:56:40 +00:00
|
|
|
case 1:
|
2024-02-11 18:08:39 +00:00
|
|
|
nwarps = 4;
|
|
|
|
rows_per_cuda_block = 1;
|
2024-02-08 20:56:40 +00:00
|
|
|
break;
|
|
|
|
case 2:
|
|
|
|
case 3:
|
|
|
|
case 4:
|
2024-02-11 18:08:39 +00:00
|
|
|
nwarps = 4;
|
|
|
|
rows_per_cuda_block = 2;
|
2024-02-08 20:56:40 +00:00
|
|
|
break;
|
2024-02-11 18:08:39 +00:00
|
|
|
case 5:
|
|
|
|
case 6:
|
|
|
|
case 7:
|
|
|
|
case 8:
|
|
|
|
nwarps = 2;
|
|
|
|
rows_per_cuda_block = 2;
|
2024-02-08 20:56:40 +00:00
|
|
|
break;
|
|
|
|
default:
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
break;
|
2024-02-11 18:08:39 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
const int64_t nblocks = (nrows_x + rows_per_cuda_block - 1) / rows_per_cuda_block;
|
|
|
|
const dim3 block_nums(nblocks, 1, 1);
|
|
|
|
const dim3 block_dims(WARP_SIZE, nwarps, 1);
|
2024-02-08 20:56:40 +00:00
|
|
|
|
2024-02-11 18:08:39 +00:00
|
|
|
switch (ncols_y) {
|
|
|
|
case 1:
|
|
|
|
mul_mat_vec_q<1, qk, qi, block_q_t, vdr, vec_dot>
|
|
|
|
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
|
|
|
|
break;
|
|
|
|
case 2:
|
|
|
|
mul_mat_vec_q<2, qk, qi, block_q_t, vdr, vec_dot>
|
|
|
|
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
|
|
|
|
break;
|
|
|
|
case 3:
|
|
|
|
mul_mat_vec_q<3, qk, qi, block_q_t, vdr, vec_dot>
|
|
|
|
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
|
|
|
|
break;
|
|
|
|
case 4:
|
|
|
|
mul_mat_vec_q<4, qk, qi, block_q_t, vdr, vec_dot>
|
|
|
|
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
|
|
|
|
break;
|
|
|
|
case 5:
|
|
|
|
mul_mat_vec_q<5, qk, qi, block_q_t, vdr, vec_dot>
|
|
|
|
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
|
|
|
|
break;
|
|
|
|
case 6:
|
|
|
|
mul_mat_vec_q<6, qk, qi, block_q_t, vdr, vec_dot>
|
|
|
|
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
|
|
|
|
break;
|
|
|
|
case 7:
|
|
|
|
mul_mat_vec_q<7, qk, qi, block_q_t, vdr, vec_dot>
|
|
|
|
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
|
|
|
|
break;
|
|
|
|
case 8:
|
|
|
|
mul_mat_vec_q<8, qk, qi, block_q_t, vdr, vec_dot>
|
|
|
|
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
|
|
|
|
break;
|
2024-02-06 13:44:06 +00:00
|
|
|
default:
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
break;
|
|
|
|
}
|
2024-01-30 13:14:12 +00:00
|
|
|
}
|
|
|
|
|
2023-07-29 21:04:44 +00:00
|
|
|
static void ggml_mul_mat_q4_0_q8_1_cuda(
|
|
|
|
const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
|
|
|
|
const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
int id;
|
|
|
|
CUDA_CHECK(cudaGetDevice(&id));
|
2023-12-24 13:34:22 +00:00
|
|
|
const int compute_capability = g_device_caps[id].cc;
|
2023-08-09 07:42:34 +00:00
|
|
|
|
2023-08-12 22:24:45 +00:00
|
|
|
int mmq_x, mmq_y, nwarps;
|
2023-09-13 09:20:24 +00:00
|
|
|
if (compute_capability >= CC_RDNA2) {
|
|
|
|
mmq_x = MMQ_X_Q4_0_RDNA2;
|
|
|
|
mmq_y = MMQ_Y_Q4_0_RDNA2;
|
|
|
|
nwarps = NWARPS_Q4_0_RDNA2;
|
|
|
|
} else if (compute_capability >= CC_OFFSET_AMD) {
|
|
|
|
mmq_x = MMQ_X_Q4_0_RDNA1;
|
|
|
|
mmq_y = MMQ_Y_Q4_0_RDNA1;
|
|
|
|
nwarps = NWARPS_Q4_0_RDNA1;
|
2023-09-30 16:12:57 +00:00
|
|
|
} else if (compute_capability >= CC_VOLTA) {
|
2023-08-12 22:24:45 +00:00
|
|
|
mmq_x = MMQ_X_Q4_0_AMPERE;
|
|
|
|
mmq_y = MMQ_Y_Q4_0_AMPERE;
|
|
|
|
nwarps = NWARPS_Q4_0_AMPERE;
|
|
|
|
} else if (compute_capability >= MIN_CC_DP4A) {
|
|
|
|
mmq_x = MMQ_X_Q4_0_PASCAL;
|
|
|
|
mmq_y = MMQ_Y_Q4_0_PASCAL;
|
|
|
|
nwarps = NWARPS_Q4_0_PASCAL;
|
2023-08-02 14:48:10 +00:00
|
|
|
} else {
|
2023-08-12 22:24:45 +00:00
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
|
|
|
|
const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
|
|
|
|
const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
|
|
|
|
const dim3 block_nums(block_num_x, block_num_y, 1);
|
|
|
|
const dim3 block_dims(WARP_SIZE, nwarps, 1);
|
|
|
|
|
|
|
|
if (nrows_x % mmq_y == 0) {
|
|
|
|
const bool need_check = false;
|
|
|
|
mul_mat_q4_0<need_check><<<block_nums, block_dims, 0, stream>>>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
} else {
|
|
|
|
const bool need_check = true;
|
|
|
|
mul_mat_q4_0<need_check><<<block_nums, block_dims, 0, stream>>>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
2023-08-02 14:48:10 +00:00
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_mul_mat_q4_1_q8_1_cuda(
|
|
|
|
const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
|
|
|
|
const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
int id;
|
|
|
|
CUDA_CHECK(cudaGetDevice(&id));
|
2023-12-24 13:34:22 +00:00
|
|
|
const int compute_capability = g_device_caps[id].cc;
|
2023-08-09 07:42:34 +00:00
|
|
|
|
2023-08-12 22:24:45 +00:00
|
|
|
int mmq_x, mmq_y, nwarps;
|
2023-09-13 09:20:24 +00:00
|
|
|
if (compute_capability >= CC_RDNA2) {
|
|
|
|
mmq_x = MMQ_X_Q4_1_RDNA2;
|
|
|
|
mmq_y = MMQ_Y_Q4_1_RDNA2;
|
|
|
|
nwarps = NWARPS_Q4_1_RDNA2;
|
|
|
|
} else if (compute_capability >= CC_OFFSET_AMD) {
|
|
|
|
mmq_x = MMQ_X_Q4_1_RDNA1;
|
|
|
|
mmq_y = MMQ_Y_Q4_1_RDNA1;
|
|
|
|
nwarps = NWARPS_Q4_1_RDNA1;
|
2023-09-30 16:12:57 +00:00
|
|
|
} else if (compute_capability >= CC_VOLTA) {
|
2023-08-12 22:24:45 +00:00
|
|
|
mmq_x = MMQ_X_Q4_1_AMPERE;
|
|
|
|
mmq_y = MMQ_Y_Q4_1_AMPERE;
|
|
|
|
nwarps = NWARPS_Q4_1_AMPERE;
|
|
|
|
} else if (compute_capability >= MIN_CC_DP4A) {
|
|
|
|
mmq_x = MMQ_X_Q4_1_PASCAL;
|
|
|
|
mmq_y = MMQ_Y_Q4_1_PASCAL;
|
|
|
|
nwarps = NWARPS_Q4_1_PASCAL;
|
2023-08-02 14:48:10 +00:00
|
|
|
} else {
|
2023-08-12 22:24:45 +00:00
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
|
|
|
|
const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
|
|
|
|
const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
|
|
|
|
const dim3 block_nums(block_num_x, block_num_y, 1);
|
|
|
|
const dim3 block_dims(WARP_SIZE, nwarps, 1);
|
2023-08-09 07:42:34 +00:00
|
|
|
|
2023-08-12 22:24:45 +00:00
|
|
|
if (nrows_x % mmq_y == 0) {
|
|
|
|
const bool need_check = false;
|
|
|
|
mul_mat_q4_1<need_check><<<block_nums, block_dims, 0, stream>>>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
} else {
|
|
|
|
const bool need_check = true;
|
|
|
|
mul_mat_q4_1<need_check><<<block_nums, block_dims, 0, stream>>>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
2023-08-02 14:48:10 +00:00
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_mul_mat_q5_0_q8_1_cuda(
|
|
|
|
const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
|
|
|
|
const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
int id;
|
|
|
|
CUDA_CHECK(cudaGetDevice(&id));
|
2023-12-24 13:34:22 +00:00
|
|
|
const int compute_capability = g_device_caps[id].cc;
|
2023-08-09 07:42:34 +00:00
|
|
|
|
2023-08-12 22:24:45 +00:00
|
|
|
int mmq_x, mmq_y, nwarps;
|
2023-09-13 09:20:24 +00:00
|
|
|
if (compute_capability >= CC_RDNA2) {
|
|
|
|
mmq_x = MMQ_X_Q5_0_RDNA2;
|
|
|
|
mmq_y = MMQ_Y_Q5_0_RDNA2;
|
|
|
|
nwarps = NWARPS_Q5_0_RDNA2;
|
|
|
|
} else if (compute_capability >= CC_OFFSET_AMD) {
|
|
|
|
mmq_x = MMQ_X_Q5_0_RDNA1;
|
|
|
|
mmq_y = MMQ_Y_Q5_0_RDNA1;
|
|
|
|
nwarps = NWARPS_Q5_0_RDNA1;
|
2023-09-30 16:12:57 +00:00
|
|
|
} else if (compute_capability >= CC_VOLTA) {
|
2023-08-12 22:24:45 +00:00
|
|
|
mmq_x = MMQ_X_Q5_0_AMPERE;
|
|
|
|
mmq_y = MMQ_Y_Q5_0_AMPERE;
|
|
|
|
nwarps = NWARPS_Q5_0_AMPERE;
|
|
|
|
} else if (compute_capability >= MIN_CC_DP4A) {
|
|
|
|
mmq_x = MMQ_X_Q5_0_PASCAL;
|
|
|
|
mmq_y = MMQ_Y_Q5_0_PASCAL;
|
|
|
|
nwarps = NWARPS_Q5_0_PASCAL;
|
2023-08-02 14:48:10 +00:00
|
|
|
} else {
|
2023-08-12 22:24:45 +00:00
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
|
|
|
|
const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
|
|
|
|
const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
|
|
|
|
const dim3 block_nums(block_num_x, block_num_y, 1);
|
|
|
|
const dim3 block_dims(WARP_SIZE, nwarps, 1);
|
|
|
|
|
|
|
|
if (nrows_x % mmq_y == 0) {
|
|
|
|
const bool need_check = false;
|
|
|
|
mul_mat_q5_0<need_check><<<block_nums, block_dims, 0, stream>>>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
} else {
|
|
|
|
const bool need_check = true;
|
|
|
|
mul_mat_q5_0<need_check><<<block_nums, block_dims, 0, stream>>>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
2023-08-02 14:48:10 +00:00
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_mul_mat_q5_1_q8_1_cuda(
|
|
|
|
const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
|
|
|
|
const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
int id;
|
|
|
|
CUDA_CHECK(cudaGetDevice(&id));
|
2023-12-24 13:34:22 +00:00
|
|
|
const int compute_capability = g_device_caps[id].cc;
|
2023-08-09 07:42:34 +00:00
|
|
|
|
2023-08-12 22:24:45 +00:00
|
|
|
int mmq_x, mmq_y, nwarps;
|
2023-09-13 09:20:24 +00:00
|
|
|
if (compute_capability >= CC_RDNA2) {
|
|
|
|
mmq_x = MMQ_X_Q5_1_RDNA2;
|
|
|
|
mmq_y = MMQ_Y_Q5_1_RDNA2;
|
|
|
|
nwarps = NWARPS_Q5_1_RDNA2;
|
|
|
|
} else if (compute_capability >= CC_OFFSET_AMD) {
|
|
|
|
mmq_x = MMQ_X_Q5_1_RDNA1;
|
|
|
|
mmq_y = MMQ_Y_Q5_1_RDNA1;
|
|
|
|
nwarps = NWARPS_Q5_1_RDNA1;
|
2023-09-30 16:12:57 +00:00
|
|
|
} else if (compute_capability >= CC_VOLTA) {
|
2023-08-12 22:24:45 +00:00
|
|
|
mmq_x = MMQ_X_Q5_1_AMPERE;
|
|
|
|
mmq_y = MMQ_Y_Q5_1_AMPERE;
|
|
|
|
nwarps = NWARPS_Q5_1_AMPERE;
|
|
|
|
} else if (compute_capability >= MIN_CC_DP4A) {
|
|
|
|
mmq_x = MMQ_X_Q5_1_PASCAL;
|
|
|
|
mmq_y = MMQ_Y_Q5_1_PASCAL;
|
|
|
|
nwarps = NWARPS_Q5_1_PASCAL;
|
2023-08-02 14:48:10 +00:00
|
|
|
} else {
|
2023-08-12 22:24:45 +00:00
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
|
|
|
|
const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
|
|
|
|
const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
|
|
|
|
const dim3 block_nums(block_num_x, block_num_y, 1);
|
|
|
|
const dim3 block_dims(WARP_SIZE, nwarps, 1);
|
|
|
|
|
|
|
|
if (nrows_x % mmq_y == 0) {
|
|
|
|
const bool need_check = false;
|
|
|
|
mul_mat_q5_1<need_check><<<block_nums, block_dims, 0, stream>>>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
} else {
|
|
|
|
const bool need_check = true;
|
|
|
|
mul_mat_q5_1<need_check><<<block_nums, block_dims, 0, stream>>>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
2023-08-02 14:48:10 +00:00
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_mul_mat_q8_0_q8_1_cuda(
|
|
|
|
const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
|
|
|
|
const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
int id;
|
|
|
|
CUDA_CHECK(cudaGetDevice(&id));
|
2023-12-24 13:34:22 +00:00
|
|
|
const int compute_capability = g_device_caps[id].cc;
|
2023-08-09 07:42:34 +00:00
|
|
|
|
2023-08-12 22:24:45 +00:00
|
|
|
int mmq_x, mmq_y, nwarps;
|
2023-09-13 09:20:24 +00:00
|
|
|
if (compute_capability >= CC_RDNA2) {
|
|
|
|
mmq_x = MMQ_X_Q8_0_RDNA2;
|
|
|
|
mmq_y = MMQ_Y_Q8_0_RDNA2;
|
|
|
|
nwarps = NWARPS_Q8_0_RDNA2;
|
|
|
|
} else if (compute_capability >= CC_OFFSET_AMD) {
|
|
|
|
mmq_x = MMQ_X_Q8_0_RDNA1;
|
|
|
|
mmq_y = MMQ_Y_Q8_0_RDNA1;
|
|
|
|
nwarps = NWARPS_Q8_0_RDNA1;
|
2023-09-30 16:12:57 +00:00
|
|
|
} else if (compute_capability >= CC_VOLTA) {
|
2023-08-12 22:24:45 +00:00
|
|
|
mmq_x = MMQ_X_Q8_0_AMPERE;
|
|
|
|
mmq_y = MMQ_Y_Q8_0_AMPERE;
|
|
|
|
nwarps = NWARPS_Q8_0_AMPERE;
|
|
|
|
} else if (compute_capability >= MIN_CC_DP4A) {
|
|
|
|
mmq_x = MMQ_X_Q8_0_PASCAL;
|
|
|
|
mmq_y = MMQ_Y_Q8_0_PASCAL;
|
|
|
|
nwarps = NWARPS_Q8_0_PASCAL;
|
2023-08-02 14:48:10 +00:00
|
|
|
} else {
|
2023-08-12 22:24:45 +00:00
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
|
|
|
|
const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
|
|
|
|
const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
|
|
|
|
const dim3 block_nums(block_num_x, block_num_y, 1);
|
|
|
|
const dim3 block_dims(WARP_SIZE, nwarps, 1);
|
|
|
|
|
|
|
|
if (nrows_x % mmq_y == 0) {
|
|
|
|
const bool need_check = false;
|
|
|
|
mul_mat_q8_0<need_check><<<block_nums, block_dims, 0, stream>>>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
} else {
|
|
|
|
const bool need_check = true;
|
|
|
|
mul_mat_q8_0<need_check><<<block_nums, block_dims, 0, stream>>>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
2023-08-02 14:48:10 +00:00
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_mul_mat_q2_K_q8_1_cuda(
|
|
|
|
const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
|
|
|
|
const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
int id;
|
|
|
|
CUDA_CHECK(cudaGetDevice(&id));
|
2023-12-24 13:34:22 +00:00
|
|
|
const int compute_capability = g_device_caps[id].cc;
|
2023-08-09 07:42:34 +00:00
|
|
|
|
2023-08-12 22:24:45 +00:00
|
|
|
int mmq_x, mmq_y, nwarps;
|
2023-09-13 09:20:24 +00:00
|
|
|
if (compute_capability >= CC_RDNA2) {
|
|
|
|
mmq_x = MMQ_X_Q2_K_RDNA2;
|
|
|
|
mmq_y = MMQ_Y_Q2_K_RDNA2;
|
|
|
|
nwarps = NWARPS_Q2_K_RDNA2;
|
|
|
|
} else if (compute_capability >= CC_OFFSET_AMD) {
|
|
|
|
mmq_x = MMQ_X_Q2_K_RDNA1;
|
|
|
|
mmq_y = MMQ_Y_Q2_K_RDNA1;
|
|
|
|
nwarps = NWARPS_Q2_K_RDNA1;
|
2023-09-30 16:12:57 +00:00
|
|
|
} else if (compute_capability >= CC_VOLTA) {
|
2023-08-12 22:24:45 +00:00
|
|
|
mmq_x = MMQ_X_Q2_K_AMPERE;
|
|
|
|
mmq_y = MMQ_Y_Q2_K_AMPERE;
|
|
|
|
nwarps = NWARPS_Q2_K_AMPERE;
|
|
|
|
} else if (compute_capability >= MIN_CC_DP4A) {
|
|
|
|
mmq_x = MMQ_X_Q2_K_PASCAL;
|
|
|
|
mmq_y = MMQ_Y_Q2_K_PASCAL;
|
|
|
|
nwarps = NWARPS_Q2_K_PASCAL;
|
2023-08-02 14:48:10 +00:00
|
|
|
} else {
|
2023-08-12 22:24:45 +00:00
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
|
|
|
|
const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
|
|
|
|
const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
|
|
|
|
const dim3 block_nums(block_num_x, block_num_y, 1);
|
|
|
|
const dim3 block_dims(WARP_SIZE, nwarps, 1);
|
|
|
|
|
|
|
|
if (nrows_x % mmq_y == 0) {
|
|
|
|
const bool need_check = false;
|
|
|
|
mul_mat_q2_K<need_check><<<block_nums, block_dims, 0, stream>>>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
} else {
|
|
|
|
const bool need_check = true;
|
|
|
|
mul_mat_q2_K<need_check><<<block_nums, block_dims, 0, stream>>>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
2023-08-02 14:48:10 +00:00
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_mul_mat_q3_K_q8_1_cuda(
|
|
|
|
const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
|
|
|
|
const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
|
|
|
|
|
2023-08-27 12:19:59 +00:00
|
|
|
#if QK_K == 256
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
int id;
|
|
|
|
CUDA_CHECK(cudaGetDevice(&id));
|
2023-12-24 13:34:22 +00:00
|
|
|
const int compute_capability = g_device_caps[id].cc;
|
2023-08-09 07:42:34 +00:00
|
|
|
|
2023-08-12 22:24:45 +00:00
|
|
|
int mmq_x, mmq_y, nwarps;
|
2023-09-13 09:20:24 +00:00
|
|
|
if (compute_capability >= CC_RDNA2) {
|
|
|
|
mmq_x = MMQ_X_Q3_K_RDNA2;
|
|
|
|
mmq_y = MMQ_Y_Q3_K_RDNA2;
|
|
|
|
nwarps = NWARPS_Q3_K_RDNA2;
|
|
|
|
} else if (compute_capability >= CC_OFFSET_AMD) {
|
|
|
|
mmq_x = MMQ_X_Q3_K_RDNA1;
|
|
|
|
mmq_y = MMQ_Y_Q3_K_RDNA1;
|
|
|
|
nwarps = NWARPS_Q3_K_RDNA1;
|
2023-09-30 16:12:57 +00:00
|
|
|
} else if (compute_capability >= CC_VOLTA) {
|
2023-08-12 22:24:45 +00:00
|
|
|
mmq_x = MMQ_X_Q3_K_AMPERE;
|
|
|
|
mmq_y = MMQ_Y_Q3_K_AMPERE;
|
|
|
|
nwarps = NWARPS_Q3_K_AMPERE;
|
|
|
|
} else if (compute_capability >= MIN_CC_DP4A) {
|
|
|
|
mmq_x = MMQ_X_Q3_K_PASCAL;
|
|
|
|
mmq_y = MMQ_Y_Q3_K_PASCAL;
|
|
|
|
nwarps = NWARPS_Q3_K_PASCAL;
|
2023-08-02 14:48:10 +00:00
|
|
|
} else {
|
2023-08-12 22:24:45 +00:00
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
|
|
|
|
const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
|
|
|
|
const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
|
|
|
|
const dim3 block_nums(block_num_x, block_num_y, 1);
|
|
|
|
const dim3 block_dims(WARP_SIZE, nwarps, 1);
|
|
|
|
|
|
|
|
if (nrows_x % mmq_y == 0) {
|
|
|
|
const bool need_check = false;
|
|
|
|
mul_mat_q3_K<need_check><<<block_nums, block_dims, 0, stream>>>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
} else {
|
|
|
|
const bool need_check = true;
|
|
|
|
mul_mat_q3_K<need_check><<<block_nums, block_dims, 0, stream>>>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
2023-08-02 14:48:10 +00:00
|
|
|
}
|
2023-08-27 12:19:59 +00:00
|
|
|
#endif
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_mul_mat_q4_K_q8_1_cuda(
|
|
|
|
const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
|
|
|
|
const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
int id;
|
|
|
|
CUDA_CHECK(cudaGetDevice(&id));
|
2023-12-24 13:34:22 +00:00
|
|
|
const int compute_capability = g_device_caps[id].cc;
|
2023-08-09 07:42:34 +00:00
|
|
|
|
2023-08-12 22:24:45 +00:00
|
|
|
int mmq_x, mmq_y, nwarps;
|
2023-09-13 09:20:24 +00:00
|
|
|
if (compute_capability >= CC_RDNA2) {
|
|
|
|
mmq_x = MMQ_X_Q4_K_RDNA2;
|
|
|
|
mmq_y = MMQ_Y_Q4_K_RDNA2;
|
|
|
|
nwarps = NWARPS_Q4_K_RDNA2;
|
|
|
|
} else if (compute_capability >= CC_OFFSET_AMD) {
|
|
|
|
mmq_x = MMQ_X_Q4_K_RDNA1;
|
|
|
|
mmq_y = MMQ_Y_Q4_K_RDNA1;
|
|
|
|
nwarps = NWARPS_Q4_K_RDNA1;
|
2023-09-30 16:12:57 +00:00
|
|
|
} else if (compute_capability >= CC_VOLTA) {
|
2023-08-12 22:24:45 +00:00
|
|
|
mmq_x = MMQ_X_Q4_K_AMPERE;
|
|
|
|
mmq_y = MMQ_Y_Q4_K_AMPERE;
|
|
|
|
nwarps = NWARPS_Q4_K_AMPERE;
|
|
|
|
} else if (compute_capability >= MIN_CC_DP4A) {
|
|
|
|
mmq_x = MMQ_X_Q4_K_PASCAL;
|
|
|
|
mmq_y = MMQ_Y_Q4_K_PASCAL;
|
|
|
|
nwarps = NWARPS_Q4_K_PASCAL;
|
2023-08-02 14:48:10 +00:00
|
|
|
} else {
|
2023-08-12 22:24:45 +00:00
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
|
|
|
|
const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
|
|
|
|
const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
|
|
|
|
const dim3 block_nums(block_num_x, block_num_y, 1);
|
|
|
|
const dim3 block_dims(WARP_SIZE, nwarps, 1);
|
|
|
|
|
|
|
|
if (nrows_x % mmq_y == 0) {
|
|
|
|
const bool need_check = false;
|
|
|
|
mul_mat_q4_K<need_check><<<block_nums, block_dims, 0, stream>>>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
} else {
|
|
|
|
const bool need_check = true;
|
|
|
|
mul_mat_q4_K<need_check><<<block_nums, block_dims, 0, stream>>>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
2023-08-02 14:48:10 +00:00
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_mul_mat_q5_K_q8_1_cuda(
|
|
|
|
const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
|
|
|
|
const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
int id;
|
|
|
|
CUDA_CHECK(cudaGetDevice(&id));
|
2023-12-24 13:34:22 +00:00
|
|
|
const int compute_capability = g_device_caps[id].cc;
|
2023-08-09 07:42:34 +00:00
|
|
|
|
2023-08-12 22:24:45 +00:00
|
|
|
int mmq_x, mmq_y, nwarps;
|
2023-09-13 09:20:24 +00:00
|
|
|
if (compute_capability >= CC_RDNA2) {
|
|
|
|
mmq_x = MMQ_X_Q5_K_RDNA2;
|
|
|
|
mmq_y = MMQ_Y_Q5_K_RDNA2;
|
|
|
|
nwarps = NWARPS_Q5_K_RDNA2;
|
|
|
|
} else if (compute_capability >= CC_OFFSET_AMD) {
|
|
|
|
mmq_x = MMQ_X_Q5_K_RDNA1;
|
|
|
|
mmq_y = MMQ_Y_Q5_K_RDNA1;
|
|
|
|
nwarps = NWARPS_Q5_K_RDNA1;
|
2023-09-30 16:12:57 +00:00
|
|
|
} else if (compute_capability >= CC_VOLTA) {
|
2023-08-12 22:24:45 +00:00
|
|
|
mmq_x = MMQ_X_Q5_K_AMPERE;
|
|
|
|
mmq_y = MMQ_Y_Q5_K_AMPERE;
|
|
|
|
nwarps = NWARPS_Q5_K_AMPERE;
|
|
|
|
} else if (compute_capability >= MIN_CC_DP4A) {
|
|
|
|
mmq_x = MMQ_X_Q5_K_PASCAL;
|
|
|
|
mmq_y = MMQ_Y_Q5_K_PASCAL;
|
|
|
|
nwarps = NWARPS_Q5_K_PASCAL;
|
2023-08-02 14:48:10 +00:00
|
|
|
} else {
|
2023-08-12 22:24:45 +00:00
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
|
|
|
|
const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
|
|
|
|
const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
|
|
|
|
const dim3 block_nums(block_num_x, block_num_y, 1);
|
|
|
|
const dim3 block_dims(WARP_SIZE, nwarps, 1);
|
|
|
|
|
|
|
|
if (nrows_x % mmq_y == 0) {
|
|
|
|
const bool need_check = false;
|
|
|
|
mul_mat_q5_K<need_check><<<block_nums, block_dims, 0, stream>>>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
} else {
|
|
|
|
const bool need_check = true;
|
|
|
|
mul_mat_q5_K<need_check><<<block_nums, block_dims, 0, stream>>>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
2023-08-02 14:48:10 +00:00
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_mul_mat_q6_K_q8_1_cuda(
|
|
|
|
const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
|
|
|
|
const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
|
|
|
|
|
2023-08-09 07:42:34 +00:00
|
|
|
int id;
|
|
|
|
CUDA_CHECK(cudaGetDevice(&id));
|
2023-12-24 13:34:22 +00:00
|
|
|
const int compute_capability = g_device_caps[id].cc;
|
2023-08-09 07:42:34 +00:00
|
|
|
|
2023-08-12 22:24:45 +00:00
|
|
|
int mmq_x, mmq_y, nwarps;
|
2023-09-13 09:20:24 +00:00
|
|
|
if (compute_capability >= CC_RDNA2) {
|
|
|
|
mmq_x = MMQ_X_Q6_K_RDNA2;
|
|
|
|
mmq_y = MMQ_Y_Q6_K_RDNA2;
|
|
|
|
nwarps = NWARPS_Q6_K_RDNA2;
|
|
|
|
} else if (compute_capability >= CC_OFFSET_AMD) {
|
|
|
|
mmq_x = MMQ_X_Q6_K_RDNA1;
|
|
|
|
mmq_y = MMQ_Y_Q6_K_RDNA1;
|
|
|
|
nwarps = NWARPS_Q6_K_RDNA1;
|
2023-09-30 16:12:57 +00:00
|
|
|
} else if (compute_capability >= CC_VOLTA) {
|
2023-08-12 22:24:45 +00:00
|
|
|
mmq_x = MMQ_X_Q6_K_AMPERE;
|
|
|
|
mmq_y = MMQ_Y_Q6_K_AMPERE;
|
|
|
|
nwarps = NWARPS_Q6_K_AMPERE;
|
|
|
|
} else if (compute_capability >= MIN_CC_DP4A) {
|
|
|
|
mmq_x = MMQ_X_Q6_K_PASCAL;
|
|
|
|
mmq_y = MMQ_Y_Q6_K_PASCAL;
|
|
|
|
nwarps = NWARPS_Q6_K_PASCAL;
|
2023-08-02 14:48:10 +00:00
|
|
|
} else {
|
2023-08-12 22:24:45 +00:00
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
|
|
|
|
const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
|
|
|
|
const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
|
|
|
|
const dim3 block_nums(block_num_x, block_num_y, 1);
|
|
|
|
const dim3 block_dims(WARP_SIZE, nwarps, 1);
|
|
|
|
|
|
|
|
if (nrows_x % mmq_y == 0) {
|
|
|
|
const bool need_check = false;
|
|
|
|
mul_mat_q6_K<need_check><<<block_nums, block_dims, 0, stream>>>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
|
|
|
} else {
|
|
|
|
const bool need_check = true;
|
|
|
|
mul_mat_q6_K<need_check><<<block_nums, block_dims, 0, stream>>>
|
|
|
|
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
2023-08-02 14:48:10 +00:00
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
2023-07-23 12:09:47 +00:00
|
|
|
static void ggml_mul_mat_p021_f16_f32_cuda(
|
|
|
|
const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x,
|
|
|
|
const int nchannels_x, const int nchannels_y, cudaStream_t stream) {
|
|
|
|
|
|
|
|
const dim3 block_nums(1, nrows_x, nchannels_y);
|
2023-06-14 17:47:19 +00:00
|
|
|
const dim3 block_dims(WARP_SIZE, 1, 1);
|
2023-07-23 12:09:47 +00:00
|
|
|
mul_mat_p021_f16_f32<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols_x, nrows_x, nchannels_x, nchannels_y);
|
2023-06-14 17:47:19 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_mul_mat_vec_nc_f16_f32_cuda(
|
|
|
|
const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x, const int row_stride_x,
|
2023-07-23 12:09:47 +00:00
|
|
|
const int nchannels_x, const int nchannels_y, const int channel_stride_x, cudaStream_t stream) {
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2023-07-23 12:09:47 +00:00
|
|
|
const dim3 block_nums(1, nrows_x, nchannels_y);
|
2023-06-14 17:47:19 +00:00
|
|
|
const dim3 block_dims(WARP_SIZE, 1, 1);
|
|
|
|
mul_mat_vec_nc_f16_f32<<<block_nums, block_dims, 0, stream>>>
|
2023-07-23 12:09:47 +00:00
|
|
|
(vx, y, dst, ncols_x, nrows_x, row_stride_x, channel_stride_x, nchannels_y/nchannels_x);
|
2023-06-14 17:47:19 +00:00
|
|
|
}
|
|
|
|
|
2024-01-29 12:37:33 +00:00
|
|
|
|
|
|
|
static void ggml_cpy_f16_f32_cuda(
|
|
|
|
const char * cx, char * cdst, const int ne,
|
|
|
|
const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
|
|
|
|
const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
|
|
|
|
|
|
|
|
const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
|
|
|
|
cpy_f32_f16<cpy_1_f16_f32><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
|
|
|
|
(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
|
|
|
|
}
|
|
|
|
|
2023-06-14 17:47:19 +00:00
|
|
|
static void ggml_cpy_f32_f32_cuda(
|
|
|
|
const char * cx, char * cdst, const int ne,
|
2024-01-29 12:37:33 +00:00
|
|
|
const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
|
|
|
|
const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
|
2023-06-14 17:47:19 +00:00
|
|
|
|
|
|
|
const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
|
|
|
|
cpy_f32_f16<cpy_1_f32_f32><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
|
2024-01-29 12:37:33 +00:00
|
|
|
(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
|
2023-06-14 17:47:19 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_cpy_f32_f16_cuda(
|
|
|
|
const char * cx, char * cdst, const int ne,
|
2024-01-29 12:37:33 +00:00
|
|
|
const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
|
|
|
|
const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
|
2023-06-14 17:47:19 +00:00
|
|
|
|
|
|
|
const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
|
|
|
|
cpy_f32_f16<cpy_1_f32_f16><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
|
2024-01-29 12:37:33 +00:00
|
|
|
(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
|
2023-06-14 17:47:19 +00:00
|
|
|
}
|
|
|
|
|
2023-12-07 11:03:17 +00:00
|
|
|
static void ggml_cpy_f32_q8_0_cuda(
|
|
|
|
const char * cx, char * cdst, const int ne,
|
2024-01-29 12:37:33 +00:00
|
|
|
const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
|
|
|
|
const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
|
2023-12-07 11:03:17 +00:00
|
|
|
|
|
|
|
GGML_ASSERT(ne % QK8_0 == 0);
|
|
|
|
const int num_blocks = ne / QK8_0;
|
|
|
|
cpy_f32_q<cpy_blck_f32_q8_0, QK8_0><<<num_blocks, 1, 0, stream>>>
|
2024-01-29 12:37:33 +00:00
|
|
|
(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
|
2023-12-07 11:03:17 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_cpy_f32_q4_0_cuda(
|
|
|
|
const char * cx, char * cdst, const int ne,
|
2024-01-29 12:37:33 +00:00
|
|
|
const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
|
|
|
|
const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
|
2023-12-07 11:03:17 +00:00
|
|
|
|
|
|
|
GGML_ASSERT(ne % QK4_0 == 0);
|
|
|
|
const int num_blocks = ne / QK4_0;
|
|
|
|
cpy_f32_q<cpy_blck_f32_q4_0, QK4_0><<<num_blocks, 1, 0, stream>>>
|
2024-01-29 12:37:33 +00:00
|
|
|
(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
|
2023-12-07 11:03:17 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_cpy_f32_q4_1_cuda(
|
|
|
|
const char * cx, char * cdst, const int ne,
|
2024-01-29 12:37:33 +00:00
|
|
|
const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
|
|
|
|
const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
|
2023-12-07 11:03:17 +00:00
|
|
|
|
|
|
|
GGML_ASSERT(ne % QK4_1 == 0);
|
|
|
|
const int num_blocks = ne / QK4_1;
|
|
|
|
cpy_f32_q<cpy_blck_f32_q4_1, QK4_1><<<num_blocks, 1, 0, stream>>>
|
2024-01-29 12:37:33 +00:00
|
|
|
(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
|
2023-12-07 11:03:17 +00:00
|
|
|
}
|
|
|
|
|
2023-11-13 14:55:52 +00:00
|
|
|
static void ggml_cpy_f16_f16_cuda(
|
|
|
|
const char * cx, char * cdst, const int ne,
|
2024-01-29 12:37:33 +00:00
|
|
|
const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
|
|
|
|
const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
|
2023-11-13 14:55:52 +00:00
|
|
|
|
|
|
|
const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
|
|
|
|
cpy_f32_f16<cpy_1_f16_f16><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
|
2024-01-29 12:37:33 +00:00
|
|
|
(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
|
2023-11-13 14:55:52 +00:00
|
|
|
}
|
|
|
|
|
2024-01-29 12:37:33 +00:00
|
|
|
|
|
|
|
|
2023-06-14 17:47:19 +00:00
|
|
|
static void scale_f32_cuda(const float * x, float * dst, const float scale, const int k, cudaStream_t stream) {
|
|
|
|
const int num_blocks = (k + CUDA_SCALE_BLOCK_SIZE - 1) / CUDA_SCALE_BLOCK_SIZE;
|
|
|
|
scale_f32<<<num_blocks, CUDA_SCALE_BLOCK_SIZE, 0, stream>>>(x, dst, scale, k);
|
|
|
|
}
|
|
|
|
|
2023-10-10 07:50:23 +00:00
|
|
|
static void clamp_f32_cuda(const float * x, float * dst, const float min, const float max, const int k, cudaStream_t stream) {
|
|
|
|
const int num_blocks = (k + CUDA_CLAMP_BLOCK_SIZE - 1) / CUDA_CLAMP_BLOCK_SIZE;
|
|
|
|
clamp_f32<<<num_blocks, CUDA_CLAMP_BLOCK_SIZE, 0, stream>>>(x, dst, min, max, k);
|
|
|
|
}
|
|
|
|
|
2023-09-28 16:04:36 +00:00
|
|
|
template<typename T>
|
2023-11-01 22:04:33 +00:00
|
|
|
static void rope_cuda(
|
|
|
|
const T * x, T * dst, int ncols, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
|
|
|
|
float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream
|
|
|
|
) {
|
2023-08-28 11:23:55 +00:00
|
|
|
GGML_ASSERT(ncols % 2 == 0);
|
|
|
|
const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
|
2023-06-06 19:33:23 +00:00
|
|
|
const int num_blocks_x = (ncols + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
|
2023-08-22 13:25:19 +00:00
|
|
|
const dim3 block_nums(nrows, num_blocks_x, 1);
|
2023-09-28 16:04:36 +00:00
|
|
|
if (pos == nullptr) {
|
2023-11-01 22:04:33 +00:00
|
|
|
rope<T, false><<<block_nums, block_dims, 0, stream>>>(
|
|
|
|
x, dst, ncols, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims
|
|
|
|
);
|
2023-09-28 16:04:36 +00:00
|
|
|
} else {
|
2023-11-01 22:04:33 +00:00
|
|
|
rope<T, true><<<block_nums, block_dims, 0, stream>>>(
|
|
|
|
x, dst, ncols, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims
|
|
|
|
);
|
2023-09-28 16:04:36 +00:00
|
|
|
}
|
2023-05-13 13:38:36 +00:00
|
|
|
}
|
|
|
|
|
2023-09-28 16:04:36 +00:00
|
|
|
template<typename T>
|
2023-11-01 22:04:33 +00:00
|
|
|
static void rope_neox_cuda(
|
2023-11-24 17:04:31 +00:00
|
|
|
const T * x, T * dst, int ncols, int n_dims, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
|
2023-11-01 22:04:33 +00:00
|
|
|
float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream
|
|
|
|
) {
|
2023-08-28 11:23:55 +00:00
|
|
|
GGML_ASSERT(ncols % 2 == 0);
|
|
|
|
const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
|
2023-08-25 08:55:59 +00:00
|
|
|
const int num_blocks_x = (ncols + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
|
|
|
|
const dim3 block_nums(nrows, num_blocks_x, 1);
|
2023-11-24 17:04:31 +00:00
|
|
|
|
|
|
|
const float theta_scale = powf(freq_base, -2.0f/n_dims);
|
|
|
|
const float inv_ndims = -1.0f / n_dims;
|
|
|
|
|
2023-09-28 16:04:36 +00:00
|
|
|
if (pos == nullptr) {
|
2023-11-01 22:04:33 +00:00
|
|
|
rope_neox<T, false><<<block_nums, block_dims, 0, stream>>>(
|
2023-11-24 17:04:31 +00:00
|
|
|
x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
|
|
|
|
theta_scale, inv_ndims
|
2023-11-01 22:04:33 +00:00
|
|
|
);
|
2023-09-28 16:04:36 +00:00
|
|
|
} else {
|
2023-11-01 22:04:33 +00:00
|
|
|
rope_neox<T, true><<<block_nums, block_dims, 0, stream>>>(
|
2023-11-24 17:04:31 +00:00
|
|
|
x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
|
|
|
|
theta_scale, inv_ndims
|
2023-11-01 22:04:33 +00:00
|
|
|
);
|
2023-09-28 16:04:36 +00:00
|
|
|
}
|
2023-08-25 08:55:59 +00:00
|
|
|
}
|
|
|
|
|
2023-11-01 22:04:33 +00:00
|
|
|
static void rope_glm_f32_cuda(
|
|
|
|
const float * x, float * dst, int ncols, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
|
|
|
|
float freq_base, int n_ctx, cudaStream_t stream
|
|
|
|
) {
|
2023-09-08 14:58:07 +00:00
|
|
|
GGML_ASSERT(ncols % 4 == 0);
|
|
|
|
const dim3 block_dims(CUDA_ROPE_BLOCK_SIZE/4, 1, 1);
|
|
|
|
const int num_blocks_x = (ncols + CUDA_ROPE_BLOCK_SIZE - 1) / CUDA_ROPE_BLOCK_SIZE;
|
2023-07-14 13:36:41 +00:00
|
|
|
const dim3 block_nums(num_blocks_x, nrows, 1);
|
2023-11-01 22:04:33 +00:00
|
|
|
rope_glm_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, pos, freq_scale, p_delta_rows, freq_base, n_ctx);
|
2023-07-14 13:36:41 +00:00
|
|
|
}
|
|
|
|
|
2023-08-22 11:22:08 +00:00
|
|
|
static void alibi_f32_cuda(const float * x, float * dst, const int ncols, const int nrows,
|
|
|
|
const int k_rows, const int n_heads_log2_floor, const float m0,
|
|
|
|
const float m1, cudaStream_t stream) {
|
|
|
|
const dim3 block_dims(CUDA_ALIBI_BLOCK_SIZE, 1, 1);
|
|
|
|
const int num_blocks_x = (ncols + CUDA_ALIBI_BLOCK_SIZE - 1) / (CUDA_ALIBI_BLOCK_SIZE);
|
|
|
|
const dim3 block_nums(num_blocks_x, nrows, 1);
|
|
|
|
alibi_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, k_rows, n_heads_log2_floor, m0, m1);
|
|
|
|
}
|
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
static void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
|
|
|
|
const dim3 block_dims(WARP_SIZE, 1, 1);
|
2024-01-31 13:10:15 +00:00
|
|
|
const dim3 block_nums(nrows, 1, 1);
|
2023-12-07 20:26:54 +00:00
|
|
|
k_sum_rows_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void argsort_f32_i32_cuda(const float * x, int * dst, const int ncols, const int nrows, ggml_sort_order order, cudaStream_t stream) {
|
|
|
|
// bitonic sort requires ncols to be power of 2
|
|
|
|
GGML_ASSERT((ncols & (ncols - 1)) == 0);
|
|
|
|
|
|
|
|
const dim3 block_dims(ncols, 1, 1);
|
|
|
|
const dim3 block_nums(1, nrows, 1);
|
2024-02-25 10:09:09 +00:00
|
|
|
if (order == GGML_SORT_ORDER_ASC) {
|
|
|
|
k_argsort_f32_i32<GGML_SORT_ORDER_ASC><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
|
|
|
|
} else if (order == GGML_SORT_ORDER_DESC) {
|
|
|
|
k_argsort_f32_i32<GGML_SORT_ORDER_DESC><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
|
2023-12-07 20:26:54 +00:00
|
|
|
} else {
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-06-14 17:47:19 +00:00
|
|
|
static void diag_mask_inf_f32_cuda(const float * x, float * dst, const int ncols_x, const int nrows_x, const int rows_per_channel, const int n_past, cudaStream_t stream) {
|
2023-08-22 13:25:19 +00:00
|
|
|
const dim3 block_dims(1, CUDA_DIAG_MASK_INF_BLOCK_SIZE, 1);
|
2023-06-14 17:47:19 +00:00
|
|
|
const int block_num_x = (ncols_x + CUDA_DIAG_MASK_INF_BLOCK_SIZE - 1) / CUDA_DIAG_MASK_INF_BLOCK_SIZE;
|
2023-08-22 13:25:19 +00:00
|
|
|
const dim3 block_nums(nrows_x, block_num_x, 1);
|
2023-06-14 17:47:19 +00:00
|
|
|
diag_mask_inf_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols_x, rows_per_channel, n_past);
|
|
|
|
}
|
|
|
|
|
2024-02-17 21:04:16 +00:00
|
|
|
static void soft_max_f32_cuda(const float * x, const float * mask, const float * pos, float * dst, const int ncols_x, const int nrows_x, const int nrows_y, const float scale, const float max_bias, cudaStream_t stream) {
|
2023-12-01 08:51:24 +00:00
|
|
|
int nth = WARP_SIZE;
|
|
|
|
while (nth < ncols_x && nth < CUDA_SOFT_MAX_BLOCK_SIZE) nth *= 2;
|
|
|
|
const dim3 block_dims(nth, 1, 1);
|
2023-08-22 13:25:19 +00:00
|
|
|
const dim3 block_nums(nrows_x, 1, 1);
|
2024-01-09 07:58:55 +00:00
|
|
|
const size_t shmem = (GGML_PAD(ncols_x, WARP_SIZE) + WARP_SIZE)*sizeof(float);
|
|
|
|
static_assert(CUDA_SOFT_MAX_BLOCK_SIZE == 1024, "These values need to be adjusted.");
|
2024-02-17 21:04:16 +00:00
|
|
|
|
|
|
|
const uint32_t n_head_kv = nrows_x/nrows_y;
|
|
|
|
const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv));
|
|
|
|
|
|
|
|
const float m0 = powf(2.0f, -(max_bias ) / n_head_log2);
|
|
|
|
const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
|
|
|
|
|
2024-01-09 07:58:55 +00:00
|
|
|
if (shmem < g_device_caps[g_main_device].smpb) {
|
|
|
|
switch (ncols_x) {
|
|
|
|
case 32:
|
2024-02-17 21:04:16 +00:00
|
|
|
soft_max_f32<true, 32, 32><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
|
2024-01-09 07:58:55 +00:00
|
|
|
break;
|
|
|
|
case 64:
|
2024-02-17 21:04:16 +00:00
|
|
|
soft_max_f32<true, 64, 64><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
|
2024-01-09 07:58:55 +00:00
|
|
|
break;
|
|
|
|
case 128:
|
2024-02-17 21:04:16 +00:00
|
|
|
soft_max_f32<true, 128, 128><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
|
2024-01-09 07:58:55 +00:00
|
|
|
break;
|
|
|
|
case 256:
|
2024-02-17 21:04:16 +00:00
|
|
|
soft_max_f32<true, 256, 256><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
|
2024-01-09 07:58:55 +00:00
|
|
|
break;
|
|
|
|
case 512:
|
2024-02-17 21:04:16 +00:00
|
|
|
soft_max_f32<true, 512, 512><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
|
2024-01-09 07:58:55 +00:00
|
|
|
break;
|
|
|
|
case 1024:
|
2024-02-17 21:04:16 +00:00
|
|
|
soft_max_f32<true, 1024, 1024><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
|
2024-01-09 07:58:55 +00:00
|
|
|
break;
|
|
|
|
case 2048:
|
2024-02-17 21:04:16 +00:00
|
|
|
soft_max_f32<true, 2048, 1024><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
|
2024-01-09 07:58:55 +00:00
|
|
|
break;
|
|
|
|
case 4096:
|
2024-02-17 21:04:16 +00:00
|
|
|
soft_max_f32<true, 4096, 1024><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
|
2024-01-09 07:58:55 +00:00
|
|
|
break;
|
|
|
|
default:
|
2024-02-17 21:04:16 +00:00
|
|
|
soft_max_f32<true, 0, 0><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
|
2024-01-09 07:58:55 +00:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
const size_t shmem_low = WARP_SIZE*sizeof(float);
|
2024-02-17 21:04:16 +00:00
|
|
|
soft_max_f32<false, 0, 0><<<block_nums, block_dims, shmem_low, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
|
2024-01-09 07:58:55 +00:00
|
|
|
}
|
2023-06-14 17:47:19 +00:00
|
|
|
}
|
|
|
|
|
2024-01-31 13:10:15 +00:00
|
|
|
template <typename T>
|
|
|
|
static void im2col_cuda(const float* x, T* dst,
|
2023-12-13 19:54:54 +00:00
|
|
|
int IW, int IH, int OW, int OH, int KW, int KH, int IC,
|
2024-01-31 13:10:15 +00:00
|
|
|
int batch, int batch_offset, int offset_delta,
|
2023-12-13 19:54:54 +00:00
|
|
|
int s0,int s1,int p0,int p1,int d0,int d1, cudaStream_t stream) {
|
|
|
|
const int parallel_elements = OW * KW * KH;
|
|
|
|
const int num_blocks = (parallel_elements + CUDA_IM2COL_BLOCK_SIZE - 1) / CUDA_IM2COL_BLOCK_SIZE;
|
2024-01-31 13:10:15 +00:00
|
|
|
dim3 block_nums(num_blocks, OH, batch * IC);
|
|
|
|
im2col_kernel<<<block_nums, CUDA_IM2COL_BLOCK_SIZE, 0, stream>>>(x, dst, batch_offset, offset_delta, IC, IW, IH, OH, OW, KW, KH, parallel_elements, (IC * KH * KW), s0, s1, p0, p1, d0, d1);
|
2023-11-13 14:55:52 +00:00
|
|
|
}
|
|
|
|
|
2023-04-21 19:59:17 +00:00
|
|
|
// buffer pool for cuda
|
2023-05-13 13:38:36 +00:00
|
|
|
#define MAX_CUDA_BUFFERS 256
|
2023-04-21 19:59:17 +00:00
|
|
|
|
|
|
|
struct scoped_spin_lock {
|
|
|
|
std::atomic_flag& lock;
|
|
|
|
scoped_spin_lock(std::atomic_flag& lock) : lock(lock) {
|
|
|
|
while (lock.test_and_set(std::memory_order_acquire)) {
|
|
|
|
; // spin
|
|
|
|
}
|
|
|
|
}
|
|
|
|
~scoped_spin_lock() {
|
|
|
|
lock.clear(std::memory_order_release);
|
|
|
|
}
|
|
|
|
scoped_spin_lock(const scoped_spin_lock&) = delete;
|
|
|
|
scoped_spin_lock& operator=(const scoped_spin_lock&) = delete;
|
|
|
|
};
|
|
|
|
|
2023-12-24 13:34:22 +00:00
|
|
|
static std::atomic_flag g_cuda_pool_lock = ATOMIC_FLAG_INIT;
|
|
|
|
|
|
|
|
// #define DEBUG_CUDA_MALLOC
|
2023-12-26 20:23:59 +00:00
|
|
|
struct ggml_cuda_buffer {
|
2023-04-21 19:59:17 +00:00
|
|
|
void * ptr = nullptr;
|
|
|
|
size_t size = 0;
|
|
|
|
};
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static ggml_cuda_buffer g_cuda_buffer_pool[GGML_CUDA_MAX_DEVICES][MAX_CUDA_BUFFERS];
|
2023-12-24 13:34:22 +00:00
|
|
|
static size_t g_cuda_pool_size[GGML_CUDA_MAX_DEVICES] = {0};
|
2023-04-21 19:59:17 +00:00
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void * ggml_cuda_pool_malloc_leg(int device, size_t size, size_t * actual_size) {
|
2023-04-21 19:59:17 +00:00
|
|
|
scoped_spin_lock lock(g_cuda_pool_lock);
|
2023-07-21 14:27:51 +00:00
|
|
|
#ifdef DEBUG_CUDA_MALLOC
|
|
|
|
int nnz = 0;
|
2023-12-24 13:34:22 +00:00
|
|
|
size_t max_size = 0;
|
2023-07-21 14:27:51 +00:00
|
|
|
#endif
|
|
|
|
size_t best_diff = 1ull << 36;
|
|
|
|
int ibest = -1;
|
2023-04-21 19:59:17 +00:00
|
|
|
for (int i = 0; i < MAX_CUDA_BUFFERS; ++i) {
|
2023-12-26 20:23:59 +00:00
|
|
|
ggml_cuda_buffer& b = g_cuda_buffer_pool[device][i];
|
2023-07-21 14:27:51 +00:00
|
|
|
if (b.ptr != nullptr) {
|
|
|
|
#ifdef DEBUG_CUDA_MALLOC
|
|
|
|
++nnz;
|
|
|
|
if (b.size > max_size) max_size = b.size;
|
|
|
|
#endif
|
|
|
|
if (b.size >= size) {
|
|
|
|
size_t diff = b.size - size;
|
|
|
|
if (diff < best_diff) {
|
|
|
|
best_diff = diff;
|
|
|
|
ibest = i;
|
|
|
|
if (!best_diff) {
|
|
|
|
void * ptr = b.ptr;
|
|
|
|
*actual_size = b.size;
|
|
|
|
b.ptr = nullptr;
|
|
|
|
b.size = 0;
|
|
|
|
return ptr;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2023-04-21 19:59:17 +00:00
|
|
|
}
|
2023-04-20 01:14:14 +00:00
|
|
|
}
|
2023-07-21 14:27:51 +00:00
|
|
|
if (ibest >= 0) {
|
2023-12-26 20:23:59 +00:00
|
|
|
ggml_cuda_buffer& b = g_cuda_buffer_pool[device][ibest];
|
2023-07-21 14:27:51 +00:00
|
|
|
void * ptr = b.ptr;
|
|
|
|
*actual_size = b.size;
|
|
|
|
b.ptr = nullptr;
|
|
|
|
b.size = 0;
|
|
|
|
return ptr;
|
|
|
|
}
|
2023-04-21 19:59:17 +00:00
|
|
|
void * ptr;
|
2023-07-21 14:27:51 +00:00
|
|
|
size_t look_ahead_size = (size_t) (1.05 * size);
|
|
|
|
look_ahead_size = 256 * ((look_ahead_size + 255)/256);
|
2023-12-26 20:23:59 +00:00
|
|
|
ggml_cuda_set_device(device);
|
2023-07-21 14:27:51 +00:00
|
|
|
CUDA_CHECK(cudaMalloc((void **) &ptr, look_ahead_size));
|
|
|
|
*actual_size = look_ahead_size;
|
2023-12-26 20:23:59 +00:00
|
|
|
g_cuda_pool_size[device] += look_ahead_size;
|
2023-12-24 13:34:22 +00:00
|
|
|
#ifdef DEBUG_CUDA_MALLOC
|
2024-02-26 14:36:38 +00:00
|
|
|
fprintf(stderr, "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, device, nnz,
|
|
|
|
(uint32_t)(max_size/1024/1024), (uint32_t)(g_cuda_pool_size[device]/1024/1024), (uint32_t)(size/1024/1024));
|
2023-12-24 13:34:22 +00:00
|
|
|
#endif
|
2023-04-21 19:59:17 +00:00
|
|
|
return ptr;
|
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_pool_free_leg(int device, void * ptr, size_t size) {
|
2023-04-21 19:59:17 +00:00
|
|
|
scoped_spin_lock lock(g_cuda_pool_lock);
|
2023-04-20 01:14:14 +00:00
|
|
|
|
2023-04-21 19:59:17 +00:00
|
|
|
for (int i = 0; i < MAX_CUDA_BUFFERS; ++i) {
|
2023-12-26 20:23:59 +00:00
|
|
|
ggml_cuda_buffer& b = g_cuda_buffer_pool[device][i];
|
2023-04-21 19:59:17 +00:00
|
|
|
if (b.ptr == nullptr) {
|
|
|
|
b.ptr = ptr;
|
|
|
|
b.size = size;
|
|
|
|
return;
|
|
|
|
}
|
2023-04-20 01:14:14 +00:00
|
|
|
}
|
2023-04-21 19:59:17 +00:00
|
|
|
fprintf(stderr, "WARNING: cuda buffer pool full, increase MAX_CUDA_BUFFERS\n");
|
2023-12-26 20:23:59 +00:00
|
|
|
ggml_cuda_set_device(device);
|
2023-04-21 19:59:17 +00:00
|
|
|
CUDA_CHECK(cudaFree(ptr));
|
2023-12-26 20:23:59 +00:00
|
|
|
g_cuda_pool_size[device] -= size;
|
2023-12-24 13:34:22 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
#if !defined(GGML_USE_HIPBLAS)
|
|
|
|
// pool with virtual memory
|
|
|
|
static CUdeviceptr g_cuda_pool_addr[GGML_CUDA_MAX_DEVICES] = {0};
|
|
|
|
static size_t g_cuda_pool_used[GGML_CUDA_MAX_DEVICES] = {0};
|
2023-12-29 16:31:19 +00:00
|
|
|
static const size_t CUDA_POOL_VMM_MAX_SIZE = 1ull << 35; // 32 GB
|
2023-12-24 13:34:22 +00:00
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void * ggml_cuda_pool_malloc_vmm(int device, size_t size, size_t * actual_size) {
|
2023-12-24 13:34:22 +00:00
|
|
|
scoped_spin_lock lock(g_cuda_pool_lock);
|
|
|
|
|
|
|
|
// round up the allocation size to the alignment to ensure that all allocations are aligned for all data types
|
|
|
|
const size_t alignment = 128;
|
|
|
|
size = alignment * ((size + alignment - 1) / alignment);
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
size_t avail = g_cuda_pool_size[device] - g_cuda_pool_used[device];
|
2023-12-24 13:34:22 +00:00
|
|
|
|
|
|
|
if (size > avail) {
|
|
|
|
// round up to the next multiple of the granularity
|
|
|
|
size_t reserve_size = size - avail;
|
2023-12-26 20:23:59 +00:00
|
|
|
const size_t granularity = g_device_caps[device].vmm_granularity;
|
2023-12-24 13:34:22 +00:00
|
|
|
reserve_size = granularity * ((reserve_size + granularity - 1) / granularity);
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
GGML_ASSERT(g_cuda_pool_size[device] + reserve_size <= CUDA_POOL_VMM_MAX_SIZE);
|
2023-12-24 13:34:22 +00:00
|
|
|
|
|
|
|
// allocate more physical memory
|
|
|
|
CUmemAllocationProp prop = {};
|
|
|
|
prop.type = CU_MEM_ALLOCATION_TYPE_PINNED;
|
|
|
|
prop.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
|
2023-12-26 20:23:59 +00:00
|
|
|
prop.location.id = device;
|
2023-12-24 13:34:22 +00:00
|
|
|
CUmemGenericAllocationHandle handle;
|
|
|
|
CU_CHECK(cuMemCreate(&handle, reserve_size, &prop, 0));
|
|
|
|
|
|
|
|
// reserve virtual address space (if not already reserved)
|
2023-12-26 20:23:59 +00:00
|
|
|
if (g_cuda_pool_addr[device] == 0) {
|
|
|
|
CU_CHECK(cuMemAddressReserve(&g_cuda_pool_addr[device], CUDA_POOL_VMM_MAX_SIZE, 0, 0, 0));
|
2023-12-24 13:34:22 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// map at the end of the pool
|
2023-12-26 20:23:59 +00:00
|
|
|
CU_CHECK(cuMemMap(g_cuda_pool_addr[device] + g_cuda_pool_size[device], reserve_size, 0, handle, 0));
|
|
|
|
|
|
|
|
// the memory allocation handle is no longer needed after mapping
|
|
|
|
CU_CHECK(cuMemRelease(handle));
|
2023-12-24 13:34:22 +00:00
|
|
|
|
|
|
|
// set access
|
|
|
|
CUmemAccessDesc access = {};
|
|
|
|
access.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
|
2023-12-26 20:23:59 +00:00
|
|
|
access.location.id = device;
|
2023-12-24 13:34:22 +00:00
|
|
|
access.flags = CU_MEM_ACCESS_FLAGS_PROT_READWRITE;
|
2023-12-26 20:23:59 +00:00
|
|
|
CU_CHECK(cuMemSetAccess(g_cuda_pool_addr[device] + g_cuda_pool_size[device], reserve_size, &access, 1));
|
2023-12-24 13:34:22 +00:00
|
|
|
|
|
|
|
// add to the pool
|
2023-12-26 20:23:59 +00:00
|
|
|
g_cuda_pool_size[device] += reserve_size;
|
2023-12-24 13:34:22 +00:00
|
|
|
|
|
|
|
//printf("cuda pool[%d]: size increased to %llu MB (reserved %llu MB)\n",
|
|
|
|
// id, (unsigned long long) (g_cuda_pool_size[id]/1024/1024),
|
|
|
|
// (unsigned long long) (reserve_size/1024/1024));
|
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
GGML_ASSERT(g_cuda_pool_addr[device] != 0);
|
2023-12-24 13:34:22 +00:00
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
void * ptr = (void *) (g_cuda_pool_addr[device] + g_cuda_pool_used[device]);
|
2023-12-24 13:34:22 +00:00
|
|
|
*actual_size = size;
|
2023-12-26 20:23:59 +00:00
|
|
|
g_cuda_pool_used[device] += size;
|
2023-12-24 13:34:22 +00:00
|
|
|
|
|
|
|
#ifdef DEBUG_CUDA_MALLOC
|
2024-02-26 14:36:38 +00:00
|
|
|
printf("cuda pool[%d]: allocated %llu bytes at %llx\n", device, (unsigned long long) size, ptr);
|
2023-12-24 13:34:22 +00:00
|
|
|
#endif
|
|
|
|
|
|
|
|
return ptr;
|
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_pool_free_vmm(int device, void * ptr, size_t size) {
|
2023-12-24 13:34:22 +00:00
|
|
|
scoped_spin_lock lock(g_cuda_pool_lock);
|
|
|
|
|
|
|
|
#ifdef DEBUG_CUDA_MALLOC
|
2024-02-26 14:36:38 +00:00
|
|
|
printf("cuda pool[%d]: freed %llu bytes at %llx\n", device, (unsigned long long) size, ptr);
|
2023-12-24 13:34:22 +00:00
|
|
|
#endif
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
g_cuda_pool_used[device] -= size;
|
2023-12-24 13:34:22 +00:00
|
|
|
|
|
|
|
// all deallocations must be in reverse order of the allocations
|
2023-12-26 20:23:59 +00:00
|
|
|
GGML_ASSERT(ptr == (void *) (g_cuda_pool_addr[device] + g_cuda_pool_used[device]));
|
2023-12-24 13:34:22 +00:00
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void * ggml_cuda_pool_malloc(int device, size_t size, size_t * actual_size) {
|
|
|
|
if (g_device_caps[device].vmm) {
|
|
|
|
return ggml_cuda_pool_malloc_vmm(device, size, actual_size);
|
2023-12-24 13:34:22 +00:00
|
|
|
} else {
|
2023-12-26 20:23:59 +00:00
|
|
|
return ggml_cuda_pool_malloc_leg(device, size, actual_size);
|
2023-12-24 13:34:22 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_pool_free(int device, void * ptr, size_t size) {
|
|
|
|
if (g_device_caps[device].vmm) {
|
|
|
|
ggml_cuda_pool_free_vmm(device, ptr, size);
|
2023-12-24 13:34:22 +00:00
|
|
|
} else {
|
2023-12-26 20:23:59 +00:00
|
|
|
ggml_cuda_pool_free_leg(device, ptr, size);
|
2023-12-24 13:34:22 +00:00
|
|
|
}
|
2023-04-21 19:59:17 +00:00
|
|
|
}
|
2023-12-24 13:34:22 +00:00
|
|
|
#else
|
|
|
|
#define ggml_cuda_pool_malloc ggml_cuda_pool_malloc_leg
|
|
|
|
#define ggml_cuda_pool_free ggml_cuda_pool_free_leg
|
|
|
|
#endif // !defined(GGML_USE_HIPBLAS)
|
|
|
|
|
|
|
|
template<typename T>
|
|
|
|
struct cuda_pool_alloc {
|
2023-12-26 20:23:59 +00:00
|
|
|
int device = -1;
|
2023-12-24 13:34:22 +00:00
|
|
|
T * ptr = nullptr;
|
|
|
|
size_t actual_size = 0;
|
|
|
|
|
|
|
|
// size is in number of elements
|
|
|
|
T * alloc(size_t size) {
|
|
|
|
GGML_ASSERT(ptr == nullptr);
|
2023-12-26 20:23:59 +00:00
|
|
|
CUDA_CHECK(cudaGetDevice(&device));
|
|
|
|
ptr = (T *) ggml_cuda_pool_malloc(device, size * sizeof(T), &this->actual_size);
|
2023-12-24 13:34:22 +00:00
|
|
|
return ptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
cuda_pool_alloc(size_t size) {
|
|
|
|
alloc(size);
|
|
|
|
}
|
|
|
|
|
|
|
|
~cuda_pool_alloc() {
|
|
|
|
if (ptr != nullptr) {
|
2023-12-26 20:23:59 +00:00
|
|
|
ggml_cuda_pool_free(device, ptr, actual_size);
|
2023-12-24 13:34:22 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
T * get() {
|
|
|
|
return ptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
cuda_pool_alloc() = default;
|
|
|
|
cuda_pool_alloc(const cuda_pool_alloc &) = delete;
|
|
|
|
cuda_pool_alloc(cuda_pool_alloc &&) = delete;
|
|
|
|
cuda_pool_alloc& operator=(const cuda_pool_alloc &) = delete;
|
|
|
|
cuda_pool_alloc& operator=(cuda_pool_alloc &&) = delete;
|
|
|
|
};
|
2023-04-21 19:59:17 +00:00
|
|
|
|
2023-11-07 06:49:08 +00:00
|
|
|
static bool g_cublas_loaded = false;
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL bool ggml_cublas_loaded(void) {
|
2023-11-07 06:49:08 +00:00
|
|
|
return g_cublas_loaded;
|
|
|
|
}
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL void ggml_init_cublas() {
|
2023-06-06 19:33:23 +00:00
|
|
|
static bool initialized = false;
|
|
|
|
|
|
|
|
if (!initialized) {
|
2023-08-25 09:09:42 +00:00
|
|
|
|
|
|
|
#ifdef __HIP_PLATFORM_AMD__
|
|
|
|
// Workaround for a rocBLAS bug when using multiple graphics cards:
|
|
|
|
// https://github.com/ROCmSoftwarePlatform/rocBLAS/issues/1346
|
|
|
|
rocblas_initialize();
|
|
|
|
CUDA_CHECK(cudaDeviceSynchronize());
|
|
|
|
#endif
|
|
|
|
|
2023-11-07 06:49:08 +00:00
|
|
|
if (cudaGetDeviceCount(&g_device_count) != cudaSuccess) {
|
|
|
|
initialized = true;
|
|
|
|
g_cublas_loaded = false;
|
2024-02-15 15:49:01 +00:00
|
|
|
fprintf(stderr, "%s: no " GGML_CUDA_NAME " devices found, " GGML_CUDA_NAME " will be disabled\n", __func__);
|
2023-11-07 06:49:08 +00:00
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2023-06-06 19:33:23 +00:00
|
|
|
GGML_ASSERT(g_device_count <= GGML_CUDA_MAX_DEVICES);
|
|
|
|
int64_t total_vram = 0;
|
2023-10-27 14:01:23 +00:00
|
|
|
#if defined(GGML_CUDA_FORCE_MMQ)
|
|
|
|
fprintf(stderr, "%s: GGML_CUDA_FORCE_MMQ: yes\n", __func__);
|
|
|
|
#else
|
|
|
|
fprintf(stderr, "%s: GGML_CUDA_FORCE_MMQ: no\n", __func__);
|
|
|
|
#endif
|
|
|
|
#if defined(CUDA_USE_TENSOR_CORES)
|
|
|
|
fprintf(stderr, "%s: CUDA_USE_TENSOR_CORES: yes\n", __func__);
|
|
|
|
#else
|
|
|
|
fprintf(stderr, "%s: CUDA_USE_TENSOR_CORES: no\n", __func__);
|
|
|
|
#endif
|
2023-08-25 09:09:42 +00:00
|
|
|
fprintf(stderr, "%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, g_device_count);
|
2023-10-24 18:51:20 +00:00
|
|
|
for (int id = 0; id < g_device_count; ++id) {
|
2023-12-24 13:34:22 +00:00
|
|
|
int device_vmm = 0;
|
|
|
|
|
|
|
|
#if !defined(GGML_USE_HIPBLAS)
|
|
|
|
CUdevice device;
|
|
|
|
CU_CHECK(cuDeviceGet(&device, id));
|
|
|
|
CU_CHECK(cuDeviceGetAttribute(&device_vmm, CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED, device));
|
|
|
|
|
|
|
|
if (device_vmm) {
|
|
|
|
CUmemAllocationProp alloc_prop = {};
|
|
|
|
alloc_prop.type = CU_MEM_ALLOCATION_TYPE_PINNED;
|
|
|
|
alloc_prop.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
|
|
|
|
alloc_prop.location.id = id;
|
2023-12-26 20:23:59 +00:00
|
|
|
CU_CHECK(cuMemGetAllocationGranularity(&g_device_caps[id].vmm_granularity, &alloc_prop, CU_MEM_ALLOC_GRANULARITY_RECOMMENDED));
|
2023-12-24 13:34:22 +00:00
|
|
|
}
|
|
|
|
#endif // !defined(GGML_USE_HIPBLAS)
|
|
|
|
g_device_caps[id].vmm = !!device_vmm;
|
|
|
|
|
2023-06-06 19:33:23 +00:00
|
|
|
cudaDeviceProp prop;
|
|
|
|
CUDA_CHECK(cudaGetDeviceProperties(&prop, id));
|
2023-12-24 13:34:22 +00:00
|
|
|
fprintf(stderr, " Device %d: %s, compute capability %d.%d, VMM: %s\n", id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no");
|
2023-07-05 12:19:42 +00:00
|
|
|
|
2024-01-12 19:07:38 +00:00
|
|
|
g_default_tensor_split[id] = total_vram;
|
2023-06-06 19:33:23 +00:00
|
|
|
total_vram += prop.totalGlobalMem;
|
2024-01-12 19:07:38 +00:00
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
2023-12-24 13:34:22 +00:00
|
|
|
g_device_caps[id].cc = 100*prop.major + 10*prop.minor + CC_OFFSET_AMD;
|
2023-09-13 09:20:24 +00:00
|
|
|
#else
|
2023-12-24 13:34:22 +00:00
|
|
|
g_device_caps[id].cc = 100*prop.major + 10*prop.minor;
|
2023-09-13 09:20:24 +00:00
|
|
|
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
2024-01-09 07:58:55 +00:00
|
|
|
g_device_caps[id].smpb = prop.sharedMemPerBlock;
|
2023-05-01 16:11:07 +00:00
|
|
|
}
|
2023-10-24 18:51:20 +00:00
|
|
|
for (int id = 0; id < g_device_count; ++id) {
|
2024-01-12 19:07:38 +00:00
|
|
|
g_default_tensor_split[id] /= total_vram;
|
2023-05-01 16:11:07 +00:00
|
|
|
}
|
2023-04-20 18:49:53 +00:00
|
|
|
|
2023-10-24 18:51:20 +00:00
|
|
|
for (int id = 0; id < g_device_count; ++id) {
|
2023-12-26 20:23:59 +00:00
|
|
|
ggml_cuda_set_device(id);
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
// create cuda streams
|
2023-10-24 18:51:20 +00:00
|
|
|
for (int is = 0; is < MAX_STREAMS; ++is) {
|
2023-09-11 17:55:51 +00:00
|
|
|
CUDA_CHECK(cudaStreamCreateWithFlags(&g_cudaStreams[id][is], cudaStreamNonBlocking));
|
|
|
|
}
|
2023-06-06 19:33:23 +00:00
|
|
|
|
|
|
|
// create cublas handle
|
|
|
|
CUBLAS_CHECK(cublasCreate(&g_cublas_handles[id]));
|
|
|
|
CUBLAS_CHECK(cublasSetMathMode(g_cublas_handles[id], CUBLAS_TF32_TENSOR_OP_MATH));
|
|
|
|
}
|
2023-04-29 00:04:18 +00:00
|
|
|
|
2023-04-21 19:59:17 +00:00
|
|
|
// configure logging to stdout
|
2023-05-01 16:11:07 +00:00
|
|
|
// CUBLAS_CHECK(cublasLoggerConfigure(1, 1, 0, nullptr));
|
2023-06-06 19:33:23 +00:00
|
|
|
|
|
|
|
initialized = true;
|
2023-11-07 06:49:08 +00:00
|
|
|
g_cublas_loaded = true;
|
2023-06-06 19:33:23 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL void * ggml_cuda_host_malloc(size_t size) {
|
2023-05-01 16:11:07 +00:00
|
|
|
if (getenv("GGML_CUDA_NO_PINNED") != nullptr) {
|
|
|
|
return nullptr;
|
2023-04-20 18:49:53 +00:00
|
|
|
}
|
2023-05-01 16:11:07 +00:00
|
|
|
|
|
|
|
void * ptr = nullptr;
|
|
|
|
cudaError_t err = cudaMallocHost((void **) &ptr, size);
|
|
|
|
if (err != cudaSuccess) {
|
2023-12-23 15:10:51 +00:00
|
|
|
// clear the error
|
2023-06-11 13:20:52 +00:00
|
|
|
cudaGetLastError();
|
2023-12-07 20:26:54 +00:00
|
|
|
fprintf(stderr, "WARNING: failed to allocate %.2f MB of pinned memory: %s\n",
|
2023-05-01 16:11:07 +00:00
|
|
|
size/1024.0/1024.0, cudaGetErrorString(err));
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
return ptr;
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL void ggml_cuda_host_free(void * ptr) {
|
2023-05-01 16:11:07 +00:00
|
|
|
CUDA_CHECK(cudaFreeHost(ptr));
|
2023-04-20 01:14:14 +00:00
|
|
|
}
|
2023-04-28 23:31:56 +00:00
|
|
|
|
2023-06-14 17:47:19 +00:00
|
|
|
static cudaError_t ggml_cuda_cpy_tensor_2d(
|
2023-06-06 19:33:23 +00:00
|
|
|
void * dst, const struct ggml_tensor * src, int64_t i3, int64_t i2, int64_t i1_low, int64_t i1_high, cudaStream_t stream) {
|
|
|
|
|
2023-06-14 17:47:19 +00:00
|
|
|
cudaMemcpyKind kind;
|
|
|
|
char * src_ptr;
|
2024-02-25 10:09:09 +00:00
|
|
|
if (src->backend == GGML_BACKEND_TYPE_CPU) {
|
2023-06-14 17:47:19 +00:00
|
|
|
kind = cudaMemcpyHostToDevice;
|
|
|
|
src_ptr = (char *) src->data;
|
2024-02-25 10:09:09 +00:00
|
|
|
} else if (src->backend == GGML_BACKEND_TYPE_GPU || src->backend == GGML_BACKEND_TYPE_GPU_SPLIT) {
|
|
|
|
GGML_ASSERT(src->backend != GGML_BACKEND_TYPE_GPU_SPLIT || (i1_low == 0 && i1_high == src->ne[1]));
|
2023-06-14 17:47:19 +00:00
|
|
|
kind = cudaMemcpyDeviceToDevice;
|
2023-10-08 17:19:14 +00:00
|
|
|
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src->extra;
|
2023-06-14 17:47:19 +00:00
|
|
|
int id;
|
|
|
|
CUDA_CHECK(cudaGetDevice(&id));
|
|
|
|
src_ptr = (char *) extra->data_device[id];
|
|
|
|
} else {
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
char * dst_ptr = (char *) dst;
|
|
|
|
|
2023-06-06 19:33:23 +00:00
|
|
|
const int64_t ne0 = src->ne[0];
|
|
|
|
const int64_t nb0 = src->nb[0];
|
|
|
|
const int64_t nb1 = src->nb[1];
|
|
|
|
const int64_t nb2 = src->nb[2];
|
|
|
|
const int64_t nb3 = src->nb[3];
|
2023-04-28 23:31:56 +00:00
|
|
|
const enum ggml_type type = src->type;
|
2023-06-06 19:33:23 +00:00
|
|
|
const int64_t ts = ggml_type_size(type);
|
|
|
|
const int64_t bs = ggml_blck_size(type);
|
2023-12-07 20:26:54 +00:00
|
|
|
int64_t i1_diff = i1_high - i1_low;
|
2023-04-28 23:31:56 +00:00
|
|
|
|
2023-06-14 17:47:19 +00:00
|
|
|
const char * x = src_ptr + i1_low*nb1 + i2*nb2 + i3*nb3;
|
2023-12-07 20:26:54 +00:00
|
|
|
if (nb0 == ts && nb1 == ts*ne0/bs) {
|
2023-06-14 17:47:19 +00:00
|
|
|
return cudaMemcpyAsync(dst_ptr, x, i1_diff*nb1, kind, stream);
|
2023-12-07 20:26:54 +00:00
|
|
|
} else if (nb0 == ts) {
|
|
|
|
return cudaMemcpy2DAsync(dst_ptr, ts*ne0/bs, x, nb1, ts*ne0/bs, i1_diff, kind, stream);
|
|
|
|
} else {
|
|
|
|
for (int64_t i1 = 0; i1 < i1_diff; i1++) {
|
|
|
|
const void * rx = (const void *) ((const char *) x + i1*nb1);
|
|
|
|
void * rd = (void *) (dst_ptr + i1*ts*ne0/bs);
|
|
|
|
// pretend the row is a matrix with cols=1
|
|
|
|
cudaError_t r = cudaMemcpy2DAsync(rd, ts/bs, rx, nb0, ts/bs, ne0, kind, stream);
|
|
|
|
if (r != cudaSuccess) return r;
|
2023-10-08 17:19:14 +00:00
|
|
|
}
|
2023-12-07 20:26:54 +00:00
|
|
|
return cudaSuccess;
|
2023-10-08 17:19:14 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_cuda_op_get_rows(
|
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_d, const float * src1_d, float * dst_d, cudaStream_t stream) {
|
2023-10-08 17:19:14 +00:00
|
|
|
|
|
|
|
GGML_ASSERT(src1->type == GGML_TYPE_I32);
|
|
|
|
GGML_ASSERT(dst->type == GGML_TYPE_F32);
|
|
|
|
|
2023-12-13 12:04:25 +00:00
|
|
|
GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
|
|
|
|
GGML_ASSERT(src1->nb[0] == ggml_type_size(src1->type));
|
|
|
|
GGML_ASSERT(dst->nb[0] == ggml_type_size(dst->type));
|
2023-10-08 17:19:14 +00:00
|
|
|
|
|
|
|
const int32_t * src1_i32 = (const int32_t *) src1_d;
|
|
|
|
|
|
|
|
switch (src0->type) {
|
|
|
|
case GGML_TYPE_F16:
|
2023-12-13 12:04:25 +00:00
|
|
|
get_rows_cuda_float(src0, src1, dst, (const half *)src0_d, src1_i32, dst_d, stream);
|
2023-10-08 17:19:14 +00:00
|
|
|
break;
|
|
|
|
case GGML_TYPE_F32:
|
2023-12-13 12:04:25 +00:00
|
|
|
get_rows_cuda_float(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
|
2023-10-08 17:19:14 +00:00
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q4_0:
|
2023-12-13 12:04:25 +00:00
|
|
|
get_rows_cuda<QK4_0, QR4_0, dequantize_q4_0>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
|
2023-10-08 17:19:14 +00:00
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q4_1:
|
2023-12-13 12:04:25 +00:00
|
|
|
get_rows_cuda<QK4_1, QR4_1, dequantize_q4_1>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
|
2023-10-08 17:19:14 +00:00
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q5_0:
|
2023-12-13 12:04:25 +00:00
|
|
|
get_rows_cuda<QK5_0, QR5_0, dequantize_q5_0>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
|
2023-10-08 17:19:14 +00:00
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q5_1:
|
2023-12-13 12:04:25 +00:00
|
|
|
get_rows_cuda<QK5_1, QR5_1, dequantize_q5_1>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
|
2023-10-08 17:19:14 +00:00
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q8_0:
|
2023-12-13 12:04:25 +00:00
|
|
|
get_rows_cuda<QK8_0, QR8_0, dequantize_q8_0>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
|
2023-10-08 17:19:14 +00:00
|
|
|
break;
|
|
|
|
default:
|
|
|
|
// TODO: k-quants
|
2023-12-21 17:06:44 +00:00
|
|
|
fprintf(stderr, "%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type));
|
2023-10-08 17:19:14 +00:00
|
|
|
GGML_ASSERT(false);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
template<class op>
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_bin_bcast(
|
2023-09-11 17:55:51 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
2023-07-12 07:54:19 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
GGML_ASSERT(src1->type == GGML_TYPE_F32);
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2023-06-28 16:35:54 +00:00
|
|
|
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
|
2023-12-07 20:26:54 +00:00
|
|
|
op()(src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream);
|
2023-06-28 16:35:54 +00:00
|
|
|
} else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
|
2023-12-07 20:26:54 +00:00
|
|
|
op()(src0, src1, dst, (const half *) src0_dd, src1_dd, (half *) dst_dd, main_stream);
|
2023-11-01 11:49:04 +00:00
|
|
|
} else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
|
2023-12-07 20:26:54 +00:00
|
|
|
op()(src0, src1, dst, (const half *) src0_dd, src1_dd, dst_dd, main_stream);
|
2023-06-28 16:35:54 +00:00
|
|
|
} else {
|
2023-12-07 20:26:54 +00:00
|
|
|
fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__,
|
|
|
|
ggml_type_name(dst->type), ggml_type_name(src0->type), ggml_type_name(src1->type));
|
2023-06-28 16:35:54 +00:00
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
2023-12-07 20:26:54 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_cuda_op_repeat(
|
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_d, const float * src1_d, float * dst_d, cudaStream_t main_stream) {
|
2023-12-07 20:26:54 +00:00
|
|
|
|
|
|
|
ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_repeat>>(dst, src0, dst, nullptr, src0_d, dst_d, main_stream);
|
2023-06-06 19:33:23 +00:00
|
|
|
|
|
|
|
(void) src1;
|
2023-12-07 20:26:54 +00:00
|
|
|
(void) src1_d;
|
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_add(
|
2023-12-07 20:26:54 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
2023-12-07 20:26:54 +00:00
|
|
|
|
|
|
|
ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_add>>(src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream);
|
2023-06-06 19:33:23 +00:00
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_acc(
|
2023-12-13 19:54:54 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
2023-12-13 19:54:54 +00:00
|
|
|
|
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT(src1->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT(dst->ne[3] == 1); // just 3D tensors supported
|
|
|
|
|
|
|
|
int nb1 = dst->op_params[0] / 4; // 4 bytes of float32
|
|
|
|
int nb2 = dst->op_params[1] / 4; // 4 bytes of float32
|
|
|
|
// int nb3 = dst->op_params[2] / 4; // 4 bytes of float32 - unused
|
|
|
|
int offset = dst->op_params[3] / 4; // offset in bytes
|
|
|
|
|
|
|
|
acc_f32_cuda(src0_dd, src1_dd, dst_dd, ggml_nelements(dst), src1->ne[0], src1->ne[1], src1->ne[2], nb1, nb2, offset, main_stream);
|
|
|
|
|
|
|
|
(void) dst;
|
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_mul(
|
2023-09-11 17:55:51 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_mul>>(src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream);
|
|
|
|
}
|
2023-07-14 18:38:24 +00:00
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_div(
|
2023-12-07 20:26:54 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_div>>(src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream);
|
cuda : loading models directly into VRAM, norm calculation on GPU, broadcasting for ggml_mul (#1483)
* Broadcasting for ggml_mul
* CUDA kernel for ggml_mul, norms in VRAM
* GPU weights not in RAM, direct loading with cuFile
* fixup! GPU weights not in RAM, direct loading with cuFile
* fixup! GPU weights not in RAM, direct loading with cuFile
* define default model path once, sync path with readme (#1366)
* ~7% faster Q5_1 AVX2 code (#1477)
* convert.py: Support models which are stored in a single pytorch_model.bin (#1469)
* Support models in a single pytorch_model.bin
* Remove spurious line with typo
* benchmark-matmul: Print the average of the test results (#1490)
* Remove unused n_parts parameter (#1509)
* Fixes #1511 lambda issue for w64devkit (mingw) (#1513)
* Fix for w64devkit and mingw
* make kv_f16 the default for api users (#1517)
* minor : fix compile warnings
* readme : adds WizardLM to the list of supported models (#1485)
* main : make reverse prompt option act as a stop token in non-interactive mode (#1032)
* Make reverse prompt option act as a stop token in non-interactive scenarios
* Making requested review changes
* Update gpt_params_parse and fix a merge error
* Revert "Update gpt_params_parse and fix a merge error"
This reverts commit 2bb2ff1748513591ad45b175a75ed1d8089d84c8.
* Update gpt_params_parse and fix a merge error take 2
* examples : add persistent chat (#1495)
* examples : add persistent chat
* examples : fix whitespace
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* tests : add missing header
* ggml : use F16 instead of F32 in Q4_0, Q4_1, Q8_0 (#1508)
* ggml : use F16 instead of F32 in Q4_0, Q4_1 and Q8_0
* llama : bump LLAMA_FILE_VERSION to 3
* cuda : update Q4 and Q8 dequantize kernels
* ggml : fix AVX dot products
* readme : update performance table + hot topics
* ggml : fix scalar implementation of Q4_1 dot
* llama : fix compile warnings in llama_set_state_data()
* llama : fix name shadowing and C4146 (#1526)
* Fix name shadowing and C4146
* Fix if macros not using defined when required
* Update llama-util.h
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
* Update llama-util.h
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
* Code style
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Fix for mingw (#1462)
* llama : add llama_init_backend() API (close #1527)
* feature : add blis and other BLAS implementation support (#1502)
* feature: add blis support
* feature: allow all BLA_VENDOR to be assigned in cmake arguments. align with whisper.cpp pr 927
* fix: version detection for BLA_SIZEOF_INTEGER, recover min version of cmake
* Fix typo in INTEGER
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Revert "feature : add blis and other BLAS implementation support (#1502)"
This reverts commit 07e9ace0f9da424d82e75df969642522880feb92.
* GPU weights not in RAM, direct loading with cuFile
* llama : code style fixes + progress print fix
* ggml : ggml_mul better broadcast support
* cmake : workarounds for cufile when CMake version < 3.25
* gg rebase fixup
* Loop in llama.cpp, fixed progress callback
* Attempt clang-tidy fix
* llama : fix vram size computation
* Add forgotten fclose()
---------
Co-authored-by: András Salamon <ott2@users.noreply.github.com>
Co-authored-by: Ilya Kurdyukov <59548320+ilyakurdyukov@users.noreply.github.com>
Co-authored-by: Tom Jobbins <784313+TheBloke@users.noreply.github.com>
Co-authored-by: rankaiyx <rankaiyx@rankaiyx.com>
Co-authored-by: Stephan Walter <stephan@walter.name>
Co-authored-by: DannyDaemonic <DannyDaemonic@gmail.com>
Co-authored-by: Erik Scholz <Green-Sky@users.noreply.github.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: David Kennedy <dakennedyd@gmail.com>
Co-authored-by: Jason McCartney <jmac@theroot.org>
Co-authored-by: Evan Jones <evan.q.jones@gmail.com>
Co-authored-by: Maxime <672982+maximegmd@users.noreply.github.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Zenix <zenixls2@gmail.com>
2023-05-20 12:19:28 +00:00
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_gelu(
|
2023-09-11 17:55:51 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
2023-07-12 17:26:18 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
2023-07-12 17:26:18 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
gelu_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
|
2023-07-12 17:26:18 +00:00
|
|
|
|
|
|
|
(void) src1;
|
|
|
|
(void) dst;
|
2023-09-11 17:55:51 +00:00
|
|
|
(void) src1_dd;
|
2023-07-12 17:26:18 +00:00
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_silu(
|
2023-09-11 17:55:51 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
silu_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
|
2023-06-06 19:33:23 +00:00
|
|
|
|
|
|
|
(void) src1;
|
|
|
|
(void) dst;
|
2023-09-11 17:55:51 +00:00
|
|
|
(void) src1_dd;
|
2023-06-06 19:33:23 +00:00
|
|
|
}
|
2023-05-01 16:11:07 +00:00
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_gelu_quick(
|
2023-12-13 19:54:54 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
2023-12-13 19:54:54 +00:00
|
|
|
|
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
|
|
|
|
|
|
|
gelu_quick_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
|
|
|
|
|
|
|
|
(void) src1;
|
|
|
|
(void) dst;
|
|
|
|
(void) src1_dd;
|
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_tanh(
|
2023-12-13 19:54:54 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
2023-12-13 19:54:54 +00:00
|
|
|
|
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
|
|
|
|
|
|
|
tanh_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
|
|
|
|
|
|
|
|
(void) src1;
|
|
|
|
(void) dst;
|
|
|
|
(void) src1_dd;
|
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_relu(
|
2023-11-13 08:58:15 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
2023-11-13 08:58:15 +00:00
|
|
|
|
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
|
|
|
|
|
|
|
relu_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
|
|
|
|
|
|
|
|
(void) src1;
|
|
|
|
(void) dst;
|
|
|
|
(void) src1_dd;
|
|
|
|
}
|
|
|
|
|
2024-01-31 13:10:15 +00:00
|
|
|
static void ggml_cuda_op_hardsigmoid(
|
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
|
|
|
|
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
|
|
|
|
|
|
|
hardsigmoid_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
|
|
|
|
|
|
|
|
(void) src1;
|
|
|
|
(void) dst;
|
|
|
|
(void) src1_dd;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_cuda_op_hardswish(
|
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
|
|
|
|
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
|
|
|
|
|
|
|
hardswish_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
|
|
|
|
|
|
|
|
(void) src1;
|
|
|
|
(void) dst;
|
|
|
|
(void) src1_dd;
|
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_leaky_relu(
|
2023-12-13 19:54:54 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
2023-12-13 19:54:54 +00:00
|
|
|
|
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
|
|
|
|
|
|
|
float negative_slope;
|
|
|
|
memcpy(&negative_slope, dst->op_params, sizeof(float));
|
|
|
|
|
|
|
|
leaky_relu_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), negative_slope, main_stream);
|
|
|
|
|
|
|
|
(void) src1;
|
|
|
|
(void) dst;
|
|
|
|
(void) src1_dd;
|
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_sqr(
|
2023-11-13 08:58:15 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
2023-11-13 08:58:15 +00:00
|
|
|
|
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
|
|
|
|
|
|
|
sqr_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
|
|
|
|
|
|
|
|
(void) src1;
|
|
|
|
(void) dst;
|
|
|
|
(void) src1_dd;
|
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_norm(
|
2023-09-11 17:55:51 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
2023-07-11 19:53:34 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
2023-07-11 19:53:34 +00:00
|
|
|
|
|
|
|
const int64_t ne00 = src0->ne[0];
|
2023-09-11 17:55:51 +00:00
|
|
|
const int64_t nrows = ggml_nrows(src0);
|
2023-07-11 19:53:34 +00:00
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
float eps;
|
|
|
|
memcpy(&eps, dst->op_params, sizeof(float));
|
|
|
|
|
|
|
|
norm_f32_cuda(src0_dd, dst_dd, ne00, nrows, eps, main_stream);
|
2023-07-11 19:53:34 +00:00
|
|
|
|
|
|
|
(void) src1;
|
|
|
|
(void) dst;
|
2023-09-11 17:55:51 +00:00
|
|
|
(void) src1_dd;
|
2023-07-11 19:53:34 +00:00
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_group_norm(
|
2023-12-13 19:54:54 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
2023-12-13 19:54:54 +00:00
|
|
|
|
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
|
|
|
|
|
|
|
int num_groups = dst->op_params[0];
|
|
|
|
int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups);
|
|
|
|
group_norm_f32_cuda(src0_dd, dst_dd, num_groups, group_size, src0->ne[0] * src0->ne[1] * src0->ne[2], main_stream);
|
|
|
|
|
|
|
|
(void) src1;
|
|
|
|
(void) dst;
|
|
|
|
(void) src1_dd;
|
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_concat(
|
2023-12-13 19:54:54 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
2023-12-13 19:54:54 +00:00
|
|
|
|
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT(src1->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT(dst->type == GGML_TYPE_F32);
|
|
|
|
|
|
|
|
for (int i3 = 0; i3 < dst->ne[3]; i3++) {
|
|
|
|
concat_f32_cuda(src0_dd + i3 * (src0->nb[3] / 4), src1_dd + i3 * (src1->nb[3] / 4), dst_dd + i3 * (dst->nb[3] / 4), dst->ne[0], dst->ne[1], dst->ne[2], src0->ne[2], main_stream);
|
|
|
|
}
|
|
|
|
|
|
|
|
(void) src1;
|
|
|
|
(void) dst;
|
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_upscale(
|
2023-12-13 19:54:54 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
2023-12-13 19:54:54 +00:00
|
|
|
|
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT(dst->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
|
|
|
|
|
|
|
|
const int scale_factor = dst->op_params[0];
|
|
|
|
|
|
|
|
upscale_f32_cuda(src0_dd, dst_dd, src0->ne[0], src0->ne[1], src0->ne[2], scale_factor, main_stream);
|
|
|
|
|
|
|
|
(void) src1;
|
|
|
|
(void) dst;
|
2023-12-18 17:27:47 +00:00
|
|
|
(void) src1_dd;
|
2023-12-13 19:54:54 +00:00
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_pad(
|
2023-12-13 19:54:54 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
2023-12-13 19:54:54 +00:00
|
|
|
|
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT(dst->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
|
|
|
|
|
|
|
|
pad_f32_cuda(src0_dd, dst_dd,
|
|
|
|
src0->ne[0], src0->ne[1], src0->ne[2],
|
|
|
|
dst->ne[0], dst->ne[1], dst->ne[2], main_stream);
|
|
|
|
|
|
|
|
(void) src1;
|
|
|
|
(void) dst;
|
2023-12-18 17:27:47 +00:00
|
|
|
(void) src1_dd;
|
2023-12-13 19:54:54 +00:00
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_rms_norm(
|
2023-09-11 17:55:51 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
2023-05-01 16:11:07 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
2023-05-01 16:11:07 +00:00
|
|
|
|
2023-06-06 19:33:23 +00:00
|
|
|
const int64_t ne00 = src0->ne[0];
|
2023-09-11 17:55:51 +00:00
|
|
|
const int64_t nrows = ggml_nrows(src0);
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2023-07-24 15:57:12 +00:00
|
|
|
float eps;
|
|
|
|
memcpy(&eps, dst->op_params, sizeof(float));
|
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
rms_norm_f32_cuda(src0_dd, dst_dd, ne00, nrows, eps, main_stream);
|
2023-06-06 19:33:23 +00:00
|
|
|
|
|
|
|
(void) src1;
|
|
|
|
(void) dst;
|
2023-09-11 17:55:51 +00:00
|
|
|
(void) src1_dd;
|
2023-06-06 19:33:23 +00:00
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_mul_mat_q(
|
2023-09-11 17:55:51 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
|
|
|
|
const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
|
2023-12-26 20:23:59 +00:00
|
|
|
const int64_t src1_padded_row_size, cudaStream_t stream) {
|
2023-07-29 21:04:44 +00:00
|
|
|
|
|
|
|
const int64_t ne00 = src0->ne[0];
|
|
|
|
|
|
|
|
const int64_t ne10 = src1->ne[0];
|
|
|
|
GGML_ASSERT(ne10 % QK8_1 == 0);
|
|
|
|
|
|
|
|
const int64_t ne0 = dst->ne[0];
|
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
const int64_t row_diff = row_high - row_low;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
|
|
|
int id;
|
|
|
|
CUDA_CHECK(cudaGetDevice(&id));
|
|
|
|
|
|
|
|
// the main device has a larger memory buffer to hold the results from all GPUs
|
2024-02-07 11:40:26 +00:00
|
|
|
// nrows_dst == nrows of the matrix that the kernel writes into
|
2024-02-25 10:09:09 +00:00
|
|
|
const int64_t nrows_dst = dst->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device ? ne0 : row_diff;
|
2023-07-29 21:04:44 +00:00
|
|
|
|
|
|
|
switch (src0->type) {
|
|
|
|
case GGML_TYPE_Q4_0:
|
2023-09-11 17:55:51 +00:00
|
|
|
ggml_mul_mat_q4_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
|
2023-07-29 21:04:44 +00:00
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q4_1:
|
2023-09-11 17:55:51 +00:00
|
|
|
ggml_mul_mat_q4_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
|
2023-07-29 21:04:44 +00:00
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q5_0:
|
2023-09-11 17:55:51 +00:00
|
|
|
ggml_mul_mat_q5_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
|
2023-07-29 21:04:44 +00:00
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q5_1:
|
2023-09-11 17:55:51 +00:00
|
|
|
ggml_mul_mat_q5_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
|
2023-07-29 21:04:44 +00:00
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q8_0:
|
2023-09-11 17:55:51 +00:00
|
|
|
ggml_mul_mat_q8_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
|
2023-07-29 21:04:44 +00:00
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q2_K:
|
2023-09-11 17:55:51 +00:00
|
|
|
ggml_mul_mat_q2_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
|
2023-07-29 21:04:44 +00:00
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q3_K:
|
2023-09-11 17:55:51 +00:00
|
|
|
ggml_mul_mat_q3_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
|
2023-07-29 21:04:44 +00:00
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q4_K:
|
2023-09-11 17:55:51 +00:00
|
|
|
ggml_mul_mat_q4_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
|
2023-07-29 21:04:44 +00:00
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q5_K:
|
2023-09-11 17:55:51 +00:00
|
|
|
ggml_mul_mat_q5_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
|
2023-07-29 21:04:44 +00:00
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q6_K:
|
2023-09-11 17:55:51 +00:00
|
|
|
ggml_mul_mat_q6_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
|
2023-07-29 21:04:44 +00:00
|
|
|
break;
|
|
|
|
default:
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
(void) src1;
|
|
|
|
(void) dst;
|
2023-09-11 17:55:51 +00:00
|
|
|
(void) src1_ddf_i;
|
2023-07-29 21:04:44 +00:00
|
|
|
}
|
|
|
|
|
2024-01-12 19:07:38 +00:00
|
|
|
static int64_t get_row_rounding(ggml_type type, const std::array<float, GGML_CUDA_MAX_DEVICES> & tensor_split) {
|
2023-09-13 09:20:24 +00:00
|
|
|
int64_t min_compute_capability = INT_MAX;
|
|
|
|
int64_t max_compute_capability = INT_MIN;
|
2023-12-26 20:23:59 +00:00
|
|
|
for (int id = 0; id < g_device_count; ++id) {
|
2024-01-12 19:07:38 +00:00
|
|
|
if (tensor_split[id] < (id + 1 < g_device_count ? tensor_split[id + 1] : 1.0f)) {
|
2023-12-24 13:34:22 +00:00
|
|
|
if (min_compute_capability > g_device_caps[id].cc) {
|
|
|
|
min_compute_capability = g_device_caps[id].cc;
|
2023-09-13 09:20:24 +00:00
|
|
|
}
|
2023-12-24 13:34:22 +00:00
|
|
|
if (max_compute_capability < g_device_caps[id].cc) {
|
|
|
|
max_compute_capability = g_device_caps[id].cc;
|
2023-09-13 09:20:24 +00:00
|
|
|
}
|
2023-08-09 07:42:34 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-09-13 09:20:24 +00:00
|
|
|
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
switch(type) {
|
|
|
|
case GGML_TYPE_Q4_0:
|
|
|
|
case GGML_TYPE_Q4_1:
|
|
|
|
case GGML_TYPE_Q5_0:
|
|
|
|
case GGML_TYPE_Q5_1:
|
|
|
|
case GGML_TYPE_Q8_0:
|
|
|
|
return max_compute_capability >= CC_RDNA2 ? 128 : 64;
|
|
|
|
case GGML_TYPE_F16:
|
2023-11-17 08:01:15 +00:00
|
|
|
case GGML_TYPE_F32:
|
2023-09-13 09:20:24 +00:00
|
|
|
return 1;
|
|
|
|
case GGML_TYPE_Q2_K:
|
|
|
|
return max_compute_capability >= CC_RDNA2 ? 128 : 32;
|
|
|
|
case GGML_TYPE_Q3_K:
|
|
|
|
return min_compute_capability < CC_RDNA2 ? 128 : 64;
|
|
|
|
case GGML_TYPE_Q4_K:
|
|
|
|
case GGML_TYPE_Q5_K:
|
|
|
|
case GGML_TYPE_Q6_K:
|
2024-01-08 15:02:32 +00:00
|
|
|
case GGML_TYPE_IQ2_XXS:
|
2024-01-11 19:39:39 +00:00
|
|
|
case GGML_TYPE_IQ2_XS:
|
2024-02-26 16:28:38 +00:00
|
|
|
case GGML_TYPE_IQ2_S:
|
2024-01-30 13:14:12 +00:00
|
|
|
case GGML_TYPE_IQ3_XXS:
|
2024-02-18 16:16:55 +00:00
|
|
|
case GGML_TYPE_IQ1_S:
|
2024-02-21 09:39:52 +00:00
|
|
|
case GGML_TYPE_IQ4_NL:
|
2024-02-27 14:34:24 +00:00
|
|
|
case GGML_TYPE_IQ4_XS:
|
2024-02-24 14:23:52 +00:00
|
|
|
case GGML_TYPE_IQ3_S:
|
2023-09-13 09:20:24 +00:00
|
|
|
return max_compute_capability >= CC_RDNA2 ? 128 : 64;
|
|
|
|
default:
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
#else
|
2023-08-09 07:42:34 +00:00
|
|
|
switch(type) {
|
|
|
|
case GGML_TYPE_Q4_0:
|
|
|
|
case GGML_TYPE_Q4_1:
|
2023-09-30 16:12:57 +00:00
|
|
|
return max_compute_capability >= CC_VOLTA ? 128 : 64;
|
2023-08-09 07:42:34 +00:00
|
|
|
case GGML_TYPE_Q5_0:
|
|
|
|
case GGML_TYPE_Q5_1:
|
|
|
|
case GGML_TYPE_Q8_0:
|
|
|
|
return 64;
|
|
|
|
case GGML_TYPE_F16:
|
2023-11-17 08:01:15 +00:00
|
|
|
case GGML_TYPE_F32:
|
2023-08-09 07:42:34 +00:00
|
|
|
return 1;
|
|
|
|
case GGML_TYPE_Q2_K:
|
|
|
|
case GGML_TYPE_Q3_K:
|
|
|
|
case GGML_TYPE_Q4_K:
|
|
|
|
case GGML_TYPE_Q5_K:
|
2024-01-08 15:02:32 +00:00
|
|
|
case GGML_TYPE_IQ2_XXS:
|
2024-01-11 19:39:39 +00:00
|
|
|
case GGML_TYPE_IQ2_XS:
|
2024-02-26 16:28:38 +00:00
|
|
|
case GGML_TYPE_IQ2_S:
|
2024-01-30 13:14:12 +00:00
|
|
|
case GGML_TYPE_IQ3_XXS:
|
2024-02-18 16:16:55 +00:00
|
|
|
case GGML_TYPE_IQ1_S:
|
2024-02-21 09:39:52 +00:00
|
|
|
case GGML_TYPE_IQ4_NL:
|
2024-02-27 14:34:24 +00:00
|
|
|
case GGML_TYPE_IQ4_XS:
|
2024-02-24 14:23:52 +00:00
|
|
|
case GGML_TYPE_IQ3_S:
|
2023-09-30 16:12:57 +00:00
|
|
|
return max_compute_capability >= CC_VOLTA ? 128 : 64;
|
2023-08-09 07:42:34 +00:00
|
|
|
case GGML_TYPE_Q6_K:
|
|
|
|
return 64;
|
|
|
|
default:
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
2023-09-13 09:20:24 +00:00
|
|
|
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
2023-08-09 07:42:34 +00:00
|
|
|
}
|
|
|
|
|
2024-01-12 19:07:38 +00:00
|
|
|
static void get_row_split(int64_t * row_low, int64_t * row_high, const ggml_tensor * tensor, const std::array<float, GGML_CUDA_MAX_DEVICES> & tensor_split, int id) {
|
|
|
|
const int64_t nrows = ggml_nrows(tensor);
|
|
|
|
const int64_t rounding = get_row_rounding(tensor->type, tensor_split);
|
|
|
|
|
|
|
|
*row_low = id == 0 ? 0 : nrows*tensor_split[id];
|
|
|
|
*row_low -= *row_low % rounding;
|
|
|
|
|
|
|
|
if (id == g_device_count - 1) {
|
|
|
|
*row_high = nrows;
|
|
|
|
} else {
|
|
|
|
*row_high = nrows*tensor_split[id + 1];
|
|
|
|
*row_high -= *row_high % rounding;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_mul_mat_vec_q(
|
2023-09-11 17:55:51 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
|
|
|
|
const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
|
2023-12-26 20:23:59 +00:00
|
|
|
const int64_t src1_padded_row_size, cudaStream_t stream) {
|
2023-06-06 19:33:23 +00:00
|
|
|
|
|
|
|
const int64_t ne00 = src0->ne[0];
|
2023-09-11 17:55:51 +00:00
|
|
|
const int64_t row_diff = row_high - row_low;
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2024-02-07 11:40:26 +00:00
|
|
|
const int64_t ne10 = src1->ne[0];
|
|
|
|
GGML_ASSERT(ne10 % QK8_1 == 0);
|
|
|
|
|
|
|
|
const int64_t ne0 = dst->ne[0];
|
|
|
|
|
|
|
|
int id;
|
|
|
|
CUDA_CHECK(cudaGetDevice(&id));
|
|
|
|
|
|
|
|
// the main device has a larger memory buffer to hold the results from all GPUs
|
|
|
|
// nrows_dst == nrows of the matrix that the kernel writes into
|
2024-02-25 10:09:09 +00:00
|
|
|
const int64_t nrows_dst = dst->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device ? ne0 : row_diff;
|
2024-02-07 11:40:26 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
switch (src0->type) {
|
|
|
|
case GGML_TYPE_Q4_0:
|
2024-02-06 13:44:06 +00:00
|
|
|
mul_mat_vec_q_cuda<QK4_0, QI4_0, block_q4_0, VDR_Q4_0_Q8_1_MMVQ, vec_dot_q4_0_q8_1>
|
2024-02-07 11:40:26 +00:00
|
|
|
(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
2023-09-11 17:55:51 +00:00
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q4_1:
|
2024-02-06 13:44:06 +00:00
|
|
|
mul_mat_vec_q_cuda<QK4_1, QI4_1, block_q4_1, VDR_Q4_1_Q8_1_MMVQ, vec_dot_q4_1_q8_1>
|
2024-02-07 11:40:26 +00:00
|
|
|
(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
2023-09-11 17:55:51 +00:00
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q5_0:
|
2024-02-06 13:44:06 +00:00
|
|
|
mul_mat_vec_q_cuda<QK5_0, QI5_0, block_q5_0, VDR_Q5_0_Q8_1_MMVQ, vec_dot_q5_0_q8_1>
|
2024-02-07 11:40:26 +00:00
|
|
|
(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
2023-09-11 17:55:51 +00:00
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q5_1:
|
2024-02-06 13:44:06 +00:00
|
|
|
mul_mat_vec_q_cuda<QK5_1, QI5_1, block_q5_1, VDR_Q5_1_Q8_1_MMVQ, vec_dot_q5_1_q8_1>
|
2024-02-07 11:40:26 +00:00
|
|
|
(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
2023-09-11 17:55:51 +00:00
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q8_0:
|
2024-02-06 13:44:06 +00:00
|
|
|
mul_mat_vec_q_cuda<QK8_0, QI8_0, block_q8_0, VDR_Q8_0_Q8_1_MMVQ, vec_dot_q8_0_q8_1>
|
2024-02-07 11:40:26 +00:00
|
|
|
(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
2023-09-11 17:55:51 +00:00
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q2_K:
|
2024-02-06 13:44:06 +00:00
|
|
|
mul_mat_vec_q_cuda<QK_K, QI2_K, block_q2_K, VDR_Q2_K_Q8_1_MMVQ, vec_dot_q2_K_q8_1>
|
2024-02-07 11:40:26 +00:00
|
|
|
(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
2023-09-11 17:55:51 +00:00
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q3_K:
|
2024-02-06 13:44:06 +00:00
|
|
|
mul_mat_vec_q_cuda<QK_K, QI3_K, block_q3_K, VDR_Q3_K_Q8_1_MMVQ, vec_dot_q3_K_q8_1>
|
2024-02-07 11:40:26 +00:00
|
|
|
(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
2023-09-11 17:55:51 +00:00
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q4_K:
|
2024-02-06 13:44:06 +00:00
|
|
|
mul_mat_vec_q_cuda<QK_K, QI4_K, block_q4_K, VDR_Q4_K_Q8_1_MMVQ, vec_dot_q4_K_q8_1>
|
2024-02-07 11:40:26 +00:00
|
|
|
(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
2023-09-11 17:55:51 +00:00
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q5_K:
|
2024-02-06 13:44:06 +00:00
|
|
|
mul_mat_vec_q_cuda<QK_K, QI5_K, block_q5_K, VDR_Q5_K_Q8_1_MMVQ, vec_dot_q5_K_q8_1>
|
2024-02-07 11:40:26 +00:00
|
|
|
(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
2023-09-11 17:55:51 +00:00
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q6_K:
|
2024-02-06 13:44:06 +00:00
|
|
|
mul_mat_vec_q_cuda<QK_K, QI6_K, block_q6_K, VDR_Q6_K_Q8_1_MMVQ, vec_dot_q6_K_q8_1>
|
2024-02-07 11:40:26 +00:00
|
|
|
(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
2023-09-11 17:55:51 +00:00
|
|
|
break;
|
2024-01-08 15:02:32 +00:00
|
|
|
case GGML_TYPE_IQ2_XXS:
|
2024-02-06 13:44:06 +00:00
|
|
|
mul_mat_vec_q_cuda<QK_K, QI2_XXS, block_iq2_xxs, 1, vec_dot_iq2_xxs_q8_1>
|
2024-02-07 11:40:26 +00:00
|
|
|
(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
2024-01-08 15:02:32 +00:00
|
|
|
break;
|
2024-01-11 19:39:39 +00:00
|
|
|
case GGML_TYPE_IQ2_XS:
|
2024-02-06 13:44:06 +00:00
|
|
|
mul_mat_vec_q_cuda<QK_K, QI2_XS, block_iq2_xs, 1, vec_dot_iq2_xs_q8_1>
|
2024-02-07 11:40:26 +00:00
|
|
|
(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
2024-01-11 19:39:39 +00:00
|
|
|
break;
|
2024-02-26 16:28:38 +00:00
|
|
|
case GGML_TYPE_IQ2_S:
|
|
|
|
mul_mat_vec_q_cuda<QK_K, QI2_S, block_iq2_s, 1, vec_dot_iq2_s_q8_1>
|
|
|
|
(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
|
|
|
break;
|
2024-01-30 13:14:12 +00:00
|
|
|
case GGML_TYPE_IQ3_XXS:
|
2024-02-06 13:44:06 +00:00
|
|
|
mul_mat_vec_q_cuda<QK_K, QI3_XXS, block_iq3_xxs, 1, vec_dot_iq3_xxs_q8_1>
|
2024-02-07 11:40:26 +00:00
|
|
|
(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
2024-01-30 13:14:12 +00:00
|
|
|
break;
|
2024-02-18 16:16:55 +00:00
|
|
|
case GGML_TYPE_IQ1_S:
|
|
|
|
mul_mat_vec_q_cuda<QK_K, QI1_S, block_iq1_s, 1, vec_dot_iq1_s_q8_1>
|
|
|
|
(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
|
|
|
break;
|
2024-02-21 09:39:52 +00:00
|
|
|
case GGML_TYPE_IQ4_NL:
|
|
|
|
mul_mat_vec_q_cuda<QK4_NL, QI4_NL, block_iq4_nl, VDR_Q4_0_Q8_1_MMVQ, vec_dot_iq4_nl_q8_1>
|
|
|
|
(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
|
|
|
break;
|
2024-02-27 14:34:24 +00:00
|
|
|
case GGML_TYPE_IQ4_XS:
|
|
|
|
mul_mat_vec_q_cuda<QK_K, QI4_XS, block_iq4_xs, 1, vec_dot_iq4_xs_q8_1>
|
|
|
|
(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
|
|
|
break;
|
2024-02-24 14:23:52 +00:00
|
|
|
case GGML_TYPE_IQ3_S:
|
|
|
|
mul_mat_vec_q_cuda<QK_K, QI3_XS, block_iq3_s, 1, vec_dot_iq3_s_q8_1>
|
|
|
|
(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
|
|
|
break;
|
2023-09-11 17:55:51 +00:00
|
|
|
default:
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
break;
|
|
|
|
}
|
2023-06-19 08:23:56 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
(void) src1;
|
|
|
|
(void) dst;
|
|
|
|
(void) src1_ddf_i;
|
|
|
|
(void) src1_ncols;
|
|
|
|
(void) src1_padded_row_size;
|
|
|
|
}
|
2023-07-05 12:19:42 +00:00
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_dequantize_mul_mat_vec(
|
2023-09-11 17:55:51 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
|
|
|
|
const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
|
2023-12-26 20:23:59 +00:00
|
|
|
const int64_t src1_padded_row_size, cudaStream_t stream) {
|
2023-07-05 12:19:42 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
const int64_t ne00 = src0->ne[0];
|
|
|
|
const int64_t row_diff = row_high - row_low;
|
|
|
|
|
2023-12-29 08:32:31 +00:00
|
|
|
GGML_ASSERT(src1->type == GGML_TYPE_F32);
|
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
// on some GPUs it is faster to convert src1 to half and to use half precision intrinsics
|
2023-07-29 21:04:44 +00:00
|
|
|
#ifdef GGML_CUDA_F16
|
2023-12-24 13:34:22 +00:00
|
|
|
cuda_pool_alloc<half> src1_dfloat_a;
|
|
|
|
half * src1_dfloat = nullptr; // dfloat == half
|
2023-09-11 17:55:51 +00:00
|
|
|
|
2023-12-07 11:03:17 +00:00
|
|
|
bool src1_convert_f16 =
|
|
|
|
src0->type == GGML_TYPE_Q4_0 || src0->type == GGML_TYPE_Q4_1 ||
|
2023-09-11 17:55:51 +00:00
|
|
|
src0->type == GGML_TYPE_Q5_0 || src0->type == GGML_TYPE_Q5_1 ||
|
|
|
|
src0->type == GGML_TYPE_Q8_0 || src0->type == GGML_TYPE_F16;
|
|
|
|
|
|
|
|
if (src1_convert_f16) {
|
2023-12-24 13:34:22 +00:00
|
|
|
src1_dfloat = src1_dfloat_a.alloc(ne00);
|
2024-02-01 17:30:17 +00:00
|
|
|
const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
|
|
|
|
GGML_ASSERT(to_fp16_cuda != nullptr);
|
|
|
|
to_fp16_cuda(src1_ddf_i, src1_dfloat, ne00, stream);
|
2023-09-11 17:55:51 +00:00
|
|
|
}
|
2023-06-19 08:23:56 +00:00
|
|
|
#else
|
2023-09-11 17:55:51 +00:00
|
|
|
const dfloat * src1_dfloat = (const dfloat *) src1_ddf_i; // dfloat == float, no conversion
|
2023-07-29 21:04:44 +00:00
|
|
|
#endif // GGML_CUDA_F16
|
2023-06-19 08:23:56 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
switch (src0->type) {
|
|
|
|
case GGML_TYPE_Q4_0:
|
|
|
|
dequantize_mul_mat_vec_q4_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
|
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q4_1:
|
|
|
|
dequantize_mul_mat_vec_q4_1_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
|
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q5_0:
|
|
|
|
dequantize_mul_mat_vec_q5_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
|
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q5_1:
|
|
|
|
dequantize_mul_mat_vec_q5_1_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
|
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q8_0:
|
|
|
|
dequantize_mul_mat_vec_q8_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
|
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q2_K:
|
|
|
|
dequantize_mul_mat_vec_q2_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
|
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q3_K:
|
|
|
|
dequantize_mul_mat_vec_q3_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
|
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q4_K:
|
|
|
|
dequantize_mul_mat_vec_q4_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
|
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q5_K:
|
|
|
|
dequantize_mul_mat_vec_q5_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
|
|
|
|
break;
|
|
|
|
case GGML_TYPE_Q6_K:
|
|
|
|
dequantize_mul_mat_vec_q6_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
|
|
|
|
break;
|
|
|
|
case GGML_TYPE_F16:
|
|
|
|
convert_mul_mat_vec_f16_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
break;
|
|
|
|
}
|
2023-05-01 11:32:22 +00:00
|
|
|
|
2023-06-06 19:33:23 +00:00
|
|
|
(void) src1;
|
|
|
|
(void) dst;
|
2023-09-11 17:55:51 +00:00
|
|
|
(void) src1_ddq_i;
|
|
|
|
(void) src1_ncols;
|
|
|
|
(void) src1_padded_row_size;
|
2023-05-01 16:11:07 +00:00
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_mul_mat_cublas(
|
2023-09-11 17:55:51 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
|
|
|
|
const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
|
2023-12-26 20:23:59 +00:00
|
|
|
const int64_t src1_padded_row_size, cudaStream_t stream) {
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2023-10-25 07:26:27 +00:00
|
|
|
GGML_ASSERT(src0_dd_i != nullptr);
|
2023-06-06 19:33:23 +00:00
|
|
|
GGML_ASSERT(src1_ddf_i != nullptr);
|
2023-10-25 07:26:27 +00:00
|
|
|
GGML_ASSERT(dst_dd_i != nullptr);
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2023-05-01 16:11:07 +00:00
|
|
|
const int64_t ne00 = src0->ne[0];
|
|
|
|
const int64_t ne10 = src1->ne[0];
|
|
|
|
|
2023-06-06 19:33:23 +00:00
|
|
|
const int64_t ne0 = dst->ne[0];
|
2023-10-25 07:26:27 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
const int64_t row_diff = row_high - row_low;
|
|
|
|
|
2023-06-06 19:33:23 +00:00
|
|
|
int id;
|
|
|
|
CUDA_CHECK(cudaGetDevice(&id));
|
|
|
|
|
|
|
|
// the main device has a larger memory buffer to hold the results from all GPUs
|
|
|
|
// ldc == nrows of the matrix that cuBLAS writes into
|
2024-02-25 10:09:09 +00:00
|
|
|
int ldc = dst->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device ? ne0 : row_diff;
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2023-12-24 13:34:22 +00:00
|
|
|
const int compute_capability = g_device_caps[id].cc;
|
2023-09-28 10:08:28 +00:00
|
|
|
|
2023-12-18 17:27:47 +00:00
|
|
|
if (compute_capability >= CC_VOLTA && (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT) {
|
2023-12-29 08:32:31 +00:00
|
|
|
//printf("this branch\n");
|
2023-09-30 16:12:57 +00:00
|
|
|
// convert src0 and src1 to fp16, multiply as fp16, convert dst to fp32
|
2023-12-24 13:34:22 +00:00
|
|
|
cuda_pool_alloc<half> src0_as_f16;
|
2023-09-30 16:12:57 +00:00
|
|
|
if (src0->type != GGML_TYPE_F16) {
|
|
|
|
const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src0->type);
|
|
|
|
GGML_ASSERT(to_fp16_cuda != nullptr);
|
|
|
|
size_t ne = row_diff*ne00;
|
2023-12-24 13:34:22 +00:00
|
|
|
src0_as_f16.alloc(ne);
|
|
|
|
to_fp16_cuda(src0_dd_i, src0_as_f16.get(), ne, stream);
|
2023-09-30 16:12:57 +00:00
|
|
|
}
|
2023-12-24 13:34:22 +00:00
|
|
|
const half * src0_ptr = src0->type == GGML_TYPE_F16 ? (const half *) src0_dd_i : src0_as_f16.get();
|
2023-09-30 16:12:57 +00:00
|
|
|
|
2023-12-24 13:34:22 +00:00
|
|
|
cuda_pool_alloc<half> src1_as_f16;
|
2023-09-28 10:08:28 +00:00
|
|
|
if (src1->type != GGML_TYPE_F16) {
|
|
|
|
const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
|
|
|
|
GGML_ASSERT(to_fp16_cuda != nullptr);
|
|
|
|
size_t ne = src1_ncols*ne10;
|
2023-12-24 13:34:22 +00:00
|
|
|
src1_as_f16.alloc(ne);
|
|
|
|
to_fp16_cuda(src1_ddf_i, src1_as_f16.get(), ne, stream);
|
2023-09-28 10:08:28 +00:00
|
|
|
}
|
2023-12-24 13:34:22 +00:00
|
|
|
const half * src1_ptr = src1->type == GGML_TYPE_F16 ? (const half *) src1_ddf_i : src1_as_f16.get();
|
|
|
|
cuda_pool_alloc<half> dst_f16(row_diff*src1_ncols);
|
2023-09-28 10:08:28 +00:00
|
|
|
|
|
|
|
const half alpha_f16 = 1.0f;
|
|
|
|
const half beta_f16 = 0.0f;
|
|
|
|
|
|
|
|
CUBLAS_CHECK(cublasSetStream(g_cublas_handles[id], stream));
|
|
|
|
CUBLAS_CHECK(
|
|
|
|
cublasGemmEx(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N,
|
|
|
|
row_diff, src1_ncols, ne10,
|
2023-12-24 13:34:22 +00:00
|
|
|
&alpha_f16, src0_ptr, CUDA_R_16F, ne00,
|
|
|
|
src1_ptr, CUDA_R_16F, ne10,
|
|
|
|
&beta_f16, dst_f16.get(), CUDA_R_16F, ldc,
|
2023-09-28 10:08:28 +00:00
|
|
|
CUBLAS_COMPUTE_16F,
|
|
|
|
CUBLAS_GEMM_DEFAULT_TENSOR_OP));
|
|
|
|
|
|
|
|
const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
|
2023-12-24 13:34:22 +00:00
|
|
|
to_fp32_cuda(dst_f16.get(), dst_dd_i, row_diff*src1_ncols, stream);
|
2023-12-29 08:32:31 +00:00
|
|
|
} else {
|
2023-12-24 13:34:22 +00:00
|
|
|
cuda_pool_alloc<float> src0_ddq_as_f32;
|
2023-12-29 08:32:31 +00:00
|
|
|
cuda_pool_alloc<float> src1_ddq_as_f32;
|
2023-09-28 10:08:28 +00:00
|
|
|
|
|
|
|
if (src0->type != GGML_TYPE_F32) {
|
|
|
|
const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(src0->type);
|
|
|
|
GGML_ASSERT(to_fp32_cuda != nullptr);
|
2023-12-24 13:34:22 +00:00
|
|
|
src0_ddq_as_f32.alloc(row_diff*ne00);
|
|
|
|
to_fp32_cuda(src0_dd_i, src0_ddq_as_f32.get(), row_diff*ne00, stream);
|
2023-09-28 10:08:28 +00:00
|
|
|
}
|
2023-12-29 08:32:31 +00:00
|
|
|
if (src1->type != GGML_TYPE_F32) {
|
|
|
|
const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(src1->type);
|
|
|
|
GGML_ASSERT(to_fp32_cuda != nullptr);
|
|
|
|
src1_ddq_as_f32.alloc(src1_ncols*ne10);
|
|
|
|
to_fp32_cuda(src1_ddf_i, src1_ddq_as_f32.get(), src1_ncols*ne10, stream);
|
|
|
|
}
|
|
|
|
|
2023-12-24 13:34:22 +00:00
|
|
|
const float * src0_ddf_i = src0->type == GGML_TYPE_F32 ? (const float *) src0_dd_i : src0_ddq_as_f32.get();
|
2023-12-29 08:32:31 +00:00
|
|
|
const float * src1_ddf1_i = src1->type == GGML_TYPE_F32 ? (const float *) src1_ddf_i : src1_ddq_as_f32.get();
|
2023-09-28 10:08:28 +00:00
|
|
|
|
|
|
|
const float alpha = 1.0f;
|
|
|
|
const float beta = 0.0f;
|
|
|
|
|
|
|
|
CUBLAS_CHECK(cublasSetStream(g_cublas_handles[id], stream));
|
|
|
|
CUBLAS_CHECK(
|
|
|
|
cublasSgemm(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N,
|
|
|
|
row_diff, src1_ncols, ne10,
|
2023-12-29 08:32:31 +00:00
|
|
|
&alpha, src0_ddf_i, ne00,
|
|
|
|
src1_ddf1_i, ne10,
|
|
|
|
&beta, dst_dd_i, ldc));
|
2023-09-12 22:15:33 +00:00
|
|
|
}
|
2023-06-06 19:33:23 +00:00
|
|
|
|
|
|
|
(void) dst;
|
2023-09-11 17:55:51 +00:00
|
|
|
(void) src1_ddq_i;
|
|
|
|
(void) src1_padded_row_size;
|
2023-06-06 19:33:23 +00:00
|
|
|
}
|
2023-05-01 16:11:07 +00:00
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_rope(
|
2023-09-11 17:55:51 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
2023-05-01 16:11:07 +00:00
|
|
|
|
2023-09-28 16:04:36 +00:00
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
|
|
|
|
GGML_ASSERT( dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
|
|
|
|
GGML_ASSERT(src0->type == dst->type);
|
2023-05-01 16:11:07 +00:00
|
|
|
|
2023-06-06 19:33:23 +00:00
|
|
|
const int64_t ne00 = src0->ne[0];
|
2023-07-31 12:32:30 +00:00
|
|
|
const int64_t ne01 = src0->ne[1];
|
2023-09-28 16:04:36 +00:00
|
|
|
const int64_t ne2 = dst->ne[2];
|
2023-09-11 17:55:51 +00:00
|
|
|
const int64_t nrows = ggml_nrows(src0);
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2023-11-01 22:04:33 +00:00
|
|
|
//const int n_past = ((int32_t *) dst->op_params)[0];
|
|
|
|
const int n_dims = ((int32_t *) dst->op_params)[1];
|
|
|
|
const int mode = ((int32_t *) dst->op_params)[2];
|
|
|
|
const int n_ctx = ((int32_t *) dst->op_params)[3];
|
|
|
|
const int n_orig_ctx = ((int32_t *) dst->op_params)[4];
|
2023-07-21 14:27:51 +00:00
|
|
|
|
2023-11-01 22:04:33 +00:00
|
|
|
// RoPE alteration for extended context
|
|
|
|
float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
|
|
|
|
memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float));
|
|
|
|
memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float));
|
|
|
|
memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float));
|
|
|
|
memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float));
|
|
|
|
memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float));
|
|
|
|
memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float));
|
2023-09-28 16:04:36 +00:00
|
|
|
|
|
|
|
const int32_t * pos = nullptr;
|
|
|
|
if ((mode & 1) == 0) {
|
|
|
|
GGML_ASSERT(src1->type == GGML_TYPE_I32);
|
|
|
|
GGML_ASSERT(src1->ne[0] == ne2);
|
|
|
|
pos = (const int32_t *) src1_dd;
|
|
|
|
}
|
2023-05-01 11:32:22 +00:00
|
|
|
|
2023-08-23 20:08:04 +00:00
|
|
|
const bool is_neox = mode & 2;
|
|
|
|
const bool is_glm = mode & 4;
|
2023-07-14 13:36:41 +00:00
|
|
|
|
2023-11-01 22:04:33 +00:00
|
|
|
rope_corr_dims corr_dims;
|
|
|
|
ggml_rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims.v);
|
|
|
|
|
2023-06-06 19:33:23 +00:00
|
|
|
// compute
|
2023-07-14 13:36:41 +00:00
|
|
|
if (is_glm) {
|
2023-09-28 16:04:36 +00:00
|
|
|
GGML_ASSERT(false);
|
2023-11-01 22:04:33 +00:00
|
|
|
rope_glm_f32_cuda(src0_dd, dst_dd, ne00, nrows, pos, freq_scale, ne01, freq_base, n_ctx, main_stream);
|
2023-08-23 20:08:04 +00:00
|
|
|
} else if (is_neox) {
|
2023-09-28 16:04:36 +00:00
|
|
|
if (src0->type == GGML_TYPE_F32) {
|
2023-11-01 22:04:33 +00:00
|
|
|
rope_neox_cuda(
|
2023-11-24 17:04:31 +00:00
|
|
|
(const float *)src0_dd, (float *)dst_dd, ne00, n_dims, nrows, pos, freq_scale, ne01, freq_base, ext_factor,
|
2023-11-01 22:04:33 +00:00
|
|
|
attn_factor, corr_dims, main_stream
|
|
|
|
);
|
2023-09-28 16:04:36 +00:00
|
|
|
} else if (src0->type == GGML_TYPE_F16) {
|
2023-11-01 22:04:33 +00:00
|
|
|
rope_neox_cuda(
|
2023-11-24 17:04:31 +00:00
|
|
|
(const half *)src0_dd, (half *)dst_dd, ne00, n_dims, nrows, pos, freq_scale, ne01, freq_base, ext_factor,
|
2023-11-01 22:04:33 +00:00
|
|
|
attn_factor, corr_dims, main_stream
|
|
|
|
);
|
2023-09-28 16:04:36 +00:00
|
|
|
} else {
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
2023-07-14 13:36:41 +00:00
|
|
|
} else {
|
2023-09-28 16:04:36 +00:00
|
|
|
if (src0->type == GGML_TYPE_F32) {
|
2023-11-01 22:04:33 +00:00
|
|
|
rope_cuda(
|
|
|
|
(const float *)src0_dd, (float *)dst_dd, ne00, nrows, pos, freq_scale, ne01, freq_base, ext_factor,
|
|
|
|
attn_factor, corr_dims, main_stream
|
|
|
|
);
|
2023-09-28 16:04:36 +00:00
|
|
|
} else if (src0->type == GGML_TYPE_F16) {
|
2023-11-01 22:04:33 +00:00
|
|
|
rope_cuda(
|
|
|
|
(const half *)src0_dd, (half *)dst_dd, ne00, nrows, pos, freq_scale, ne01, freq_base, ext_factor,
|
|
|
|
attn_factor, corr_dims, main_stream
|
|
|
|
);
|
2023-09-28 16:04:36 +00:00
|
|
|
} else {
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
2023-07-14 13:36:41 +00:00
|
|
|
}
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2023-07-23 12:36:02 +00:00
|
|
|
(void) src1;
|
2023-06-06 19:33:23 +00:00
|
|
|
(void) dst;
|
2023-09-11 17:55:51 +00:00
|
|
|
(void) src1_dd;
|
2023-04-29 00:04:18 +00:00
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_alibi(
|
2023-09-11 17:55:51 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
2023-08-22 11:22:08 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
2023-08-22 11:22:08 +00:00
|
|
|
|
|
|
|
const int64_t ne00 = src0->ne[0];
|
|
|
|
const int64_t ne01 = src0->ne[1];
|
|
|
|
const int64_t ne02 = src0->ne[2];
|
2023-09-11 17:55:51 +00:00
|
|
|
const int64_t nrows = ggml_nrows(src0);
|
2023-08-22 11:22:08 +00:00
|
|
|
|
2023-10-10 07:50:23 +00:00
|
|
|
//const int n_past = ((int32_t *) dst->op_params)[0];
|
2023-08-22 11:22:08 +00:00
|
|
|
const int n_head = ((int32_t *) dst->op_params)[1];
|
|
|
|
float max_bias;
|
|
|
|
memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
|
|
|
|
|
2023-10-10 07:50:23 +00:00
|
|
|
//GGML_ASSERT(ne01 + n_past == ne00);
|
2023-08-22 11:22:08 +00:00
|
|
|
GGML_ASSERT(n_head == ne02);
|
|
|
|
|
|
|
|
const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
|
|
|
|
|
|
|
|
const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
|
|
|
|
const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor);
|
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
alibi_f32_cuda(src0_dd, dst_dd, ne00, nrows, ne01, n_heads_log2_floor, m0, m1, main_stream);
|
2023-08-22 11:22:08 +00:00
|
|
|
|
|
|
|
(void) src1;
|
2023-09-11 17:55:51 +00:00
|
|
|
(void) src1_dd;
|
2023-08-22 11:22:08 +00:00
|
|
|
}
|
|
|
|
|
2024-01-31 13:10:15 +00:00
|
|
|
static void ggml_cuda_op_pool2d(
|
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
|
|
|
|
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
|
|
|
|
|
|
|
const int32_t * opts = (const int32_t *)dst->op_params;
|
|
|
|
enum ggml_op_pool op = static_cast<ggml_op_pool>(opts[0]);
|
|
|
|
const int k0 = opts[1];
|
|
|
|
const int k1 = opts[2];
|
|
|
|
const int s0 = opts[3];
|
|
|
|
const int s1 = opts[4];
|
|
|
|
const int p0 = opts[5];
|
|
|
|
const int p1 = opts[6];
|
|
|
|
|
|
|
|
const int64_t IH = src0->ne[1];
|
|
|
|
const int64_t IW = src0->ne[0];
|
|
|
|
|
|
|
|
const int64_t N = dst->ne[3];
|
|
|
|
const int64_t OC = dst->ne[2];
|
|
|
|
const int64_t OH = dst->ne[1];
|
|
|
|
const int64_t OW = dst->ne[0];
|
|
|
|
|
|
|
|
const int parallel_elements = N * OC * OH * OW;
|
|
|
|
const int num_blocks = (parallel_elements + CUDA_POOL2D_BLOCK_SIZE - 1) / CUDA_POOL2D_BLOCK_SIZE;
|
|
|
|
dim3 block_nums(num_blocks);
|
|
|
|
pool2d_nchw_kernel<<<block_nums, CUDA_IM2COL_BLOCK_SIZE, 0, main_stream>>>(IH, IW, OH, OW, k1, k0, s1, s0, p1, p0, parallel_elements, src0_dd, dst_dd, op);
|
|
|
|
|
|
|
|
(void) src1;
|
|
|
|
(void) src1_dd;
|
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_im2col(
|
2023-11-13 14:55:52 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
2023-11-13 14:55:52 +00:00
|
|
|
|
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F16);
|
|
|
|
GGML_ASSERT(src1->type == GGML_TYPE_F32);
|
2024-01-31 13:10:15 +00:00
|
|
|
GGML_ASSERT( dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
|
2023-11-13 14:55:52 +00:00
|
|
|
|
|
|
|
const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
|
|
|
|
const int32_t s1 = ((const int32_t*)(dst->op_params))[1];
|
|
|
|
const int32_t p0 = ((const int32_t*)(dst->op_params))[2];
|
|
|
|
const int32_t p1 = ((const int32_t*)(dst->op_params))[3];
|
|
|
|
const int32_t d0 = ((const int32_t*)(dst->op_params))[4];
|
|
|
|
const int32_t d1 = ((const int32_t*)(dst->op_params))[5];
|
|
|
|
|
|
|
|
const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1;
|
|
|
|
|
|
|
|
const int64_t IC = src1->ne[is_2D ? 2 : 1];
|
|
|
|
const int64_t IH = is_2D ? src1->ne[1] : 1;
|
|
|
|
const int64_t IW = src1->ne[0];
|
|
|
|
|
|
|
|
const int64_t KH = is_2D ? src0->ne[1] : 1;
|
|
|
|
const int64_t KW = src0->ne[0];
|
|
|
|
|
|
|
|
const int64_t OH = is_2D ? dst->ne[2] : 1;
|
|
|
|
const int64_t OW = dst->ne[1];
|
|
|
|
|
2023-12-13 19:54:54 +00:00
|
|
|
const size_t delta_offset = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
|
2024-01-31 13:10:15 +00:00
|
|
|
const int64_t batch = src1->ne[3];
|
|
|
|
const size_t batch_offset = src1->nb[3] / 4; // nb is byte offset, src is type float32
|
2023-11-13 14:55:52 +00:00
|
|
|
|
2024-01-31 13:10:15 +00:00
|
|
|
if(dst->type == GGML_TYPE_F16) {
|
|
|
|
im2col_cuda(src1_dd, (half*) dst_dd, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, main_stream);
|
|
|
|
} else {
|
|
|
|
im2col_cuda(src1_dd, (float*) dst_dd, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, main_stream);
|
|
|
|
}
|
2023-11-13 14:55:52 +00:00
|
|
|
|
|
|
|
(void) src0;
|
|
|
|
(void) src0_dd;
|
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_sum_rows(
|
2023-12-07 20:26:54 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
2023-12-07 20:26:54 +00:00
|
|
|
|
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
|
|
|
|
|
|
|
const int64_t ncols = src0->ne[0];
|
|
|
|
const int64_t nrows = ggml_nrows(src0);
|
|
|
|
|
|
|
|
sum_rows_f32_cuda(src0_dd, dst_dd, ncols, nrows, main_stream);
|
|
|
|
|
|
|
|
(void) src1;
|
|
|
|
(void) dst;
|
|
|
|
(void) src1_dd;
|
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_argsort(
|
2023-12-07 20:26:54 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
2023-12-07 20:26:54 +00:00
|
|
|
|
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT( dst->type == GGML_TYPE_I32);
|
|
|
|
|
|
|
|
const int64_t ncols = src0->ne[0];
|
|
|
|
const int64_t nrows = ggml_nrows(src0);
|
|
|
|
|
|
|
|
enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0];
|
|
|
|
|
|
|
|
argsort_f32_i32_cuda(src0_dd, (int *)dst_dd, ncols, nrows, order, main_stream);
|
|
|
|
|
|
|
|
(void) src1;
|
|
|
|
(void) dst;
|
|
|
|
(void) src1_dd;
|
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_diag_mask_inf(
|
2023-09-11 17:55:51 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
2023-06-14 17:47:19 +00:00
|
|
|
|
|
|
|
const int64_t ne00 = src0->ne[0];
|
|
|
|
const int64_t ne01 = src0->ne[1];
|
2023-09-11 17:55:51 +00:00
|
|
|
const int nrows0 = ggml_nrows(src0);
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2023-07-23 12:36:02 +00:00
|
|
|
const int n_past = ((int32_t *) dst->op_params)[0];
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
diag_mask_inf_f32_cuda(src0_dd, dst_dd, ne00, nrows0, ne01, n_past, main_stream);
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2023-07-23 12:36:02 +00:00
|
|
|
(void) src1;
|
2023-06-14 17:47:19 +00:00
|
|
|
(void) dst;
|
2023-09-11 17:55:51 +00:00
|
|
|
(void) src1_dd;
|
2023-06-14 17:47:19 +00:00
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_soft_max(
|
2023-09-11 17:55:51 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2023-12-01 08:51:24 +00:00
|
|
|
GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F32); // src1 contains mask and it is optional
|
|
|
|
|
2024-02-17 21:04:16 +00:00
|
|
|
const int64_t ne00 = src0->ne[0];
|
2023-12-01 08:51:24 +00:00
|
|
|
const int64_t nrows_x = ggml_nrows(src0);
|
2024-02-17 21:04:16 +00:00
|
|
|
const int64_t nrows_y = src0->ne[1];
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2024-02-17 21:04:16 +00:00
|
|
|
float scale = 1.0f;
|
|
|
|
float max_bias = 0.0f;
|
2023-12-01 08:51:24 +00:00
|
|
|
|
2024-02-17 21:04:16 +00:00
|
|
|
memcpy(&scale, (float *) dst->op_params + 0, sizeof(float));
|
|
|
|
memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float));
|
2024-01-09 07:58:55 +00:00
|
|
|
|
2024-02-17 21:04:16 +00:00
|
|
|
// positions tensor
|
2024-02-19 08:04:45 +00:00
|
|
|
float * src2_dd = nullptr;
|
2024-02-17 21:04:16 +00:00
|
|
|
cuda_pool_alloc<float> src2_f;
|
|
|
|
|
|
|
|
ggml_tensor * src2 = dst->src[2];
|
|
|
|
const bool use_src2 = src2 != nullptr;
|
|
|
|
|
|
|
|
if (use_src2) {
|
2024-02-25 10:09:09 +00:00
|
|
|
const bool src2_on_device = src2->backend == GGML_BACKEND_TYPE_GPU;
|
2024-02-17 21:04:16 +00:00
|
|
|
|
|
|
|
if (src2_on_device) {
|
2024-02-19 08:04:45 +00:00
|
|
|
ggml_tensor_extra_gpu * src2_extra = (ggml_tensor_extra_gpu *) src2->extra;
|
2024-02-17 21:04:16 +00:00
|
|
|
src2_dd = (float *) src2_extra->data_device[g_main_device];
|
|
|
|
} else {
|
|
|
|
src2_dd = src2_f.alloc(ggml_nelements(src2));
|
|
|
|
CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src2_dd, src2, 0, 0, 0, 1, main_stream));
|
|
|
|
}
|
2024-01-09 07:58:55 +00:00
|
|
|
}
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2024-02-17 21:04:16 +00:00
|
|
|
soft_max_f32_cuda(src0_dd, src1 ? src1_dd : nullptr, src2_dd, dst_dd, ne00, nrows_x, nrows_y, scale, max_bias, main_stream);
|
2023-06-14 17:47:19 +00:00
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_scale(
|
2023-09-11 17:55:51 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2023-12-22 11:12:53 +00:00
|
|
|
float scale;
|
|
|
|
memcpy(&scale, dst->op_params, sizeof(float));
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
scale_f32_cuda(src0_dd, dst_dd, scale, ggml_nelements(src0), main_stream);
|
2023-06-14 17:47:19 +00:00
|
|
|
CUDA_CHECK(cudaGetLastError());
|
|
|
|
|
|
|
|
(void) src1;
|
|
|
|
(void) dst;
|
2023-09-11 17:55:51 +00:00
|
|
|
(void) src1_dd;
|
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
static void ggml_cuda_op_clamp(
|
2023-10-10 07:50:23 +00:00
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
|
2023-12-26 20:23:59 +00:00
|
|
|
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
|
2023-10-10 07:50:23 +00:00
|
|
|
|
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F32);
|
|
|
|
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
|
|
|
|
2023-11-01 22:10:09 +00:00
|
|
|
float min;
|
|
|
|
float max;
|
|
|
|
memcpy(&min, dst->op_params, sizeof(float));
|
|
|
|
memcpy(&max, (float *) dst->op_params + 1, sizeof(float));
|
2023-10-10 07:50:23 +00:00
|
|
|
|
|
|
|
clamp_f32_cuda(src0_dd, dst_dd, min, max, ggml_nelements(src0), main_stream);
|
|
|
|
CUDA_CHECK(cudaGetLastError());
|
|
|
|
|
|
|
|
(void) src1;
|
|
|
|
(void) dst;
|
|
|
|
(void) src1_dd;
|
|
|
|
}
|
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
static void ggml_cuda_op_flatten(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const ggml_cuda_op_flatten_t op) {
|
|
|
|
const int64_t nrows0 = ggml_nrows(src0);
|
|
|
|
|
|
|
|
const bool use_src1 = src1 != nullptr;
|
|
|
|
const int64_t nrows1 = use_src1 ? ggml_nrows(src1) : 1;
|
|
|
|
|
2024-02-25 10:09:09 +00:00
|
|
|
GGML_ASSERT(!use_src1 || src1->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
|
|
|
|
GGML_ASSERT( dst->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
|
2023-09-11 17:55:51 +00:00
|
|
|
|
2023-10-08 17:19:14 +00:00
|
|
|
ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
|
|
|
|
ggml_tensor_extra_gpu * src1_extra = use_src1 ? (ggml_tensor_extra_gpu *) src1->extra : nullptr;
|
|
|
|
ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
|
2023-09-11 17:55:51 +00:00
|
|
|
|
2024-02-25 10:09:09 +00:00
|
|
|
const bool src0_on_device = src0->backend == GGML_BACKEND_TYPE_GPU || src0->backend == GGML_BACKEND_TYPE_GPU_SPLIT;
|
|
|
|
const bool src1_on_device = use_src1 && src1->backend == GGML_BACKEND_TYPE_GPU;
|
|
|
|
const bool dst_on_device = dst->backend == GGML_BACKEND_TYPE_GPU;
|
2023-09-11 17:55:51 +00:00
|
|
|
|
|
|
|
// dd = data device
|
|
|
|
float * src0_ddf = nullptr;
|
|
|
|
float * src1_ddf = nullptr;
|
|
|
|
float * dst_ddf = nullptr;
|
|
|
|
|
2023-12-24 13:34:22 +00:00
|
|
|
cuda_pool_alloc<float> src0_f;
|
|
|
|
cuda_pool_alloc<float> src1_f;
|
|
|
|
cuda_pool_alloc<float> dst_f;
|
2023-09-11 17:55:51 +00:00
|
|
|
|
|
|
|
ggml_cuda_set_device(g_main_device);
|
2023-12-24 13:34:22 +00:00
|
|
|
cudaStream_t main_stream = g_cudaStreams[g_main_device][0];
|
2023-09-11 17:55:51 +00:00
|
|
|
|
|
|
|
if (src0_on_device) {
|
|
|
|
src0_ddf = (float *) src0_extra->data_device[g_main_device];
|
|
|
|
} else {
|
2023-12-24 13:34:22 +00:00
|
|
|
src0_ddf = src0_f.alloc(ggml_nelements(src0));
|
2023-09-11 17:55:51 +00:00
|
|
|
CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src0_ddf, src0, 0, 0, 0, nrows0, main_stream));
|
|
|
|
}
|
|
|
|
|
2023-12-21 21:20:49 +00:00
|
|
|
if (use_src1) {
|
2023-09-11 17:55:51 +00:00
|
|
|
if (src1_on_device) {
|
|
|
|
src1_ddf = (float *) src1_extra->data_device[g_main_device];
|
|
|
|
} else {
|
2023-12-24 13:34:22 +00:00
|
|
|
src1_ddf = src1_f.alloc(ggml_nelements(src1));
|
2023-09-11 17:55:51 +00:00
|
|
|
CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src1_ddf, src1, 0, 0, 0, nrows1, main_stream));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (dst_on_device) {
|
|
|
|
dst_ddf = (float *) dst_extra->data_device[g_main_device];
|
|
|
|
} else {
|
2023-12-24 13:34:22 +00:00
|
|
|
dst_ddf = dst_f.alloc(ggml_nelements(dst));
|
2023-09-11 17:55:51 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// do the computation
|
|
|
|
op(src0, src1, dst, src0_ddf, src1_ddf, dst_ddf, main_stream);
|
|
|
|
CUDA_CHECK(cudaGetLastError());
|
|
|
|
|
|
|
|
// copy dst to host if necessary
|
|
|
|
if (!dst_on_device) {
|
|
|
|
CUDA_CHECK(cudaMemcpyAsync(dst->data, dst_ddf, ggml_nbytes(dst), cudaMemcpyDeviceToHost, main_stream));
|
|
|
|
}
|
|
|
|
|
2024-02-25 10:09:09 +00:00
|
|
|
if (dst->backend == GGML_BACKEND_TYPE_CPU) {
|
2023-09-11 17:55:51 +00:00
|
|
|
CUDA_CHECK(cudaDeviceSynchronize());
|
|
|
|
}
|
2023-06-14 17:47:19 +00:00
|
|
|
}
|
|
|
|
|
2023-09-28 16:04:36 +00:00
|
|
|
static void ggml_cuda_set_peer_access(const int n_tokens) {
|
2023-09-17 14:37:53 +00:00
|
|
|
static bool peer_access_enabled = false;
|
|
|
|
|
|
|
|
const bool enable_peer_access = n_tokens <= GGML_CUDA_PEER_MAX_BATCH_SIZE;
|
|
|
|
|
|
|
|
if (peer_access_enabled == enable_peer_access) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
#ifdef NDEBUG
|
2023-12-20 14:41:22 +00:00
|
|
|
for (int id = 0; id < g_device_count; ++id) {
|
2023-12-26 20:23:59 +00:00
|
|
|
ggml_cuda_set_device(id);
|
2023-12-20 14:41:22 +00:00
|
|
|
CUDA_CHECK(cudaDeviceSynchronize());
|
|
|
|
}
|
|
|
|
|
2023-09-17 14:37:53 +00:00
|
|
|
for (int id = 0; id < g_device_count; ++id) {
|
2023-12-26 20:23:59 +00:00
|
|
|
ggml_cuda_set_device(id);
|
2023-09-17 14:37:53 +00:00
|
|
|
|
|
|
|
for (int id_other = 0; id_other < g_device_count; ++id_other) {
|
|
|
|
if (id == id_other) {
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
if (id != g_main_device && id_other != g_main_device) {
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
2023-09-17 21:35:20 +00:00
|
|
|
int can_access_peer;
|
|
|
|
CUDA_CHECK(cudaDeviceCanAccessPeer(&can_access_peer, id, id_other));
|
|
|
|
if (can_access_peer) {
|
|
|
|
if (enable_peer_access) {
|
2024-02-19 22:40:26 +00:00
|
|
|
cudaError_t err = cudaDeviceEnablePeerAccess(id_other, 0);
|
|
|
|
if (err != cudaErrorPeerAccessAlreadyEnabled) {
|
|
|
|
CUDA_CHECK(err);
|
|
|
|
}
|
2023-09-17 21:35:20 +00:00
|
|
|
} else {
|
2024-02-19 22:40:26 +00:00
|
|
|
cudaError_t err = cudaDeviceDisablePeerAccess(id_other);
|
|
|
|
if (err != cudaErrorPeerAccessNotEnabled) {
|
|
|
|
CUDA_CHECK(err);
|
|
|
|
}
|
2023-09-17 21:35:20 +00:00
|
|
|
}
|
2023-09-17 14:37:53 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
#endif // NDEBUG
|
|
|
|
|
|
|
|
peer_access_enabled = enable_peer_access;
|
|
|
|
}
|
|
|
|
|
2024-01-12 19:07:38 +00:00
|
|
|
// FIXME: move this somewhere else
|
|
|
|
struct ggml_backend_cuda_split_buffer_type_context {
|
|
|
|
std::array<float, GGML_CUDA_MAX_DEVICES> tensor_split;
|
|
|
|
};
|
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
static void ggml_cuda_op_mul_mat(
|
|
|
|
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, ggml_cuda_op_mul_mat_t op,
|
|
|
|
const bool convert_src1_to_q8_1) {
|
|
|
|
|
2023-05-01 16:11:07 +00:00
|
|
|
const int64_t ne00 = src0->ne[0];
|
|
|
|
const int64_t ne01 = src0->ne[1];
|
|
|
|
const int64_t ne02 = src0->ne[2];
|
|
|
|
const int64_t ne03 = src0->ne[3];
|
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
const int64_t ne10 = src1->ne[0];
|
|
|
|
const int64_t ne11 = src1->ne[1];
|
|
|
|
const int64_t ne12 = src1->ne[2];
|
|
|
|
const int64_t ne13 = src1->ne[3];
|
|
|
|
const int64_t nrows1 = ggml_nrows(src1);
|
2023-07-23 12:09:47 +00:00
|
|
|
|
|
|
|
GGML_ASSERT(ne03 == ne13);
|
2023-06-06 19:33:23 +00:00
|
|
|
|
|
|
|
const int64_t ne0 = dst->ne[0];
|
|
|
|
const int64_t ne1 = dst->ne[1];
|
2023-05-01 16:11:07 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
const int nb2 = dst->nb[2];
|
|
|
|
const int nb3 = dst->nb[3];
|
2023-05-01 16:11:07 +00:00
|
|
|
|
2024-02-25 10:09:09 +00:00
|
|
|
GGML_ASSERT(dst->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
|
|
|
|
GGML_ASSERT(src1->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
|
2023-12-29 08:32:31 +00:00
|
|
|
GGML_ASSERT(src1->type == GGML_TYPE_F32 || (src1->ne[2] == 1 && src1->ne[3] == 1));
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
GGML_ASSERT(ne12 >= ne02 && ne12 % ne02 == 0);
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
const int64_t i02_divisor = ne12 / ne02;
|
2023-07-23 12:09:47 +00:00
|
|
|
|
2023-06-06 19:33:23 +00:00
|
|
|
const size_t src0_ts = ggml_type_size(src0->type);
|
|
|
|
const size_t src0_bs = ggml_blck_size(src0->type);
|
2023-09-11 17:55:51 +00:00
|
|
|
const size_t q8_1_ts = sizeof(block_q8_1);
|
|
|
|
const size_t q8_1_bs = QK8_1;
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2023-10-08 17:19:14 +00:00
|
|
|
ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
|
|
|
|
ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
|
|
|
|
ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2024-02-25 10:09:09 +00:00
|
|
|
const bool src0_on_device = src0->backend == GGML_BACKEND_TYPE_GPU || src0->backend == GGML_BACKEND_TYPE_GPU_SPLIT;
|
2023-06-14 17:47:19 +00:00
|
|
|
const bool src0_is_contiguous = ggml_is_contiguous(src0);
|
2023-09-11 17:55:51 +00:00
|
|
|
const bool src1_is_contiguous = ggml_is_contiguous(src1);
|
2023-12-07 11:03:17 +00:00
|
|
|
|
|
|
|
const int64_t src1_padded_col_size = GGML_PAD(ne10, MATRIX_ROW_PADDING);
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2024-02-25 10:09:09 +00:00
|
|
|
const bool split = src0->backend == GGML_BACKEND_TYPE_GPU_SPLIT;
|
2023-09-11 17:55:51 +00:00
|
|
|
GGML_ASSERT(!(split && ne02 > 1));
|
|
|
|
GGML_ASSERT(!(split && ne03 > 1));
|
2023-07-23 12:09:47 +00:00
|
|
|
GGML_ASSERT(!(split && ne02 < ne12));
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2024-01-12 19:07:38 +00:00
|
|
|
std::array<float, GGML_CUDA_MAX_DEVICES> tensor_split;
|
|
|
|
if (split) {
|
2024-02-25 10:09:09 +00:00
|
|
|
// TODO: check that src0->buffer->buft is a split buffer type, replace GGML_BACKEND_TYPE_GPU_SPLIT check
|
2024-01-12 19:07:38 +00:00
|
|
|
// GGML_ASSERT(src0->buffer != nullptr && src0->buffer->buft == ...);
|
|
|
|
ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) src0->buffer->buft->context;
|
|
|
|
tensor_split = buft_ctx->tensor_split;
|
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
struct dev_data {
|
|
|
|
cuda_pool_alloc<char> src0_dd_alloc;
|
|
|
|
cuda_pool_alloc<float> src1_ddf_alloc;
|
|
|
|
cuda_pool_alloc<char> src1_ddq_alloc;
|
|
|
|
cuda_pool_alloc<float> dst_dd_alloc;
|
|
|
|
|
|
|
|
char * src0_dd = nullptr;
|
|
|
|
float * src1_ddf = nullptr; // float
|
|
|
|
char * src1_ddq = nullptr; // q8_1
|
|
|
|
float * dst_dd = nullptr;
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
int64_t row_low;
|
|
|
|
int64_t row_high;
|
|
|
|
};
|
2023-06-17 17:15:02 +00:00
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
dev_data dev[GGML_CUDA_MAX_DEVICES];
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2023-11-05 15:08:57 +00:00
|
|
|
int used_devices = 0;
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
for (int id = 0; id < g_device_count; ++id) {
|
2023-09-11 17:55:51 +00:00
|
|
|
// by default, use all rows
|
2023-12-26 20:23:59 +00:00
|
|
|
dev[id].row_low = 0;
|
|
|
|
dev[id].row_high = ne01;
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
// for multi GPU, get the row boundaries from tensor split
|
|
|
|
// and round to mul_mat_q tile sizes
|
2023-06-06 19:33:23 +00:00
|
|
|
if (split) {
|
2024-01-12 19:07:38 +00:00
|
|
|
const int64_t rounding = get_row_rounding(src0->type, tensor_split);
|
2023-08-09 07:42:34 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
if (id != 0) {
|
2024-01-12 19:07:38 +00:00
|
|
|
dev[id].row_low = ne01*tensor_split[id];
|
2023-12-26 20:23:59 +00:00
|
|
|
if (dev[id].row_low < ne01) {
|
|
|
|
dev[id].row_low -= dev[id].row_low % rounding;
|
2023-12-23 08:16:33 +00:00
|
|
|
}
|
2023-09-11 17:55:51 +00:00
|
|
|
}
|
2023-07-29 21:04:44 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
if (id != g_device_count - 1) {
|
2024-01-12 19:07:38 +00:00
|
|
|
dev[id].row_high = ne01*tensor_split[id + 1];
|
2023-12-26 20:23:59 +00:00
|
|
|
if (dev[id].row_high < ne01) {
|
|
|
|
dev[id].row_high -= dev[id].row_high % rounding;
|
2023-12-23 08:16:33 +00:00
|
|
|
}
|
2023-08-02 14:48:10 +00:00
|
|
|
}
|
2023-06-06 19:33:23 +00:00
|
|
|
}
|
2023-09-11 17:55:51 +00:00
|
|
|
}
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
for (int id = 0; id < g_device_count; ++id) {
|
|
|
|
if ((!split && id != g_main_device) || dev[id].row_low == dev[id].row_high) {
|
2023-06-06 19:33:23 +00:00
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
2023-11-05 15:08:57 +00:00
|
|
|
used_devices++;
|
|
|
|
|
2024-02-25 10:09:09 +00:00
|
|
|
const bool src1_on_device = src1->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device;
|
|
|
|
const bool dst_on_device = dst->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device;
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
ggml_cuda_set_device(id);
|
2023-12-26 20:23:59 +00:00
|
|
|
cudaStream_t stream = g_cudaStreams[id][0];
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2023-06-14 17:47:19 +00:00
|
|
|
if (src0_on_device && src0_is_contiguous) {
|
2023-12-26 20:23:59 +00:00
|
|
|
dev[id].src0_dd = (char *) src0_extra->data_device[id];
|
2023-06-06 19:33:23 +00:00
|
|
|
} else {
|
2023-12-26 20:23:59 +00:00
|
|
|
dev[id].src0_dd = dev[id].src0_dd_alloc.alloc(ggml_nbytes(src0));
|
2023-06-06 19:33:23 +00:00
|
|
|
}
|
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
if (src1_on_device && src1_is_contiguous) {
|
2023-12-26 20:23:59 +00:00
|
|
|
dev[id].src1_ddf = (float *) src1_extra->data_device[id];
|
2023-09-11 17:55:51 +00:00
|
|
|
} else {
|
2023-12-26 20:23:59 +00:00
|
|
|
dev[id].src1_ddf = dev[id].src1_ddf_alloc.alloc(ggml_nelements(src1));
|
2023-06-06 19:33:23 +00:00
|
|
|
}
|
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
if (convert_src1_to_q8_1) {
|
2023-12-26 20:23:59 +00:00
|
|
|
dev[id].src1_ddq = dev[id].src1_ddq_alloc.alloc(nrows1*src1_padded_col_size*q8_1_ts/q8_1_bs);
|
2023-09-11 17:55:51 +00:00
|
|
|
|
2023-10-08 17:19:14 +00:00
|
|
|
if (src1_on_device && src1_is_contiguous) {
|
2023-12-26 20:23:59 +00:00
|
|
|
quantize_row_q8_1_cuda(dev[id].src1_ddf, dev[id].src1_ddq, ne10, nrows1, src1_padded_col_size, stream);
|
2023-11-05 07:12:13 +00:00
|
|
|
CUDA_CHECK(cudaGetLastError());
|
2023-05-13 13:38:36 +00:00
|
|
|
}
|
2023-06-06 19:33:23 +00:00
|
|
|
}
|
2023-09-11 17:55:51 +00:00
|
|
|
|
2023-06-06 19:33:23 +00:00
|
|
|
if (dst_on_device) {
|
2023-12-26 20:23:59 +00:00
|
|
|
dev[id].dst_dd = (float *) dst_extra->data_device[id];
|
2023-06-06 19:33:23 +00:00
|
|
|
} else {
|
2023-12-26 20:23:59 +00:00
|
|
|
const size_t size_dst_ddf = split ? (dev[id].row_high - dev[id].row_low)*ne1 : ggml_nelements(dst);
|
|
|
|
dev[id].dst_dd = dev[id].dst_dd_alloc.alloc(size_dst_ddf);
|
2023-06-06 19:33:23 +00:00
|
|
|
}
|
2023-09-11 17:55:51 +00:00
|
|
|
}
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
// if multiple devices are used they need to wait for the main device
|
|
|
|
// here an event is recorded that signals that the main device has finished calculating the input data
|
2023-11-05 15:08:57 +00:00
|
|
|
if (split && used_devices > 1) {
|
2023-12-26 20:23:59 +00:00
|
|
|
ggml_cuda_set_device(g_main_device);
|
2023-09-11 17:55:51 +00:00
|
|
|
CUDA_CHECK(cudaEventRecord(src0_extra->events[g_main_device][0], g_cudaStreams[g_main_device][0]));
|
|
|
|
}
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2023-11-05 15:08:57 +00:00
|
|
|
const int64_t src1_col_stride = split && used_devices > 1 ? MUL_MAT_SRC1_COL_STRIDE : ne11;
|
2023-09-11 17:55:51 +00:00
|
|
|
for (int64_t src1_col_0 = 0; src1_col_0 < ne11; src1_col_0 += src1_col_stride) {
|
|
|
|
const int64_t is = split ? (src1_col_0/src1_col_stride) % MAX_STREAMS : 0;
|
|
|
|
const int64_t src1_ncols = src1_col_0 + src1_col_stride > ne11 ? ne11 - src1_col_0 : src1_col_stride;
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
for (int id = 0; id < g_device_count; ++id) {
|
|
|
|
if ((!split && id != g_main_device) || dev[id].row_low == dev[id].row_high) {
|
2023-09-11 17:55:51 +00:00
|
|
|
continue;
|
|
|
|
}
|
2023-06-06 19:33:23 +00:00
|
|
|
|
2024-02-25 10:09:09 +00:00
|
|
|
const bool src1_on_device = src1->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device;
|
|
|
|
const bool dst_on_device = dst->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device;
|
2023-12-26 20:23:59 +00:00
|
|
|
const int64_t row_diff = dev[id].row_high - dev[id].row_low;
|
2023-06-09 11:58:15 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
ggml_cuda_set_device(id);
|
2023-12-26 20:23:59 +00:00
|
|
|
cudaStream_t stream = g_cudaStreams[id][is];
|
2023-06-09 11:58:15 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
// wait for main GPU data if necessary
|
|
|
|
if (split && (id != g_main_device || is != 0)) {
|
2023-09-16 14:55:43 +00:00
|
|
|
CUDA_CHECK(cudaStreamWaitEvent(stream, src0_extra->events[g_main_device][0], 0));
|
2023-09-11 17:55:51 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
for (int64_t i0 = 0; i0 < ne13*ne12; ++i0) {
|
|
|
|
const int64_t i03 = i0 / ne12;
|
|
|
|
const int64_t i02 = i0 % ne12;
|
|
|
|
|
|
|
|
const size_t src1_ddq_i_offset = (i0*ne11 + src1_col_0) * src1_padded_col_size*q8_1_ts/q8_1_bs;
|
2023-06-06 19:33:23 +00:00
|
|
|
|
|
|
|
// for split tensors the data begins at i0 == i0_offset_low
|
2023-12-26 20:23:59 +00:00
|
|
|
char * src0_dd_i = dev[id].src0_dd + (i0/i02_divisor) * (ne01*ne00*src0_ts)/src0_bs;
|
|
|
|
float * src1_ddf_i = dev[id].src1_ddf + (i0*ne11 + src1_col_0) * ne10;
|
|
|
|
char * src1_ddq_i = dev[id].src1_ddq + src1_ddq_i_offset;
|
|
|
|
float * dst_dd_i = dev[id].dst_dd + (i0*ne1 + src1_col_0) * (dst_on_device ? ne0 : row_diff);
|
2023-06-06 19:33:23 +00:00
|
|
|
|
|
|
|
// the main device memory buffer can be on VRAM scratch, with space for all partial results
|
|
|
|
// in that case an offset on dst_ddf_i is needed
|
2024-02-25 10:09:09 +00:00
|
|
|
if (dst->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device) {
|
2023-12-26 20:23:59 +00:00
|
|
|
dst_dd_i += dev[id].row_low; // offset is 0 if no tensor split
|
2023-06-06 19:33:23 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// copy src0, src1 to device if necessary
|
2024-02-25 10:09:09 +00:00
|
|
|
if (src1->backend == GGML_BACKEND_TYPE_GPU && src1_is_contiguous) {
|
2023-09-11 17:55:51 +00:00
|
|
|
if (id != g_main_device) {
|
|
|
|
if (convert_src1_to_q8_1) {
|
2023-12-26 20:23:59 +00:00
|
|
|
char * src1_ddq_i_source = dev[g_main_device].src1_ddq + src1_ddq_i_offset;
|
|
|
|
CUDA_CHECK(cudaMemcpyPeerAsync(src1_ddq_i, id, src1_ddq_i_source, g_main_device,
|
|
|
|
src1_ncols*src1_padded_col_size*q8_1_ts/q8_1_bs, stream));
|
2023-09-11 17:55:51 +00:00
|
|
|
} else {
|
2023-06-06 19:33:23 +00:00
|
|
|
float * src1_ddf_i_source = (float *) src1_extra->data_device[g_main_device];
|
2023-09-11 17:55:51 +00:00
|
|
|
src1_ddf_i_source += (i0*ne11 + src1_col_0) * ne10;
|
2023-12-26 20:23:59 +00:00
|
|
|
CUDA_CHECK(cudaMemcpyPeerAsync(src1_ddf_i, id, src1_ddf_i_source, g_main_device,
|
|
|
|
src1_ncols*ne10*sizeof(float), stream));
|
2023-06-06 19:33:23 +00:00
|
|
|
}
|
|
|
|
}
|
2024-02-25 10:09:09 +00:00
|
|
|
} else if (src1->backend == GGML_BACKEND_TYPE_CPU || (src1_on_device && !src1_is_contiguous)) {
|
2023-09-11 17:55:51 +00:00
|
|
|
CUDA_CHECK(ggml_cuda_cpy_tensor_2d(
|
2023-12-26 20:23:59 +00:00
|
|
|
src1_ddf_i, src1, i03, i02, src1_col_0, src1_col_0+src1_ncols, stream));
|
2023-09-11 17:55:51 +00:00
|
|
|
} else {
|
|
|
|
GGML_ASSERT(false);
|
2023-06-06 19:33:23 +00:00
|
|
|
}
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2024-02-25 10:09:09 +00:00
|
|
|
if (convert_src1_to_q8_1 && (src1->backend == GGML_BACKEND_TYPE_CPU || !src1_is_contiguous)) {
|
2023-09-11 17:55:51 +00:00
|
|
|
quantize_row_q8_1_cuda(src1_ddf_i, src1_ddq_i, ne10, src1_ncols, src1_padded_col_size, stream);
|
|
|
|
CUDA_CHECK(cudaGetLastError());
|
2023-06-06 19:33:23 +00:00
|
|
|
}
|
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
if (src1_col_0 == 0 && (!src0_on_device || !src0_is_contiguous) && i02 % i02_divisor == 0) {
|
2023-12-26 20:23:59 +00:00
|
|
|
CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src0_dd_i, src0, i03, i02/i02_divisor, dev[id].row_low, dev[id].row_high, stream));
|
2023-06-06 19:33:23 +00:00
|
|
|
}
|
2023-05-01 16:11:07 +00:00
|
|
|
|
2023-06-06 19:33:23 +00:00
|
|
|
// do the computation
|
2023-09-11 17:55:51 +00:00
|
|
|
op(src0, src1, dst, src0_dd_i, src1_ddf_i, src1_ddq_i, dst_dd_i,
|
2023-12-26 20:23:59 +00:00
|
|
|
dev[id].row_low, dev[id].row_high, src1_ncols, src1_padded_col_size, stream);
|
2023-07-01 19:49:44 +00:00
|
|
|
CUDA_CHECK(cudaGetLastError());
|
2023-06-06 19:33:23 +00:00
|
|
|
|
|
|
|
// copy dst to host or other device if necessary
|
|
|
|
if (!dst_on_device) {
|
|
|
|
void * dst_off_device;
|
|
|
|
cudaMemcpyKind kind;
|
2024-02-25 10:09:09 +00:00
|
|
|
if (dst->backend == GGML_BACKEND_TYPE_CPU) {
|
2023-06-06 19:33:23 +00:00
|
|
|
dst_off_device = dst->data;
|
|
|
|
kind = cudaMemcpyDeviceToHost;
|
2024-02-25 10:09:09 +00:00
|
|
|
} else if (dst->backend == GGML_BACKEND_TYPE_GPU) {
|
2023-06-06 19:33:23 +00:00
|
|
|
dst_off_device = dst_extra->data_device[g_main_device];
|
|
|
|
kind = cudaMemcpyDeviceToDevice;
|
|
|
|
} else {
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
if (split) {
|
|
|
|
// src0 = weight matrix is saved as a transposed matrix for better memory layout.
|
|
|
|
// dst is NOT transposed.
|
2023-07-29 21:04:10 +00:00
|
|
|
// The outputs of matrix matrix multiplications can therefore NOT simply be concatenated for >1 GPU.
|
2023-06-06 19:33:23 +00:00
|
|
|
// Instead they need to be copied to the correct slice in ne0 = dst row index.
|
|
|
|
// If dst is a vector with ne0 == 1 then you don't have to do this but it still produces correct results.
|
2023-09-11 17:55:51 +00:00
|
|
|
float * dhf_dst_i = (float *) ((char *) dst_off_device + i02*nb2 + i03*nb3);
|
|
|
|
GGML_ASSERT(dst->nb[1] == ne0*sizeof(float));
|
2023-12-26 20:23:59 +00:00
|
|
|
dhf_dst_i += src1_col_0*ne0 + dev[id].row_low;
|
|
|
|
#if !defined(GGML_USE_HIPBLAS)
|
|
|
|
if (kind == cudaMemcpyDeviceToDevice) {
|
|
|
|
// cudaMemcpy2DAsync may fail with copies between vmm pools of different devices
|
|
|
|
cudaMemcpy3DPeerParms p = {};
|
|
|
|
p.dstDevice = g_main_device;
|
|
|
|
p.dstPtr = make_cudaPitchedPtr(dhf_dst_i, ne0*sizeof(float), row_diff, src1_ncols);
|
|
|
|
p.srcDevice = id;
|
|
|
|
p.srcPtr = make_cudaPitchedPtr(dst_dd_i, row_diff*sizeof(float), row_diff, src1_ncols);
|
|
|
|
p.extent = make_cudaExtent(row_diff*sizeof(float), src1_ncols, 1);
|
|
|
|
CUDA_CHECK(cudaMemcpy3DPeerAsync(&p, stream));
|
|
|
|
} else
|
|
|
|
#endif
|
|
|
|
{
|
|
|
|
CUDA_CHECK(cudaMemcpy2DAsync(dhf_dst_i, ne0*sizeof(float),
|
|
|
|
dst_dd_i, row_diff*sizeof(float),
|
|
|
|
row_diff*sizeof(float), src1_ncols,
|
|
|
|
kind, stream));
|
|
|
|
}
|
2023-06-06 19:33:23 +00:00
|
|
|
} else {
|
|
|
|
float * dhf_dst_i = (float *) ((char *) dst_off_device + i02*nb2 + i03*nb3);
|
2023-09-11 17:55:51 +00:00
|
|
|
GGML_ASSERT(dst->nb[1] == ne0*sizeof(float));
|
|
|
|
dhf_dst_i += src1_col_0*ne0;
|
|
|
|
CUDA_CHECK(cudaMemcpyAsync(dhf_dst_i, dst_dd_i, src1_ncols*ne0*sizeof(float), kind, stream));
|
2023-06-06 19:33:23 +00:00
|
|
|
}
|
|
|
|
}
|
2023-07-01 19:49:44 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
// add event for the main device to wait on until other device is done
|
|
|
|
if (split && (id != g_main_device || is != 0)) {
|
|
|
|
CUDA_CHECK(cudaEventRecord(src0_extra->events[id][is], stream));
|
2023-07-01 19:49:44 +00:00
|
|
|
}
|
2023-06-06 19:33:23 +00:00
|
|
|
}
|
2023-05-01 16:11:07 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-07-01 19:49:44 +00:00
|
|
|
// main device waits for all other devices to be finished
|
|
|
|
if (split && g_device_count > 1) {
|
2023-09-11 17:55:51 +00:00
|
|
|
int64_t is_max = (ne11 + MUL_MAT_SRC1_COL_STRIDE - 1) / MUL_MAT_SRC1_COL_STRIDE;
|
|
|
|
is_max = is_max <= MAX_STREAMS ? is_max : MAX_STREAMS;
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
ggml_cuda_set_device(g_main_device);
|
|
|
|
for (int id = 0; id < g_device_count; ++id) {
|
|
|
|
if (dev[id].row_low == dev[id].row_high) {
|
2023-11-05 15:08:57 +00:00
|
|
|
continue;
|
|
|
|
}
|
2023-09-11 17:55:51 +00:00
|
|
|
for (int64_t is = 0; is < is_max; ++is) {
|
2023-09-16 14:55:43 +00:00
|
|
|
CUDA_CHECK(cudaStreamWaitEvent(g_cudaStreams[g_main_device][0], src0_extra->events[id][is], 0));
|
2023-07-01 19:49:44 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2024-02-25 10:09:09 +00:00
|
|
|
if (dst->backend == GGML_BACKEND_TYPE_CPU) {
|
2023-12-26 20:23:59 +00:00
|
|
|
ggml_cuda_set_device(g_main_device);
|
2023-07-01 19:49:44 +00:00
|
|
|
CUDA_CHECK(cudaDeviceSynchronize());
|
|
|
|
}
|
2023-05-01 16:11:07 +00:00
|
|
|
}
|
|
|
|
|
2023-10-08 17:19:14 +00:00
|
|
|
static void ggml_cuda_repeat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_repeat);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_cuda_get_rows(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_get_rows);
|
|
|
|
}
|
|
|
|
|
2023-09-28 16:04:36 +00:00
|
|
|
static void ggml_cuda_add(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
2023-09-11 17:55:51 +00:00
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_add);
|
2023-06-06 19:33:23 +00:00
|
|
|
}
|
|
|
|
|
2023-12-13 19:54:54 +00:00
|
|
|
static void ggml_cuda_acc(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_acc);
|
|
|
|
}
|
|
|
|
|
2023-09-28 16:04:36 +00:00
|
|
|
static void ggml_cuda_mul(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
2023-09-11 17:55:51 +00:00
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_mul);
|
2023-06-06 19:33:23 +00:00
|
|
|
}
|
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
static void ggml_cuda_div(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_div);
|
|
|
|
}
|
|
|
|
|
2023-09-28 16:04:36 +00:00
|
|
|
static void ggml_cuda_gelu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
2023-09-11 17:55:51 +00:00
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_gelu);
|
2023-07-12 17:26:18 +00:00
|
|
|
}
|
|
|
|
|
2023-09-28 16:04:36 +00:00
|
|
|
static void ggml_cuda_silu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
2023-09-11 17:55:51 +00:00
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_silu);
|
2023-06-06 19:33:23 +00:00
|
|
|
}
|
|
|
|
|
2023-12-13 19:54:54 +00:00
|
|
|
static void ggml_cuda_gelu_quick(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_gelu_quick);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_cuda_tanh(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_tanh);
|
|
|
|
}
|
|
|
|
|
2023-11-13 08:58:15 +00:00
|
|
|
static void ggml_cuda_relu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_relu);
|
|
|
|
}
|
|
|
|
|
2024-01-31 13:10:15 +00:00
|
|
|
static void ggml_cuda_hardsigmoid(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_hardsigmoid);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_cuda_hardswish(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_hardswish);
|
|
|
|
}
|
2023-12-13 19:54:54 +00:00
|
|
|
static void ggml_cuda_leaky_relu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_leaky_relu);
|
|
|
|
}
|
|
|
|
|
2023-11-13 08:58:15 +00:00
|
|
|
static void ggml_cuda_sqr(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_sqr);
|
|
|
|
}
|
|
|
|
|
2023-09-28 16:04:36 +00:00
|
|
|
static void ggml_cuda_norm(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
2023-09-11 17:55:51 +00:00
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_norm);
|
2023-07-11 19:53:34 +00:00
|
|
|
}
|
|
|
|
|
2023-12-13 19:54:54 +00:00
|
|
|
static void ggml_cuda_group_norm(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_group_norm);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_cuda_concat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_concat);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_cuda_upscale(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_upscale);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_cuda_pad(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_pad);
|
|
|
|
}
|
|
|
|
|
2023-09-28 16:04:36 +00:00
|
|
|
static void ggml_cuda_rms_norm(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
2023-09-11 17:55:51 +00:00
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_rms_norm);
|
cuda : loading models directly into VRAM, norm calculation on GPU, broadcasting for ggml_mul (#1483)
* Broadcasting for ggml_mul
* CUDA kernel for ggml_mul, norms in VRAM
* GPU weights not in RAM, direct loading with cuFile
* fixup! GPU weights not in RAM, direct loading with cuFile
* fixup! GPU weights not in RAM, direct loading with cuFile
* define default model path once, sync path with readme (#1366)
* ~7% faster Q5_1 AVX2 code (#1477)
* convert.py: Support models which are stored in a single pytorch_model.bin (#1469)
* Support models in a single pytorch_model.bin
* Remove spurious line with typo
* benchmark-matmul: Print the average of the test results (#1490)
* Remove unused n_parts parameter (#1509)
* Fixes #1511 lambda issue for w64devkit (mingw) (#1513)
* Fix for w64devkit and mingw
* make kv_f16 the default for api users (#1517)
* minor : fix compile warnings
* readme : adds WizardLM to the list of supported models (#1485)
* main : make reverse prompt option act as a stop token in non-interactive mode (#1032)
* Make reverse prompt option act as a stop token in non-interactive scenarios
* Making requested review changes
* Update gpt_params_parse and fix a merge error
* Revert "Update gpt_params_parse and fix a merge error"
This reverts commit 2bb2ff1748513591ad45b175a75ed1d8089d84c8.
* Update gpt_params_parse and fix a merge error take 2
* examples : add persistent chat (#1495)
* examples : add persistent chat
* examples : fix whitespace
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* tests : add missing header
* ggml : use F16 instead of F32 in Q4_0, Q4_1, Q8_0 (#1508)
* ggml : use F16 instead of F32 in Q4_0, Q4_1 and Q8_0
* llama : bump LLAMA_FILE_VERSION to 3
* cuda : update Q4 and Q8 dequantize kernels
* ggml : fix AVX dot products
* readme : update performance table + hot topics
* ggml : fix scalar implementation of Q4_1 dot
* llama : fix compile warnings in llama_set_state_data()
* llama : fix name shadowing and C4146 (#1526)
* Fix name shadowing and C4146
* Fix if macros not using defined when required
* Update llama-util.h
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
* Update llama-util.h
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
* Code style
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Fix for mingw (#1462)
* llama : add llama_init_backend() API (close #1527)
* feature : add blis and other BLAS implementation support (#1502)
* feature: add blis support
* feature: allow all BLA_VENDOR to be assigned in cmake arguments. align with whisper.cpp pr 927
* fix: version detection for BLA_SIZEOF_INTEGER, recover min version of cmake
* Fix typo in INTEGER
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Revert "feature : add blis and other BLAS implementation support (#1502)"
This reverts commit 07e9ace0f9da424d82e75df969642522880feb92.
* GPU weights not in RAM, direct loading with cuFile
* llama : code style fixes + progress print fix
* ggml : ggml_mul better broadcast support
* cmake : workarounds for cufile when CMake version < 3.25
* gg rebase fixup
* Loop in llama.cpp, fixed progress callback
* Attempt clang-tidy fix
* llama : fix vram size computation
* Add forgotten fclose()
---------
Co-authored-by: András Salamon <ott2@users.noreply.github.com>
Co-authored-by: Ilya Kurdyukov <59548320+ilyakurdyukov@users.noreply.github.com>
Co-authored-by: Tom Jobbins <784313+TheBloke@users.noreply.github.com>
Co-authored-by: rankaiyx <rankaiyx@rankaiyx.com>
Co-authored-by: Stephan Walter <stephan@walter.name>
Co-authored-by: DannyDaemonic <DannyDaemonic@gmail.com>
Co-authored-by: Erik Scholz <Green-Sky@users.noreply.github.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: David Kennedy <dakennedyd@gmail.com>
Co-authored-by: Jason McCartney <jmac@theroot.org>
Co-authored-by: Evan Jones <evan.q.jones@gmail.com>
Co-authored-by: Maxime <672982+maximegmd@users.noreply.github.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Zenix <zenixls2@gmail.com>
2023-05-20 12:19:28 +00:00
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
|
2023-12-07 20:26:54 +00:00
|
|
|
if (!g_cublas_loaded) return false;
|
2023-11-07 06:49:08 +00:00
|
|
|
|
2023-05-01 16:11:07 +00:00
|
|
|
const int64_t ne10 = src1->ne[0];
|
|
|
|
|
|
|
|
const int64_t ne0 = dst->ne[0];
|
|
|
|
const int64_t ne1 = dst->ne[1];
|
|
|
|
|
|
|
|
// TODO: find the optimal values for these
|
2023-09-28 19:42:38 +00:00
|
|
|
return (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
|
|
|
|
src1->type == GGML_TYPE_F32 &&
|
|
|
|
dst->type == GGML_TYPE_F32 &&
|
|
|
|
(ne0 >= 32 && ne1 >= 32 && ne10 >= 32);
|
2023-05-01 16:11:07 +00:00
|
|
|
}
|
|
|
|
|
2023-09-28 16:04:36 +00:00
|
|
|
static void ggml_cuda_mul_mat_vec_p021(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst){
|
2023-06-14 17:47:19 +00:00
|
|
|
GGML_ASSERT(ggml_is_permuted(src0) && ggml_is_permuted(src1));
|
2024-02-25 10:09:09 +00:00
|
|
|
GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
|
2023-06-14 17:47:19 +00:00
|
|
|
GGML_ASSERT(src0->nb[0] <= src0->nb[1] && src0->nb[2] <= src0->nb[3]); // 0213 permutation
|
|
|
|
GGML_ASSERT(src1->nb[0] <= src1->nb[1] && src1->nb[2] <= src1->nb[3]); // 0213 permutation
|
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F16);
|
|
|
|
GGML_ASSERT(src1->type == GGML_TYPE_F32);
|
|
|
|
|
|
|
|
const int64_t ne00 = src0->ne[0];
|
|
|
|
const int64_t ne01 = src0->ne[1];
|
|
|
|
const int64_t ne02 = src0->ne[2];
|
|
|
|
|
2023-07-23 12:09:47 +00:00
|
|
|
const int64_t ne12 = src1->ne[2];
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
ggml_cuda_set_device(g_main_device);
|
2023-09-11 17:55:51 +00:00
|
|
|
cudaStream_t main_stream = g_cudaStreams[g_main_device][0];
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2023-10-08 17:19:14 +00:00
|
|
|
ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
|
2023-06-14 17:47:19 +00:00
|
|
|
void * src0_ddq = src0_extra->data_device[g_main_device];
|
|
|
|
|
2023-10-08 17:19:14 +00:00
|
|
|
ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
|
2023-06-14 17:47:19 +00:00
|
|
|
float * src1_ddf = (float *) src1_extra->data_device[g_main_device];
|
|
|
|
|
2023-10-08 17:19:14 +00:00
|
|
|
ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
|
2023-06-14 17:47:19 +00:00
|
|
|
float * dst_ddf = (float *) dst_extra->data_device[g_main_device];
|
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
ggml_mul_mat_p021_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, ne02, ne12, main_stream);
|
2023-06-14 17:47:19 +00:00
|
|
|
}
|
|
|
|
|
2023-09-28 16:04:36 +00:00
|
|
|
static void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst){
|
2023-10-24 13:48:37 +00:00
|
|
|
GGML_ASSERT(!ggml_is_transposed(src0));
|
|
|
|
GGML_ASSERT(!ggml_is_transposed(src1));
|
2023-06-14 17:47:19 +00:00
|
|
|
GGML_ASSERT(!ggml_is_permuted(src0));
|
2024-02-25 10:09:09 +00:00
|
|
|
GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
|
2023-06-14 17:47:19 +00:00
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F16);
|
|
|
|
GGML_ASSERT(src1->type == GGML_TYPE_F32);
|
|
|
|
|
|
|
|
const int64_t ne00 = src0->ne[0];
|
|
|
|
const int64_t ne01 = src0->ne[1];
|
|
|
|
const int64_t ne02 = src0->ne[2];
|
|
|
|
|
|
|
|
const int64_t nb01 = src0->nb[1];
|
|
|
|
const int64_t nb02 = src0->nb[2];
|
|
|
|
|
2023-10-24 13:48:37 +00:00
|
|
|
const int64_t ne12 = src1->ne[2];
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
ggml_cuda_set_device(g_main_device);
|
2023-09-11 17:55:51 +00:00
|
|
|
cudaStream_t main_stream = g_cudaStreams[g_main_device][0];
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2023-10-08 17:19:14 +00:00
|
|
|
ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
|
2023-06-14 17:47:19 +00:00
|
|
|
void * src0_ddq = src0_extra->data_device[g_main_device];
|
|
|
|
|
2023-10-08 17:19:14 +00:00
|
|
|
ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
|
2023-06-14 17:47:19 +00:00
|
|
|
float * src1_ddf = (float *) src1_extra->data_device[g_main_device];
|
|
|
|
|
2023-10-08 17:19:14 +00:00
|
|
|
ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
|
2023-06-14 17:47:19 +00:00
|
|
|
float * dst_ddf = (float *) dst_extra->data_device[g_main_device];
|
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
const int64_t row_stride_x = nb01 / sizeof(half);
|
|
|
|
const int64_t channel_stride_x = nb02 / sizeof(half);
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
ggml_mul_mat_vec_nc_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x, main_stream);
|
2023-06-14 17:47:19 +00:00
|
|
|
}
|
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
static __global__ void k_compute_batched_ptrs(
|
2023-12-18 17:27:47 +00:00
|
|
|
const half * src0_as_f16, const half * src1_as_f16, char * dst,
|
2023-11-02 18:32:11 +00:00
|
|
|
const void ** ptrs_src, void ** ptrs_dst,
|
2023-12-18 17:27:47 +00:00
|
|
|
int64_t ne12, int64_t ne13,
|
|
|
|
int64_t ne23,
|
|
|
|
size_t nb02, size_t nb03,
|
|
|
|
size_t nb12, size_t nb13,
|
|
|
|
size_t nbd2, size_t nbd3,
|
|
|
|
int64_t r2, int64_t r3) {
|
|
|
|
int64_t i13 = blockIdx.x * blockDim.x + threadIdx.x;
|
|
|
|
int64_t i12 = blockIdx.y * blockDim.y + threadIdx.y;
|
2023-11-01 22:10:09 +00:00
|
|
|
|
|
|
|
if (i13 >= ne13 || i12 >= ne12) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2023-12-18 17:27:47 +00:00
|
|
|
int64_t i03 = i13 / r3;
|
|
|
|
int64_t i02 = i12 / r2;
|
2023-11-01 22:10:09 +00:00
|
|
|
|
2023-12-29 08:32:31 +00:00
|
|
|
ptrs_src[0*ne23 + i12 + i13*ne12] = (const char *) src0_as_f16 + i02*nb02 + i03*nb03;
|
|
|
|
ptrs_src[1*ne23 + i12 + i13*ne12] = (const char *) src1_as_f16 + i12*nb12 + i13*nb13;
|
|
|
|
ptrs_dst[0*ne23 + i12 + i13*ne12] = ( char *) dst + i12*nbd2 + i13*nbd3;
|
2023-11-01 22:10:09 +00:00
|
|
|
}
|
|
|
|
|
2024-02-11 18:08:39 +00:00
|
|
|
static void ggml_cuda_mul_mat_batched_cublas(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
2023-10-24 13:48:37 +00:00
|
|
|
GGML_ASSERT(!ggml_is_transposed(src0));
|
|
|
|
GGML_ASSERT(!ggml_is_transposed(src1));
|
2023-10-27 14:01:23 +00:00
|
|
|
|
2024-02-25 10:09:09 +00:00
|
|
|
GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
|
2023-10-24 13:48:37 +00:00
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F16);
|
|
|
|
|
2023-12-29 08:32:31 +00:00
|
|
|
GGML_TENSOR_BINARY_OP_LOCALS
|
2023-10-24 13:48:37 +00:00
|
|
|
|
2023-12-29 08:32:31 +00:00
|
|
|
const int64_t ne_dst = ggml_nelements(dst);
|
2023-10-24 13:48:37 +00:00
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
ggml_cuda_set_device(g_main_device);
|
2023-10-24 13:48:37 +00:00
|
|
|
cudaStream_t main_stream = g_cudaStreams[g_main_device][0];
|
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
CUBLAS_CHECK(cublasSetStream(g_cublas_handles[g_main_device], main_stream));
|
2023-10-24 13:48:37 +00:00
|
|
|
|
|
|
|
ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
|
|
|
|
void * src0_ddq = src0_extra->data_device[g_main_device];
|
2023-12-29 08:32:31 +00:00
|
|
|
half * src0_f16 = (half *) src0_ddq;
|
2023-10-24 13:48:37 +00:00
|
|
|
|
|
|
|
ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
|
|
|
|
float * src1_ddf = (float *) src1_extra->data_device[g_main_device];
|
|
|
|
|
|
|
|
ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
|
|
|
|
float * dst_ddf = (float *) dst_extra->data_device[g_main_device];
|
|
|
|
|
|
|
|
// convert src1 to fp16
|
2023-12-29 08:32:31 +00:00
|
|
|
cuda_pool_alloc<half> src1_f16_alloc;
|
|
|
|
if (src1->type != GGML_TYPE_F16) {
|
|
|
|
const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
|
|
|
|
const int64_t ne_src1 = ggml_nelements(src1);
|
|
|
|
src1_f16_alloc.alloc(ne_src1);
|
|
|
|
GGML_ASSERT(to_fp16_cuda != nullptr);
|
|
|
|
to_fp16_cuda(src1_ddf, src1_f16_alloc.get(), ne_src1, main_stream);
|
|
|
|
}
|
|
|
|
half * src1_f16 = src1->type == GGML_TYPE_F16 ? (half *) src1_ddf : src1_f16_alloc.get();
|
2023-10-24 13:48:37 +00:00
|
|
|
|
2023-12-24 13:34:22 +00:00
|
|
|
cuda_pool_alloc<half> dst_f16;
|
|
|
|
char * dst_t;
|
2023-12-18 17:27:47 +00:00
|
|
|
|
|
|
|
cublasComputeType_t cu_compute_type = CUBLAS_COMPUTE_16F;
|
|
|
|
cudaDataType_t cu_data_type = CUDA_R_16F;
|
|
|
|
|
|
|
|
// dst strides
|
|
|
|
size_t nbd2 = dst->nb[2];
|
|
|
|
size_t nbd3 = dst->nb[3];
|
|
|
|
|
|
|
|
const half alpha_f16 = 1.0f;
|
|
|
|
const half beta_f16 = 0.0f;
|
|
|
|
|
|
|
|
const float alpha_f32 = 1.0f;
|
|
|
|
const float beta_f32 = 0.0f;
|
|
|
|
|
|
|
|
const void * alpha = &alpha_f16;
|
|
|
|
const void * beta = &beta_f16;
|
|
|
|
|
|
|
|
if (dst->op_params[0] == GGML_PREC_DEFAULT) {
|
2023-12-29 08:32:31 +00:00
|
|
|
dst_t = (char *) dst_f16.alloc(ne_dst);
|
2023-12-18 17:27:47 +00:00
|
|
|
|
|
|
|
nbd2 /= sizeof(float) / sizeof(half);
|
|
|
|
nbd3 /= sizeof(float) / sizeof(half);
|
|
|
|
} else {
|
|
|
|
dst_t = (char *) dst_ddf;
|
|
|
|
|
|
|
|
cu_compute_type = CUBLAS_COMPUTE_32F;
|
|
|
|
cu_data_type = CUDA_R_32F;
|
|
|
|
|
|
|
|
alpha = &alpha_f32;
|
|
|
|
beta = &beta_f32;
|
|
|
|
}
|
2023-10-24 13:48:37 +00:00
|
|
|
|
|
|
|
GGML_ASSERT(ne12 % ne02 == 0);
|
|
|
|
GGML_ASSERT(ne13 % ne03 == 0);
|
|
|
|
|
|
|
|
// broadcast factors
|
|
|
|
const int64_t r2 = ne12/ne02;
|
|
|
|
const int64_t r3 = ne13/ne03;
|
|
|
|
|
|
|
|
#if 0
|
|
|
|
// use cublasGemmEx
|
|
|
|
{
|
|
|
|
for (int i13 = 0; i13 < ne13; ++i13) {
|
|
|
|
for (int i12 = 0; i12 < ne12; ++i12) {
|
|
|
|
int i03 = i13 / r3;
|
|
|
|
int i02 = i12 / r2;
|
|
|
|
|
|
|
|
CUBLAS_CHECK(
|
2023-12-18 17:27:47 +00:00
|
|
|
cublasGemmEx(g_cublas_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N,
|
2023-10-24 13:48:37 +00:00
|
|
|
ne01, ne11, ne10,
|
2023-12-18 17:27:47 +00:00
|
|
|
alpha, (const char *) src0_as_f16 + i02*src0->nb[2] + i03*src0->nb[3] , CUDA_R_16F, nb01/sizeof(half),
|
|
|
|
(const char *) src1_as_f16 + i12*src1->nb[2]/2 + i13*src1->nb[3]/2, CUDA_R_16F, nb11/sizeof(float),
|
|
|
|
beta, ( char *) dst_t + i12*nbd2 + i13*nbd3, cu_data_type, ne01,
|
|
|
|
cu_compute_type,
|
2023-10-24 13:48:37 +00:00
|
|
|
CUBLAS_GEMM_DEFAULT_TENSOR_OP));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
|
|
if (r2 == 1 && r3 == 1 && src0->nb[2]*src0->ne[2] == src0->nb[3] && src1->nb[2]*src1->ne[2] == src1->nb[3]) {
|
|
|
|
// there is no broadcast and src0, src1 are contiguous across dims 2, 3
|
|
|
|
// use cublasGemmStridedBatchedEx
|
|
|
|
CUBLAS_CHECK(
|
2023-12-07 20:26:54 +00:00
|
|
|
cublasGemmStridedBatchedEx(g_cublas_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N,
|
2023-10-24 13:48:37 +00:00
|
|
|
ne01, ne11, ne10,
|
2023-12-29 08:32:31 +00:00
|
|
|
alpha, (const char *) src0_f16, CUDA_R_16F, nb01/nb00, nb02/nb00, // strideA
|
|
|
|
(const char *) src1_f16, CUDA_R_16F, nb11/nb10, nb12/nb10, // strideB
|
|
|
|
beta, ( char *) dst_t, cu_data_type, ne01, nb2/nb0, // strideC
|
2023-10-24 13:48:37 +00:00
|
|
|
ne12*ne13,
|
2023-12-18 17:27:47 +00:00
|
|
|
cu_compute_type,
|
2023-10-24 13:48:37 +00:00
|
|
|
CUBLAS_GEMM_DEFAULT_TENSOR_OP));
|
|
|
|
} else {
|
|
|
|
// use cublasGemmBatchedEx
|
|
|
|
const int ne23 = ne12*ne13;
|
2023-11-02 18:32:11 +00:00
|
|
|
|
2023-12-24 13:34:22 +00:00
|
|
|
cuda_pool_alloc<const void *> ptrs_src(2*ne23);
|
|
|
|
cuda_pool_alloc< void *> ptrs_dst(1*ne23);
|
2023-11-01 22:10:09 +00:00
|
|
|
|
|
|
|
dim3 block_dims(ne13, ne12);
|
|
|
|
k_compute_batched_ptrs<<<1, block_dims, 0, main_stream>>>(
|
2023-12-29 08:32:31 +00:00
|
|
|
src0_f16, src1_f16, dst_t,
|
2023-12-24 13:34:22 +00:00
|
|
|
ptrs_src.get(), ptrs_dst.get(),
|
2023-11-01 22:10:09 +00:00
|
|
|
ne12, ne13,
|
|
|
|
ne23,
|
|
|
|
nb02, nb03,
|
2023-12-29 08:32:31 +00:00
|
|
|
src1->type == GGML_TYPE_F16 ? nb12 : nb12/2,
|
|
|
|
src1->type == GGML_TYPE_F16 ? nb13 : nb13/2,
|
2023-12-18 17:27:47 +00:00
|
|
|
nbd2, nbd3,
|
2023-11-01 22:10:09 +00:00
|
|
|
r2, r3);
|
|
|
|
CUDA_CHECK(cudaGetLastError());
|
2023-11-05 07:12:13 +00:00
|
|
|
|
2023-10-24 13:48:37 +00:00
|
|
|
CUBLAS_CHECK(
|
2023-12-07 20:26:54 +00:00
|
|
|
cublasGemmBatchedEx(g_cublas_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N,
|
2023-10-24 13:48:37 +00:00
|
|
|
ne01, ne11, ne10,
|
2023-12-29 08:32:31 +00:00
|
|
|
alpha, (const void **) (ptrs_src.get() + 0*ne23), CUDA_R_16F, nb01/nb00,
|
|
|
|
(const void **) (ptrs_src.get() + 1*ne23), CUDA_R_16F, nb11/nb10,
|
2023-12-24 13:34:22 +00:00
|
|
|
beta, ( void **) (ptrs_dst.get() + 0*ne23), cu_data_type, ne01,
|
2023-10-24 13:48:37 +00:00
|
|
|
ne23,
|
2023-12-18 17:27:47 +00:00
|
|
|
cu_compute_type,
|
2023-10-24 13:48:37 +00:00
|
|
|
CUBLAS_GEMM_DEFAULT_TENSOR_OP));
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
2023-12-18 17:27:47 +00:00
|
|
|
if (dst->op_params[0] == GGML_PREC_DEFAULT) {
|
|
|
|
const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
|
2023-12-29 08:32:31 +00:00
|
|
|
to_fp32_cuda(dst_f16.get(), dst_ddf, ne_dst, main_stream);
|
2023-12-18 17:27:47 +00:00
|
|
|
}
|
2023-10-24 13:48:37 +00:00
|
|
|
}
|
|
|
|
|
2023-09-28 16:04:36 +00:00
|
|
|
static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
2023-10-27 14:01:23 +00:00
|
|
|
const bool all_on_device =
|
2024-02-25 10:09:09 +00:00
|
|
|
(src0->backend == GGML_BACKEND_TYPE_GPU || src0->backend == GGML_BACKEND_TYPE_GPU_SPLIT) &&
|
|
|
|
(src1->backend == GGML_BACKEND_TYPE_GPU) &&
|
|
|
|
( dst->backend == GGML_BACKEND_TYPE_GPU);
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2024-02-25 10:09:09 +00:00
|
|
|
const bool split = src0->backend == GGML_BACKEND_TYPE_GPU_SPLIT;
|
2023-11-05 17:45:16 +00:00
|
|
|
|
2023-09-11 17:55:51 +00:00
|
|
|
int64_t min_compute_capability = INT_MAX;
|
2024-01-12 19:07:38 +00:00
|
|
|
|
2024-02-11 18:08:39 +00:00
|
|
|
bool any_pascal_with_slow_fp16 = false;
|
2024-01-12 19:07:38 +00:00
|
|
|
if (split) {
|
|
|
|
ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) src0->buffer->buft->context;
|
|
|
|
auto & tensor_split = buft_ctx->tensor_split;
|
|
|
|
for (int id = 0; id < g_device_count; ++id) {
|
2024-02-11 18:08:39 +00:00
|
|
|
// skip devices that are not going to do any work:
|
|
|
|
if (tensor_split[id] >= (id + 1 < g_device_count ? tensor_split[id + 1] : 1.0f)) {
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (min_compute_capability > g_device_caps[id].cc) {
|
2024-01-12 19:07:38 +00:00
|
|
|
min_compute_capability = g_device_caps[id].cc;
|
|
|
|
}
|
2024-02-11 18:08:39 +00:00
|
|
|
if (g_device_caps[id].cc == 610) {
|
|
|
|
any_pascal_with_slow_fp16 = true;
|
|
|
|
}
|
2023-09-11 17:55:51 +00:00
|
|
|
}
|
2024-01-12 19:07:38 +00:00
|
|
|
} else {
|
2024-02-11 18:08:39 +00:00
|
|
|
min_compute_capability = g_device_caps[g_main_device].cc;
|
|
|
|
any_pascal_with_slow_fp16 = g_device_caps[g_main_device].cc == 610;
|
2023-09-11 17:55:51 +00:00
|
|
|
}
|
|
|
|
|
2024-02-11 18:08:39 +00:00
|
|
|
// check data types and tensor shapes for custom matrix multiplication kernels:
|
|
|
|
bool use_dequantize_mul_mat_vec = (ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F16)
|
|
|
|
&& src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
|
|
|
|
&& src0->ne[0] % GGML_CUDA_DMMV_X == 0 && src1->ne[1] == 1;
|
|
|
|
|
|
|
|
bool use_mul_mat_vec_q = ggml_is_quantized(src0->type)
|
|
|
|
&& src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
|
|
|
|
&& src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
|
|
|
|
|
|
|
|
bool use_mul_mat_q = ggml_cuda_supports_mmq(src0->type)
|
|
|
|
&& src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
|
|
|
|
|
2023-12-29 22:12:53 +00:00
|
|
|
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
|
|
|
|
2023-12-30 12:52:01 +00:00
|
|
|
const bool fp16_performance_good = min_compute_capability >= CC_RDNA1;
|
2024-02-11 18:08:39 +00:00
|
|
|
|
2023-12-30 12:52:01 +00:00
|
|
|
#ifdef CUDA_USE_TENSOR_CORES
|
|
|
|
use_mul_mat_q = use_mul_mat_q && min_compute_capability < CC_RDNA3;
|
|
|
|
#endif // CUDA_USE_TENSOR_CORES
|
2023-12-29 22:12:53 +00:00
|
|
|
|
|
|
|
#else
|
|
|
|
|
2024-02-11 18:08:39 +00:00
|
|
|
// fp16 performance is good on Volta or newer and on P100 (compute capability 6.0)
|
|
|
|
const bool fp16_performance_good = min_compute_capability >= CC_PASCAL && !any_pascal_with_slow_fp16;
|
|
|
|
|
|
|
|
// mmvq and mmq need the __dp4a instruction which on NVIDIA is only available for CC >= 6.1
|
|
|
|
use_mul_mat_vec_q = use_mul_mat_vec_q && min_compute_capability >= MIN_CC_DP4A;
|
|
|
|
use_mul_mat_q = use_mul_mat_q && min_compute_capability >= MIN_CC_DP4A;
|
|
|
|
|
2023-12-29 22:12:53 +00:00
|
|
|
#ifdef CUDA_USE_TENSOR_CORES
|
|
|
|
// when tensor cores are available, use them for large batch size
|
|
|
|
// ref: https://github.com/ggerganov/llama.cpp/pull/3776
|
2024-02-11 18:08:39 +00:00
|
|
|
use_mul_mat_q = use_mul_mat_q && (!fp16_performance_good || src1->ne[1] <= MMQ_MAX_BATCH_SIZE);
|
2023-12-29 22:12:53 +00:00
|
|
|
#endif // CUDA_USE_TENSOR_CORES
|
|
|
|
|
|
|
|
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
|
2023-10-27 14:01:23 +00:00
|
|
|
|
2024-02-11 18:08:39 +00:00
|
|
|
// if mmvq is available it's a better choice than dmmv:
|
|
|
|
#ifndef GGML_CUDA_FORCE_DMMV
|
|
|
|
use_dequantize_mul_mat_vec = use_dequantize_mul_mat_vec && !use_mul_mat_vec_q;
|
|
|
|
#endif // GGML_CUDA_FORCE_DMMV
|
2024-01-08 15:02:32 +00:00
|
|
|
|
2023-10-24 13:48:37 +00:00
|
|
|
// debug helpers
|
|
|
|
//printf("src0: %8d %8d %8d %8d\n", src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3]);
|
|
|
|
//printf(" %8d %8d %8d %8d\n", src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3]);
|
|
|
|
//printf("src1: %8d %8d %8d %8d\n", src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3]);
|
|
|
|
//printf(" %8d %8d %8d %8d\n", src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3]);
|
|
|
|
//printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
|
|
|
|
//printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
|
|
|
|
|
2023-12-29 22:12:53 +00:00
|
|
|
if (!split && all_on_device && !fp16_performance_good && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
|
2023-10-25 07:26:27 +00:00
|
|
|
// KQ single-batch
|
2023-06-14 17:47:19 +00:00
|
|
|
ggml_cuda_mul_mat_vec_p021(src0, src1, dst);
|
2023-12-29 22:12:53 +00:00
|
|
|
} else if (!split && all_on_device && !fp16_performance_good && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
|
2023-10-25 07:26:27 +00:00
|
|
|
// KQV single-batch
|
2023-06-14 17:47:19 +00:00
|
|
|
ggml_cuda_mul_mat_vec_nc(src0, src1, dst);
|
2024-01-12 19:07:38 +00:00
|
|
|
} else if (!split && all_on_device && fp16_performance_good && src0->type == GGML_TYPE_F16 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
|
2023-10-25 07:26:27 +00:00
|
|
|
// KQ + KQV multi-batch
|
2024-02-11 18:08:39 +00:00
|
|
|
ggml_cuda_mul_mat_batched_cublas(src0, src1, dst);
|
|
|
|
} else if (use_dequantize_mul_mat_vec) {
|
|
|
|
ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, false);
|
|
|
|
} else if (use_mul_mat_vec_q) {
|
|
|
|
ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_vec_q, true);
|
|
|
|
} else if (use_mul_mat_q) {
|
|
|
|
ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_q, true);
|
2023-06-06 19:33:23 +00:00
|
|
|
} else {
|
2024-02-11 18:08:39 +00:00
|
|
|
ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, false);
|
2023-06-06 19:33:23 +00:00
|
|
|
}
|
|
|
|
}
|
2023-05-01 16:11:07 +00:00
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
#if 0
|
|
|
|
template<typename ... Srcs>
|
|
|
|
static __global__ void k_compute_batched_ptrs_id(
|
|
|
|
const void ** ptrs_src, void ** ptrs_dst,
|
|
|
|
int ne12, int ne13,
|
|
|
|
int ne23,
|
|
|
|
int nb02, int nb03,
|
|
|
|
int nb12, int nb13,
|
|
|
|
int nb2, int nb3,
|
|
|
|
int r2, int r3,
|
|
|
|
ggml_type src0_type, half * src0_as_f16, int64_t src0_ne,
|
|
|
|
const half * src1_f16, half * dst_f16,
|
|
|
|
const int32_t * ids, const int id,
|
|
|
|
Srcs... src0s) {
|
|
|
|
|
|
|
|
int i = ids[id];
|
|
|
|
|
|
|
|
half * src0_f16;
|
|
|
|
const void * srcs_ar[] = { (const half *) src0s... };
|
|
|
|
if (src0_type == GGML_TYPE_F16) {
|
|
|
|
src0_f16 = (half *) srcs_ar[i];
|
|
|
|
} else {
|
|
|
|
src0_f16 = src0_as_f16;
|
|
|
|
if (threadIdx.x == 0 && threadIdx.y == 0) {
|
|
|
|
const to_fp16_cuda_t to_fp16 = ggml_get_to_fp16_cuda(src0_type);
|
|
|
|
to_fp16(srcs_ar[i], src0_f16, src0_ne, cudaStreamFireAndForget);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
int i13 = blockIdx.x * blockDim.x + threadIdx.x;
|
|
|
|
int i12 = blockIdx.y * blockDim.y + threadIdx.y;
|
|
|
|
|
|
|
|
if (i13 >= ne13 || i12 >= ne12) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
int i03 = i13 / r3;
|
|
|
|
int i02 = i12 / r2;
|
|
|
|
|
|
|
|
ptrs_src[0*ne23 + i12 + i13*ne12] = (const char *) src0_f16 + i02*nb02 + i03*nb03;
|
|
|
|
ptrs_src[1*ne23 + i12 + i13*ne12] = (const char *) src1_f16 + i12*nb12/2 + i13*nb13/2;
|
|
|
|
ptrs_dst[0*ne23 + i12 + i13*ne12] = ( char *) dst_f16 + i12* nb2/2 + i13* nb3/2;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_cuda_mul_mat_id_cublas(ggml_tensor * dst) {
|
|
|
|
const struct ggml_tensor * ids = dst->src[0];
|
|
|
|
const struct ggml_tensor * src1 = dst->src[1];
|
|
|
|
const struct ggml_tensor * src00 = dst->src[2];
|
|
|
|
|
|
|
|
const int id = dst->op_params[0];
|
|
|
|
|
|
|
|
GGML_ASSERT(!ggml_is_transposed(src00));
|
|
|
|
GGML_ASSERT(!ggml_is_transposed(src1));
|
|
|
|
|
2024-02-25 10:09:09 +00:00
|
|
|
GGML_ASSERT(src00->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
|
2023-12-07 20:26:54 +00:00
|
|
|
GGML_ASSERT(src1->type == GGML_TYPE_F32);
|
|
|
|
|
|
|
|
const int64_t ne00 = src00->ne[0]; GGML_UNUSED(ne00);
|
|
|
|
const int64_t ne01 = src00->ne[1];
|
|
|
|
const int64_t ne02 = src00->ne[2];
|
|
|
|
const int64_t ne03 = src00->ne[3];
|
|
|
|
|
|
|
|
//const int64_t nb01 = src00->nb[1];
|
|
|
|
const int64_t nb02 = src00->nb[2]; GGML_UNUSED(nb02);
|
|
|
|
const int64_t nb03 = src00->nb[3]; GGML_UNUSED(nb03);
|
|
|
|
|
|
|
|
const int64_t ne10 = src1->ne[0];
|
|
|
|
const int64_t ne11 = src1->ne[1];
|
|
|
|
const int64_t ne12 = src1->ne[2];
|
|
|
|
const int64_t ne13 = src1->ne[3];
|
|
|
|
|
|
|
|
//const int64_t nb11 = src1->nb[1];
|
|
|
|
const int64_t nb12 = src1->nb[2]; GGML_UNUSED(nb12);
|
|
|
|
const int64_t nb13 = src1->nb[3]; GGML_UNUSED(nb13);
|
|
|
|
|
|
|
|
const int64_t ne1 = ggml_nelements(src1);
|
|
|
|
const int64_t ne = ggml_nelements(dst);
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
ggml_cuda_set_device(g_main_device);
|
2023-12-07 20:26:54 +00:00
|
|
|
cudaStream_t main_stream = g_cudaStreams[g_main_device][0];
|
|
|
|
|
|
|
|
CUBLAS_CHECK(cublasSetStream(g_cublas_handles[g_main_device], main_stream));
|
|
|
|
|
|
|
|
//ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
|
|
|
|
//void * src0_ddq = src0_extra->data_device[g_main_device];
|
|
|
|
//half * src0_as_f16 = (half *) src0_ddq;
|
|
|
|
|
|
|
|
ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
|
|
|
|
float * src1_ddf = (float *) src1_extra->data_device[g_main_device];
|
|
|
|
|
|
|
|
ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
|
|
|
|
float * dst_ddf = (float *) dst_extra->data_device[g_main_device];
|
|
|
|
|
|
|
|
// convert src1 to fp16
|
|
|
|
const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
|
|
|
|
GGML_ASSERT(to_fp16_cuda != nullptr);
|
|
|
|
|
|
|
|
size_t src1_as = 0;
|
|
|
|
half * src1_as_f16 = (half *) ggml_cuda_pool_malloc(ne1 * sizeof(half), &src1_as);
|
|
|
|
to_fp16_cuda(src1_ddf, src1_as_f16, ne1, main_stream);
|
|
|
|
|
|
|
|
size_t dst_as = 0;
|
|
|
|
half * dst_f16 = (half *) ggml_cuda_pool_malloc(ne * sizeof(half), &dst_as);
|
|
|
|
|
|
|
|
GGML_ASSERT(ne12 % ne02 == 0);
|
|
|
|
GGML_ASSERT(ne13 % ne03 == 0);
|
|
|
|
|
|
|
|
// broadcast factors
|
|
|
|
const int64_t r2 = ne12/ne02;
|
|
|
|
const int64_t r3 = ne13/ne03;
|
|
|
|
|
|
|
|
const half alpha_f16 = 1.0f;
|
|
|
|
const half beta_f16 = 0.0f;
|
|
|
|
|
|
|
|
// use cublasGemmBatchedEx
|
|
|
|
const int ne23 = ne12*ne13;
|
|
|
|
|
|
|
|
const void ** ptrs_src = nullptr;
|
|
|
|
void ** ptrs_dst = nullptr;
|
|
|
|
|
|
|
|
size_t ptrs_src_s = 0;
|
|
|
|
size_t ptrs_dst_s = 0;
|
|
|
|
|
|
|
|
ptrs_src = (const void **) ggml_cuda_pool_malloc(2*ne23*sizeof(void *), &ptrs_src_s);
|
|
|
|
ptrs_dst = ( void **) ggml_cuda_pool_malloc(1*ne23*sizeof(void *), &ptrs_dst_s);
|
|
|
|
|
|
|
|
int64_t src0_ne = ggml_nelements(src00);
|
|
|
|
half * src0_as_f16 = nullptr;
|
|
|
|
size_t src0_as = 0;
|
|
|
|
if (src00->type != GGML_TYPE_F16) {
|
|
|
|
src0_as_f16 = (half *) ggml_cuda_pool_malloc(src0_ne * sizeof(half), &src0_as);
|
|
|
|
}
|
|
|
|
|
|
|
|
static_assert(GGML_MAX_SRC == 6, "GGML_MAX_SRC == 6");
|
|
|
|
dim3 block_dims(ne13, ne12);
|
|
|
|
k_compute_batched_ptrs_id<<<1, block_dims, 0, main_stream>>>(
|
|
|
|
ptrs_src, ptrs_dst,
|
|
|
|
ne12, ne13,
|
|
|
|
ne23,
|
|
|
|
ne00*ne01*sizeof(half), ne00*ne01*ne02*sizeof(half),
|
|
|
|
nb12, nb13,
|
|
|
|
dst->nb[2], dst->nb[3],
|
|
|
|
r2, r3,
|
|
|
|
src00->type, src0_as_f16, src0_ne,
|
|
|
|
src1_as_f16, dst_f16,
|
|
|
|
(const int *)((ggml_tensor_extra_gpu *)ids->extra)->data_device[g_main_device], id,
|
|
|
|
dst->src[2] ? (const half *)((ggml_tensor_extra_gpu *)dst->src[2]->extra)->data_device[g_main_device] : nullptr,
|
|
|
|
dst->src[3] ? (const half *)((ggml_tensor_extra_gpu *)dst->src[3]->extra)->data_device[g_main_device] : nullptr,
|
|
|
|
dst->src[4] ? (const half *)((ggml_tensor_extra_gpu *)dst->src[4]->extra)->data_device[g_main_device] : nullptr,
|
|
|
|
dst->src[5] ? (const half *)((ggml_tensor_extra_gpu *)dst->src[5]->extra)->data_device[g_main_device] : nullptr
|
|
|
|
);
|
|
|
|
CUDA_CHECK(cudaGetLastError());
|
|
|
|
|
|
|
|
CUBLAS_CHECK(
|
|
|
|
cublasGemmBatchedEx(g_cublas_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N,
|
|
|
|
ne01, ne11, ne10,
|
|
|
|
&alpha_f16, (const void **) (ptrs_src + 0*ne23), CUDA_R_16F, ne00,
|
|
|
|
(const void **) (ptrs_src + 1*ne23), CUDA_R_16F, ne10,
|
|
|
|
&beta_f16, ( void **) (ptrs_dst + 0*ne23), CUDA_R_16F, ne01,
|
|
|
|
ne23,
|
|
|
|
CUBLAS_COMPUTE_16F,
|
|
|
|
CUBLAS_GEMM_DEFAULT_TENSOR_OP));
|
|
|
|
|
|
|
|
if (src0_as != 0) {
|
|
|
|
ggml_cuda_pool_free(src0_as_f16, src0_as);
|
|
|
|
}
|
|
|
|
if (ptrs_src_s != 0) {
|
|
|
|
ggml_cuda_pool_free(ptrs_src, ptrs_src_s);
|
|
|
|
}
|
|
|
|
if (ptrs_dst_s != 0) {
|
|
|
|
ggml_cuda_pool_free(ptrs_dst, ptrs_dst_s);
|
|
|
|
}
|
|
|
|
|
|
|
|
const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
|
|
|
|
to_fp32_cuda(dst_f16, dst_ddf, ne, main_stream);
|
|
|
|
|
|
|
|
ggml_cuda_pool_free(src1_as_f16, src1_as);
|
|
|
|
ggml_cuda_pool_free(dst_f16, dst_as);
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
2023-12-13 12:04:25 +00:00
|
|
|
static void ggml_cuda_mul_mat_id(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
2023-12-07 20:26:54 +00:00
|
|
|
#if 0
|
|
|
|
ggml_cuda_mul_mat_id_cublas(dst);
|
|
|
|
// TODO: mmq/mmv support
|
2023-12-13 12:04:25 +00:00
|
|
|
#endif
|
2023-12-07 20:26:54 +00:00
|
|
|
|
2024-01-26 17:59:43 +00:00
|
|
|
const size_t nb11 = src1->nb[1];
|
|
|
|
const size_t nb1 = dst->nb[1];
|
2023-12-20 14:41:22 +00:00
|
|
|
|
2023-12-13 12:04:25 +00:00
|
|
|
const struct ggml_tensor * ids = src0;
|
|
|
|
const int32_t id = ((int32_t *) dst->op_params)[0];
|
|
|
|
const int32_t n_as = ((int32_t *) dst->op_params)[1];
|
2023-12-07 20:26:54 +00:00
|
|
|
|
2023-12-13 12:04:25 +00:00
|
|
|
std::vector<char> ids_host(ggml_nbytes(ids));
|
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
cudaStream_t stream = g_cudaStreams[g_main_device][0];
|
2023-12-20 14:41:22 +00:00
|
|
|
|
2024-02-25 10:09:09 +00:00
|
|
|
if (ids->backend == GGML_BACKEND_TYPE_GPU) {
|
2023-12-13 12:04:25 +00:00
|
|
|
const char * ids_dev = (const char *)((const ggml_tensor_extra_gpu *)ids->extra)->data_device[g_main_device];
|
2023-12-20 14:41:22 +00:00
|
|
|
CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids_dev, ggml_nbytes(ids), cudaMemcpyDeviceToHost, stream));
|
|
|
|
CUDA_CHECK(cudaStreamSynchronize(stream));
|
2023-12-13 12:04:25 +00:00
|
|
|
} else {
|
|
|
|
memcpy(ids_host.data(), ids->data, ggml_nbytes(ids));
|
|
|
|
}
|
|
|
|
|
|
|
|
const ggml_tensor_extra_gpu * src1_extra = (const ggml_tensor_extra_gpu *) src1->extra;
|
|
|
|
const ggml_tensor_extra_gpu * dst_extra = (const ggml_tensor_extra_gpu *) dst->extra;
|
|
|
|
|
|
|
|
ggml_tensor_extra_gpu src1_row_extra;
|
|
|
|
ggml_tensor_extra_gpu dst_row_extra;
|
|
|
|
|
|
|
|
ggml_tensor src1_row = *src1;
|
|
|
|
ggml_tensor dst_row = *dst;
|
|
|
|
|
2024-02-25 10:09:09 +00:00
|
|
|
src1_row.backend = GGML_BACKEND_TYPE_GPU;
|
|
|
|
dst_row.backend = GGML_BACKEND_TYPE_GPU;
|
2023-12-21 17:42:59 +00:00
|
|
|
|
2023-12-20 14:41:22 +00:00
|
|
|
src1_row.extra = &src1_row_extra;
|
|
|
|
dst_row.extra = &dst_row_extra;
|
2023-12-13 12:04:25 +00:00
|
|
|
|
2024-02-25 10:09:09 +00:00
|
|
|
char * src1_original = src1->backend == GGML_BACKEND_TYPE_CPU ?
|
2023-12-21 17:42:59 +00:00
|
|
|
(char *) src1->data : (char *) src1_extra->data_device[g_main_device];
|
2024-02-25 10:09:09 +00:00
|
|
|
char * dst_original = dst->backend == GGML_BACKEND_TYPE_CPU ?
|
2023-12-21 17:42:59 +00:00
|
|
|
(char *) dst->data : (char *) dst_extra->data_device[g_main_device];
|
2023-12-13 12:04:25 +00:00
|
|
|
|
2023-12-20 14:41:22 +00:00
|
|
|
if (src1->ne[1] == 1) {
|
2024-02-25 10:09:09 +00:00
|
|
|
GGML_ASSERT(src1->backend == GGML_BACKEND_TYPE_GPU);
|
|
|
|
GGML_ASSERT(dst->backend == GGML_BACKEND_TYPE_GPU);
|
2023-12-21 17:42:59 +00:00
|
|
|
|
2023-12-20 14:41:22 +00:00
|
|
|
for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
|
|
|
|
//int32_t row_id;
|
|
|
|
//CUDA_CHECK(cudaMemcpyAsync(&row_id, ids_dev + i01*ids->nb[1] + id*ids->nb[0], sizeof(int32_t), cudaMemcpyDeviceToHost, g_cudaStreams[g_main_device][0]));
|
|
|
|
//CUDA_CHECK(cudaStreamSynchronize(g_cudaStreams[g_main_device][0]));
|
2023-12-13 12:04:25 +00:00
|
|
|
|
2023-12-20 14:41:22 +00:00
|
|
|
const int32_t row_id = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]);
|
2023-12-07 20:26:54 +00:00
|
|
|
|
2023-12-20 14:41:22 +00:00
|
|
|
GGML_ASSERT(row_id >= 0 && row_id < n_as);
|
2023-12-07 20:26:54 +00:00
|
|
|
|
2023-12-20 14:41:22 +00:00
|
|
|
const struct ggml_tensor * src0_row = dst->src[row_id + 2];
|
2023-12-13 12:04:25 +00:00
|
|
|
|
2023-12-20 14:41:22 +00:00
|
|
|
src1_row_extra.data_device[g_main_device] = src1_original + i01*src1->nb[1];
|
|
|
|
src1_row.data = (char *) src1->data + i01*src1->nb[1]; // TODO why is this set?
|
2023-12-13 12:04:25 +00:00
|
|
|
|
2023-12-20 14:41:22 +00:00
|
|
|
dst_row_extra.data_device[g_main_device] = dst_original + i01*dst->nb[1];
|
|
|
|
dst_row.data = (char *) dst->data + i01*dst->nb[1]; // TODO why is this set?
|
2023-12-13 12:04:25 +00:00
|
|
|
|
2023-12-20 14:41:22 +00:00
|
|
|
ggml_cuda_mul_mat(src0_row, &src1_row, &dst_row);
|
|
|
|
}
|
|
|
|
} else {
|
2023-12-24 13:34:22 +00:00
|
|
|
cuda_pool_alloc<char> src1_contiguous(sizeof(float)*ggml_nelements(src1));
|
|
|
|
cuda_pool_alloc<char> dst_contiguous(sizeof(float)*ggml_nelements(dst));
|
2023-12-13 12:04:25 +00:00
|
|
|
|
2023-12-24 13:34:22 +00:00
|
|
|
src1_row_extra.data_device[g_main_device] = src1_contiguous.get();
|
|
|
|
dst_row_extra.data_device[g_main_device] = dst_contiguous.get();
|
2023-12-20 14:41:22 +00:00
|
|
|
|
2024-02-25 10:09:09 +00:00
|
|
|
const cudaMemcpyKind src1_kind = src1->backend == GGML_BACKEND_TYPE_CPU ?
|
2023-12-21 17:42:59 +00:00
|
|
|
cudaMemcpyHostToDevice : cudaMemcpyDeviceToDevice;
|
2024-02-25 10:09:09 +00:00
|
|
|
const cudaMemcpyKind dst_kind = dst->backend == GGML_BACKEND_TYPE_CPU ?
|
2023-12-22 13:34:05 +00:00
|
|
|
cudaMemcpyDeviceToHost : cudaMemcpyDeviceToDevice;
|
2023-12-21 17:42:59 +00:00
|
|
|
|
2023-12-20 14:41:22 +00:00
|
|
|
for (int32_t row_id = 0; row_id < n_as; ++row_id) {
|
|
|
|
const struct ggml_tensor * src0_row = dst->src[row_id + 2];
|
|
|
|
|
|
|
|
int64_t num_src1_rows = 0;
|
|
|
|
for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
|
|
|
|
const int32_t row_id_i = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]);
|
|
|
|
|
|
|
|
if (row_id_i != row_id) {
|
|
|
|
continue;
|
|
|
|
}
|
2023-12-13 12:04:25 +00:00
|
|
|
|
2023-12-20 14:41:22 +00:00
|
|
|
GGML_ASSERT(row_id >= 0 && row_id < n_as);
|
2023-12-13 12:04:25 +00:00
|
|
|
|
2023-12-24 13:34:22 +00:00
|
|
|
CUDA_CHECK(cudaMemcpyAsync(src1_contiguous.get() + num_src1_rows*nb11, src1_original + i01*nb11,
|
2023-12-21 17:42:59 +00:00
|
|
|
nb11, src1_kind, stream));
|
2023-12-20 14:41:22 +00:00
|
|
|
num_src1_rows++;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (num_src1_rows == 0) {
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
src1_row.ne[1] = num_src1_rows;
|
|
|
|
dst_row.ne[1] = num_src1_rows;
|
|
|
|
|
|
|
|
src1_row.nb[1] = nb11;
|
|
|
|
src1_row.nb[2] = num_src1_rows*nb11;
|
|
|
|
src1_row.nb[3] = num_src1_rows*nb11;
|
|
|
|
|
|
|
|
dst_row.nb[1] = nb1;
|
|
|
|
dst_row.nb[2] = num_src1_rows*nb1;
|
|
|
|
dst_row.nb[3] = num_src1_rows*nb1;
|
|
|
|
|
|
|
|
ggml_cuda_mul_mat(src0_row, &src1_row, &dst_row);
|
|
|
|
|
|
|
|
num_src1_rows = 0;
|
|
|
|
for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
|
|
|
|
const int32_t row_id_i = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]);
|
|
|
|
|
|
|
|
if (row_id_i != row_id) {
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_ASSERT(row_id >= 0 && row_id < n_as);
|
|
|
|
|
2023-12-24 13:34:22 +00:00
|
|
|
CUDA_CHECK(cudaMemcpyAsync(dst_original + i01*nb1, dst_contiguous.get() + num_src1_rows*nb1,
|
2023-12-21 17:42:59 +00:00
|
|
|
nb1, dst_kind, stream));
|
2023-12-20 14:41:22 +00:00
|
|
|
num_src1_rows++;
|
|
|
|
}
|
|
|
|
}
|
2023-12-13 12:04:25 +00:00
|
|
|
}
|
2023-12-21 17:42:59 +00:00
|
|
|
|
2024-02-25 10:09:09 +00:00
|
|
|
if (dst->backend == GGML_BACKEND_TYPE_CPU) {
|
2023-12-21 17:42:59 +00:00
|
|
|
CUDA_CHECK(cudaStreamSynchronize(stream));
|
|
|
|
}
|
2023-12-07 20:26:54 +00:00
|
|
|
}
|
|
|
|
|
2023-09-28 16:04:36 +00:00
|
|
|
static void ggml_cuda_scale(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
2023-09-11 17:55:51 +00:00
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_scale);
|
2023-06-14 17:47:19 +00:00
|
|
|
}
|
|
|
|
|
2023-10-10 07:50:23 +00:00
|
|
|
static void ggml_cuda_clamp(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_clamp);
|
|
|
|
}
|
|
|
|
|
2023-09-28 16:04:36 +00:00
|
|
|
static void ggml_cuda_cpy(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
2023-06-14 17:47:19 +00:00
|
|
|
const int64_t ne = ggml_nelements(src0);
|
|
|
|
GGML_ASSERT(ne == ggml_nelements(src1));
|
|
|
|
|
2024-02-25 10:09:09 +00:00
|
|
|
GGML_ASSERT(src0->backend == GGML_BACKEND_TYPE_GPU);
|
|
|
|
GGML_ASSERT(src1->backend == GGML_BACKEND_TYPE_GPU);
|
2023-06-14 17:47:19 +00:00
|
|
|
|
|
|
|
GGML_ASSERT(ggml_nbytes(src0) <= INT_MAX);
|
|
|
|
GGML_ASSERT(ggml_nbytes(src1) <= INT_MAX);
|
|
|
|
|
|
|
|
const int64_t ne00 = src0->ne[0];
|
|
|
|
const int64_t ne01 = src0->ne[1];
|
2024-01-29 12:37:33 +00:00
|
|
|
const int64_t ne02 = src0->ne[2];
|
2024-01-30 14:21:57 +00:00
|
|
|
|
2024-01-29 12:37:33 +00:00
|
|
|
//GGML_ASSERT(src0->ne[3] == 1);
|
2023-06-14 17:47:19 +00:00
|
|
|
|
|
|
|
const int64_t nb00 = src0->nb[0];
|
|
|
|
const int64_t nb01 = src0->nb[1];
|
|
|
|
const int64_t nb02 = src0->nb[2];
|
2024-01-29 12:37:33 +00:00
|
|
|
const int64_t nb03 = src0->nb[3];
|
2023-06-14 17:47:19 +00:00
|
|
|
|
|
|
|
const int64_t ne10 = src1->ne[0];
|
|
|
|
const int64_t ne11 = src1->ne[1];
|
2024-01-29 12:37:33 +00:00
|
|
|
const int64_t ne12 = src1->ne[2];
|
|
|
|
|
|
|
|
//GGML_ASSERT(src1->ne[3] == 1);
|
2023-06-14 17:47:19 +00:00
|
|
|
|
|
|
|
const int64_t nb10 = src1->nb[0];
|
|
|
|
const int64_t nb11 = src1->nb[1];
|
|
|
|
const int64_t nb12 = src1->nb[2];
|
2024-01-29 12:37:33 +00:00
|
|
|
const int64_t nb13 = src1->nb[3];
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2023-12-26 20:23:59 +00:00
|
|
|
ggml_cuda_set_device(g_main_device);
|
2023-09-11 17:55:51 +00:00
|
|
|
cudaStream_t main_stream = g_cudaStreams[g_main_device][0];
|
2023-06-14 17:47:19 +00:00
|
|
|
|
2023-10-08 17:19:14 +00:00
|
|
|
const ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
|
|
|
|
const ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
|
2023-06-14 17:47:19 +00:00
|
|
|
|
|
|
|
char * src0_ddc = (char *) src0_extra->data_device[g_main_device];
|
|
|
|
char * src1_ddc = (char *) src1_extra->data_device[g_main_device];
|
|
|
|
|
|
|
|
if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
|
2024-01-29 12:37:33 +00:00
|
|
|
ggml_cpy_f32_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
|
2023-06-14 17:47:19 +00:00
|
|
|
} else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
|
2024-01-29 12:37:33 +00:00
|
|
|
ggml_cpy_f32_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
|
2023-12-07 11:03:17 +00:00
|
|
|
} else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
|
2024-01-29 12:37:33 +00:00
|
|
|
ggml_cpy_f32_q8_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
|
2023-12-07 11:03:17 +00:00
|
|
|
} else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) {
|
2024-01-29 12:37:33 +00:00
|
|
|
ggml_cpy_f32_q4_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
|
2023-12-07 11:03:17 +00:00
|
|
|
} else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_1) {
|
2024-01-29 12:37:33 +00:00
|
|
|
ggml_cpy_f32_q4_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
|
2023-11-13 14:55:52 +00:00
|
|
|
} else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
|
2024-01-29 12:37:33 +00:00
|
|
|
ggml_cpy_f16_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
|
|
|
|
} else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
|
|
|
|
ggml_cpy_f16_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
|
2023-06-14 17:47:19 +00:00
|
|
|
} else {
|
2023-09-28 19:42:38 +00:00
|
|
|
fprintf(stderr, "%s: unsupported type combination (%s to %s)\n", __func__,
|
|
|
|
ggml_type_name(src0->type), ggml_type_name(src1->type));
|
2023-06-14 17:47:19 +00:00
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
|
|
|
|
(void) dst;
|
|
|
|
}
|
|
|
|
|
2023-09-28 16:04:36 +00:00
|
|
|
static void ggml_cuda_dup(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
2023-12-07 11:03:17 +00:00
|
|
|
// TODO: why do we pass dst as src1 here?
|
2023-07-17 17:39:29 +00:00
|
|
|
ggml_cuda_cpy(src0, dst, nullptr);
|
|
|
|
(void) src1;
|
|
|
|
}
|
|
|
|
|
2023-09-28 16:04:36 +00:00
|
|
|
static void ggml_cuda_diag_mask_inf(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
2023-09-11 17:55:51 +00:00
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_diag_mask_inf);
|
2023-06-14 17:47:19 +00:00
|
|
|
}
|
|
|
|
|
2023-09-28 16:04:36 +00:00
|
|
|
static void ggml_cuda_soft_max(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
2023-09-11 17:55:51 +00:00
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_soft_max);
|
2023-06-14 17:47:19 +00:00
|
|
|
}
|
|
|
|
|
2023-09-28 16:04:36 +00:00
|
|
|
static void ggml_cuda_rope(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
2023-08-27 13:40:48 +00:00
|
|
|
GGML_ASSERT(ggml_is_contiguous(src0)); // TODO: this restriction is temporary until non-cont support is implemented
|
2023-09-11 17:55:51 +00:00
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_rope);
|
2023-06-06 19:33:23 +00:00
|
|
|
}
|
2023-05-01 16:11:07 +00:00
|
|
|
|
2023-09-28 16:04:36 +00:00
|
|
|
static void ggml_cuda_alibi(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
2023-09-11 17:55:51 +00:00
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_alibi);
|
2023-08-22 11:22:08 +00:00
|
|
|
}
|
|
|
|
|
2024-01-31 13:10:15 +00:00
|
|
|
static void ggml_cuda_pool2d(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_pool2d);
|
|
|
|
}
|
|
|
|
|
2023-11-15 12:58:13 +00:00
|
|
|
static void ggml_cuda_im2col(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
2023-11-13 14:55:52 +00:00
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_im2col);
|
|
|
|
}
|
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
static void ggml_cuda_sum_rows(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
GGML_ASSERT(ggml_is_contiguous(src0));
|
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_sum_rows);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_cuda_argsort(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
GGML_ASSERT(ggml_is_contiguous(src0));
|
|
|
|
ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_argsort);
|
|
|
|
}
|
|
|
|
|
2023-09-28 16:04:36 +00:00
|
|
|
static void ggml_cuda_nop(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
2023-06-06 19:33:23 +00:00
|
|
|
(void) src0;
|
|
|
|
(void) src1;
|
|
|
|
(void) dst;
|
2023-05-01 16:11:07 +00:00
|
|
|
}
|
|
|
|
|
2023-12-14 19:05:21 +00:00
|
|
|
static size_t ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split) {
|
|
|
|
static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
|
|
|
|
|
|
|
|
return nrows_split*ggml_row_size(tensor->type, tensor->ne[0]);
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static void ggml_cuda_set_main_device(const int main_device) {
|
2023-06-15 17:29:59 +00:00
|
|
|
if (main_device >= g_device_count) {
|
2023-06-06 19:33:23 +00:00
|
|
|
fprintf(stderr, "warning: cannot set main_device=%d because there are only %d devices. Using device %d instead.\n",
|
|
|
|
main_device, g_device_count, g_main_device);
|
|
|
|
return;
|
|
|
|
}
|
2023-12-07 20:26:54 +00:00
|
|
|
|
|
|
|
if (g_main_device != main_device && g_device_count > 1) {
|
|
|
|
g_main_device = main_device;
|
2024-01-12 19:07:38 +00:00
|
|
|
//cudaDeviceProp prop;
|
|
|
|
//CUDA_CHECK(cudaGetDeviceProperties(&prop, g_main_device));
|
|
|
|
//fprintf(stderr, "%s: using device %d (%s) as main device\n", __func__, g_main_device, prop.name);
|
2023-06-14 17:47:19 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor) {
|
2023-12-07 20:26:54 +00:00
|
|
|
if (!g_cublas_loaded) return false;
|
2023-11-07 06:49:08 +00:00
|
|
|
|
2023-06-06 19:33:23 +00:00
|
|
|
ggml_cuda_func_t func;
|
2024-02-25 10:09:09 +00:00
|
|
|
const bool any_on_device = tensor->backend == GGML_BACKEND_TYPE_GPU
|
|
|
|
|| (tensor->src[0] != nullptr && (tensor->src[0]->backend == GGML_BACKEND_TYPE_GPU || tensor->src[0]->backend == GGML_BACKEND_TYPE_GPU_SPLIT))
|
|
|
|
|| (tensor->src[1] != nullptr && tensor->src[1]->backend == GGML_BACKEND_TYPE_GPU);
|
cuda : loading models directly into VRAM, norm calculation on GPU, broadcasting for ggml_mul (#1483)
* Broadcasting for ggml_mul
* CUDA kernel for ggml_mul, norms in VRAM
* GPU weights not in RAM, direct loading with cuFile
* fixup! GPU weights not in RAM, direct loading with cuFile
* fixup! GPU weights not in RAM, direct loading with cuFile
* define default model path once, sync path with readme (#1366)
* ~7% faster Q5_1 AVX2 code (#1477)
* convert.py: Support models which are stored in a single pytorch_model.bin (#1469)
* Support models in a single pytorch_model.bin
* Remove spurious line with typo
* benchmark-matmul: Print the average of the test results (#1490)
* Remove unused n_parts parameter (#1509)
* Fixes #1511 lambda issue for w64devkit (mingw) (#1513)
* Fix for w64devkit and mingw
* make kv_f16 the default for api users (#1517)
* minor : fix compile warnings
* readme : adds WizardLM to the list of supported models (#1485)
* main : make reverse prompt option act as a stop token in non-interactive mode (#1032)
* Make reverse prompt option act as a stop token in non-interactive scenarios
* Making requested review changes
* Update gpt_params_parse and fix a merge error
* Revert "Update gpt_params_parse and fix a merge error"
This reverts commit 2bb2ff1748513591ad45b175a75ed1d8089d84c8.
* Update gpt_params_parse and fix a merge error take 2
* examples : add persistent chat (#1495)
* examples : add persistent chat
* examples : fix whitespace
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* tests : add missing header
* ggml : use F16 instead of F32 in Q4_0, Q4_1, Q8_0 (#1508)
* ggml : use F16 instead of F32 in Q4_0, Q4_1 and Q8_0
* llama : bump LLAMA_FILE_VERSION to 3
* cuda : update Q4 and Q8 dequantize kernels
* ggml : fix AVX dot products
* readme : update performance table + hot topics
* ggml : fix scalar implementation of Q4_1 dot
* llama : fix compile warnings in llama_set_state_data()
* llama : fix name shadowing and C4146 (#1526)
* Fix name shadowing and C4146
* Fix if macros not using defined when required
* Update llama-util.h
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
* Update llama-util.h
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
* Code style
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Fix for mingw (#1462)
* llama : add llama_init_backend() API (close #1527)
* feature : add blis and other BLAS implementation support (#1502)
* feature: add blis support
* feature: allow all BLA_VENDOR to be assigned in cmake arguments. align with whisper.cpp pr 927
* fix: version detection for BLA_SIZEOF_INTEGER, recover min version of cmake
* Fix typo in INTEGER
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Revert "feature : add blis and other BLAS implementation support (#1502)"
This reverts commit 07e9ace0f9da424d82e75df969642522880feb92.
* GPU weights not in RAM, direct loading with cuFile
* llama : code style fixes + progress print fix
* ggml : ggml_mul better broadcast support
* cmake : workarounds for cufile when CMake version < 3.25
* gg rebase fixup
* Loop in llama.cpp, fixed progress callback
* Attempt clang-tidy fix
* llama : fix vram size computation
* Add forgotten fclose()
---------
Co-authored-by: András Salamon <ott2@users.noreply.github.com>
Co-authored-by: Ilya Kurdyukov <59548320+ilyakurdyukov@users.noreply.github.com>
Co-authored-by: Tom Jobbins <784313+TheBloke@users.noreply.github.com>
Co-authored-by: rankaiyx <rankaiyx@rankaiyx.com>
Co-authored-by: Stephan Walter <stephan@walter.name>
Co-authored-by: DannyDaemonic <DannyDaemonic@gmail.com>
Co-authored-by: Erik Scholz <Green-Sky@users.noreply.github.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: David Kennedy <dakennedyd@gmail.com>
Co-authored-by: Jason McCartney <jmac@theroot.org>
Co-authored-by: Evan Jones <evan.q.jones@gmail.com>
Co-authored-by: Maxime <672982+maximegmd@users.noreply.github.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Zenix <zenixls2@gmail.com>
2023-05-20 12:19:28 +00:00
|
|
|
|
2023-12-21 17:42:59 +00:00
|
|
|
if (!any_on_device && tensor->op != GGML_OP_MUL_MAT && tensor->op != GGML_OP_MUL_MAT_ID) {
|
2023-10-08 17:19:14 +00:00
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2023-11-13 14:55:52 +00:00
|
|
|
if (tensor->op == GGML_OP_MUL_MAT) {
|
|
|
|
if (tensor->src[0]->ne[3] != tensor->src[1]->ne[3]) {
|
|
|
|
#ifndef NDEBUG
|
2023-12-21 20:07:46 +00:00
|
|
|
fprintf(stderr, "%s: cannot compute %s: src0->ne[3] = %" PRId64 ", src1->ne[3] = %" PRId64 " - fallback to CPU\n", __func__, tensor->name, tensor->src[0]->ne[3], tensor->src[1]->ne[3]);
|
2023-11-13 14:55:52 +00:00
|
|
|
#endif
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-06-06 19:33:23 +00:00
|
|
|
switch (tensor->op) {
|
2023-10-08 17:19:14 +00:00
|
|
|
case GGML_OP_REPEAT:
|
|
|
|
func = ggml_cuda_repeat;
|
|
|
|
break;
|
|
|
|
case GGML_OP_GET_ROWS:
|
|
|
|
func = ggml_cuda_get_rows;
|
|
|
|
break;
|
2023-07-17 17:39:29 +00:00
|
|
|
case GGML_OP_DUP:
|
|
|
|
func = ggml_cuda_dup;
|
|
|
|
break;
|
2023-06-06 19:33:23 +00:00
|
|
|
case GGML_OP_ADD:
|
|
|
|
func = ggml_cuda_add;
|
|
|
|
break;
|
2023-12-13 19:54:54 +00:00
|
|
|
case GGML_OP_ACC:
|
|
|
|
func = ggml_cuda_acc;
|
|
|
|
break;
|
2023-06-06 19:33:23 +00:00
|
|
|
case GGML_OP_MUL:
|
|
|
|
func = ggml_cuda_mul;
|
|
|
|
break;
|
2023-12-07 20:26:54 +00:00
|
|
|
case GGML_OP_DIV:
|
|
|
|
func = ggml_cuda_div;
|
|
|
|
break;
|
2023-07-24 11:46:21 +00:00
|
|
|
case GGML_OP_UNARY:
|
|
|
|
switch (ggml_get_unary_op(tensor)) {
|
|
|
|
case GGML_UNARY_OP_GELU:
|
|
|
|
func = ggml_cuda_gelu;
|
|
|
|
break;
|
|
|
|
case GGML_UNARY_OP_SILU:
|
|
|
|
func = ggml_cuda_silu;
|
|
|
|
break;
|
2023-12-13 19:54:54 +00:00
|
|
|
case GGML_UNARY_OP_GELU_QUICK:
|
|
|
|
func = ggml_cuda_gelu_quick;
|
|
|
|
break;
|
|
|
|
case GGML_UNARY_OP_TANH:
|
|
|
|
func = ggml_cuda_tanh;
|
|
|
|
break;
|
2023-11-13 08:58:15 +00:00
|
|
|
case GGML_UNARY_OP_RELU:
|
|
|
|
func = ggml_cuda_relu;
|
|
|
|
break;
|
2024-01-31 13:10:15 +00:00
|
|
|
case GGML_UNARY_OP_HARDSIGMOID:
|
|
|
|
func = ggml_cuda_hardsigmoid;
|
|
|
|
break;
|
|
|
|
case GGML_UNARY_OP_HARDSWISH:
|
|
|
|
func = ggml_cuda_hardswish;
|
|
|
|
break;
|
2023-07-24 11:46:21 +00:00
|
|
|
default:
|
|
|
|
return false;
|
2023-12-07 20:26:54 +00:00
|
|
|
}
|
|
|
|
break;
|
2023-07-11 19:53:34 +00:00
|
|
|
case GGML_OP_NORM:
|
|
|
|
func = ggml_cuda_norm;
|
|
|
|
break;
|
2023-12-13 19:54:54 +00:00
|
|
|
case GGML_OP_GROUP_NORM:
|
|
|
|
func = ggml_cuda_group_norm;
|
|
|
|
break;
|
|
|
|
case GGML_OP_CONCAT:
|
|
|
|
func = ggml_cuda_concat;
|
|
|
|
break;
|
|
|
|
case GGML_OP_UPSCALE:
|
|
|
|
func = ggml_cuda_upscale;
|
|
|
|
break;
|
|
|
|
case GGML_OP_PAD:
|
|
|
|
func = ggml_cuda_pad;
|
|
|
|
break;
|
|
|
|
case GGML_OP_LEAKY_RELU:
|
|
|
|
func = ggml_cuda_leaky_relu;
|
|
|
|
break;
|
2023-06-06 19:33:23 +00:00
|
|
|
case GGML_OP_RMS_NORM:
|
|
|
|
func = ggml_cuda_rms_norm;
|
|
|
|
break;
|
|
|
|
case GGML_OP_MUL_MAT:
|
2023-07-11 16:31:10 +00:00
|
|
|
if (!any_on_device && !ggml_cuda_can_mul_mat(tensor->src[0], tensor->src[1], tensor)) {
|
2023-06-06 19:33:23 +00:00
|
|
|
return false;
|
|
|
|
}
|
|
|
|
func = ggml_cuda_mul_mat;
|
|
|
|
break;
|
2023-12-07 20:26:54 +00:00
|
|
|
case GGML_OP_MUL_MAT_ID:
|
|
|
|
if (!any_on_device && !ggml_cuda_can_mul_mat(tensor->src[2], tensor->src[1], tensor)) {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
func = ggml_cuda_mul_mat_id;
|
|
|
|
break;
|
2023-06-14 17:47:19 +00:00
|
|
|
case GGML_OP_SCALE:
|
|
|
|
func = ggml_cuda_scale;
|
|
|
|
break;
|
2023-11-13 08:58:15 +00:00
|
|
|
case GGML_OP_SQR:
|
|
|
|
func = ggml_cuda_sqr;
|
|
|
|
break;
|
2023-10-10 07:50:23 +00:00
|
|
|
case GGML_OP_CLAMP:
|
|
|
|
func = ggml_cuda_clamp;
|
|
|
|
break;
|
2023-06-14 17:47:19 +00:00
|
|
|
case GGML_OP_CPY:
|
|
|
|
func = ggml_cuda_cpy;
|
|
|
|
break;
|
2023-07-17 17:39:29 +00:00
|
|
|
case GGML_OP_CONT:
|
|
|
|
func = ggml_cuda_dup;
|
|
|
|
break;
|
2023-12-13 19:54:54 +00:00
|
|
|
case GGML_OP_NONE:
|
2023-06-06 19:33:23 +00:00
|
|
|
case GGML_OP_RESHAPE:
|
2023-06-14 17:47:19 +00:00
|
|
|
case GGML_OP_VIEW:
|
|
|
|
case GGML_OP_PERMUTE:
|
|
|
|
case GGML_OP_TRANSPOSE:
|
2023-06-06 19:33:23 +00:00
|
|
|
func = ggml_cuda_nop;
|
|
|
|
break;
|
2023-06-14 17:47:19 +00:00
|
|
|
case GGML_OP_DIAG_MASK_INF:
|
|
|
|
func = ggml_cuda_diag_mask_inf;
|
|
|
|
break;
|
|
|
|
case GGML_OP_SOFT_MAX:
|
|
|
|
func = ggml_cuda_soft_max;
|
|
|
|
break;
|
2023-06-06 19:33:23 +00:00
|
|
|
case GGML_OP_ROPE:
|
|
|
|
func = ggml_cuda_rope;
|
|
|
|
break;
|
2023-08-22 11:22:08 +00:00
|
|
|
case GGML_OP_ALIBI:
|
|
|
|
func = ggml_cuda_alibi;
|
|
|
|
break;
|
2023-11-13 14:55:52 +00:00
|
|
|
case GGML_OP_IM2COL:
|
|
|
|
func = ggml_cuda_im2col;
|
|
|
|
break;
|
2024-01-31 13:10:15 +00:00
|
|
|
case GGML_OP_POOL_2D:
|
|
|
|
func = ggml_cuda_pool2d;
|
|
|
|
break;
|
2023-12-07 20:26:54 +00:00
|
|
|
case GGML_OP_SUM_ROWS:
|
|
|
|
func = ggml_cuda_sum_rows;
|
|
|
|
break;
|
|
|
|
case GGML_OP_ARGSORT:
|
|
|
|
func = ggml_cuda_argsort;
|
|
|
|
break;
|
2023-06-06 19:33:23 +00:00
|
|
|
default:
|
|
|
|
return false;
|
cuda : loading models directly into VRAM, norm calculation on GPU, broadcasting for ggml_mul (#1483)
* Broadcasting for ggml_mul
* CUDA kernel for ggml_mul, norms in VRAM
* GPU weights not in RAM, direct loading with cuFile
* fixup! GPU weights not in RAM, direct loading with cuFile
* fixup! GPU weights not in RAM, direct loading with cuFile
* define default model path once, sync path with readme (#1366)
* ~7% faster Q5_1 AVX2 code (#1477)
* convert.py: Support models which are stored in a single pytorch_model.bin (#1469)
* Support models in a single pytorch_model.bin
* Remove spurious line with typo
* benchmark-matmul: Print the average of the test results (#1490)
* Remove unused n_parts parameter (#1509)
* Fixes #1511 lambda issue for w64devkit (mingw) (#1513)
* Fix for w64devkit and mingw
* make kv_f16 the default for api users (#1517)
* minor : fix compile warnings
* readme : adds WizardLM to the list of supported models (#1485)
* main : make reverse prompt option act as a stop token in non-interactive mode (#1032)
* Make reverse prompt option act as a stop token in non-interactive scenarios
* Making requested review changes
* Update gpt_params_parse and fix a merge error
* Revert "Update gpt_params_parse and fix a merge error"
This reverts commit 2bb2ff1748513591ad45b175a75ed1d8089d84c8.
* Update gpt_params_parse and fix a merge error take 2
* examples : add persistent chat (#1495)
* examples : add persistent chat
* examples : fix whitespace
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* tests : add missing header
* ggml : use F16 instead of F32 in Q4_0, Q4_1, Q8_0 (#1508)
* ggml : use F16 instead of F32 in Q4_0, Q4_1 and Q8_0
* llama : bump LLAMA_FILE_VERSION to 3
* cuda : update Q4 and Q8 dequantize kernels
* ggml : fix AVX dot products
* readme : update performance table + hot topics
* ggml : fix scalar implementation of Q4_1 dot
* llama : fix compile warnings in llama_set_state_data()
* llama : fix name shadowing and C4146 (#1526)
* Fix name shadowing and C4146
* Fix if macros not using defined when required
* Update llama-util.h
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
* Update llama-util.h
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
* Code style
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Fix for mingw (#1462)
* llama : add llama_init_backend() API (close #1527)
* feature : add blis and other BLAS implementation support (#1502)
* feature: add blis support
* feature: allow all BLA_VENDOR to be assigned in cmake arguments. align with whisper.cpp pr 927
* fix: version detection for BLA_SIZEOF_INTEGER, recover min version of cmake
* Fix typo in INTEGER
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Revert "feature : add blis and other BLAS implementation support (#1502)"
This reverts commit 07e9ace0f9da424d82e75df969642522880feb92.
* GPU weights not in RAM, direct loading with cuFile
* llama : code style fixes + progress print fix
* ggml : ggml_mul better broadcast support
* cmake : workarounds for cufile when CMake version < 3.25
* gg rebase fixup
* Loop in llama.cpp, fixed progress callback
* Attempt clang-tidy fix
* llama : fix vram size computation
* Add forgotten fclose()
---------
Co-authored-by: András Salamon <ott2@users.noreply.github.com>
Co-authored-by: Ilya Kurdyukov <59548320+ilyakurdyukov@users.noreply.github.com>
Co-authored-by: Tom Jobbins <784313+TheBloke@users.noreply.github.com>
Co-authored-by: rankaiyx <rankaiyx@rankaiyx.com>
Co-authored-by: Stephan Walter <stephan@walter.name>
Co-authored-by: DannyDaemonic <DannyDaemonic@gmail.com>
Co-authored-by: Erik Scholz <Green-Sky@users.noreply.github.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: David Kennedy <dakennedyd@gmail.com>
Co-authored-by: Jason McCartney <jmac@theroot.org>
Co-authored-by: Evan Jones <evan.q.jones@gmail.com>
Co-authored-by: Maxime <672982+maximegmd@users.noreply.github.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Zenix <zenixls2@gmail.com>
2023-05-20 12:19:28 +00:00
|
|
|
}
|
|
|
|
|
2024-02-25 10:09:09 +00:00
|
|
|
if (tensor->src[0] != nullptr && tensor->src[0]->backend == GGML_BACKEND_TYPE_GPU_SPLIT) {
|
2023-12-20 14:41:22 +00:00
|
|
|
ggml_cuda_set_peer_access(tensor->src[1]->ne[1]);
|
|
|
|
}
|
|
|
|
|
2023-06-06 19:33:23 +00:00
|
|
|
if (params->ith != 0) {
|
|
|
|
return true;
|
|
|
|
}
|
2024-02-25 10:09:09 +00:00
|
|
|
if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
|
2023-06-06 19:33:23 +00:00
|
|
|
return true;
|
|
|
|
}
|
2023-07-11 16:31:10 +00:00
|
|
|
func(tensor->src[0], tensor->src[1], tensor);
|
2023-06-06 19:33:23 +00:00
|
|
|
return true;
|
cuda : loading models directly into VRAM, norm calculation on GPU, broadcasting for ggml_mul (#1483)
* Broadcasting for ggml_mul
* CUDA kernel for ggml_mul, norms in VRAM
* GPU weights not in RAM, direct loading with cuFile
* fixup! GPU weights not in RAM, direct loading with cuFile
* fixup! GPU weights not in RAM, direct loading with cuFile
* define default model path once, sync path with readme (#1366)
* ~7% faster Q5_1 AVX2 code (#1477)
* convert.py: Support models which are stored in a single pytorch_model.bin (#1469)
* Support models in a single pytorch_model.bin
* Remove spurious line with typo
* benchmark-matmul: Print the average of the test results (#1490)
* Remove unused n_parts parameter (#1509)
* Fixes #1511 lambda issue for w64devkit (mingw) (#1513)
* Fix for w64devkit and mingw
* make kv_f16 the default for api users (#1517)
* minor : fix compile warnings
* readme : adds WizardLM to the list of supported models (#1485)
* main : make reverse prompt option act as a stop token in non-interactive mode (#1032)
* Make reverse prompt option act as a stop token in non-interactive scenarios
* Making requested review changes
* Update gpt_params_parse and fix a merge error
* Revert "Update gpt_params_parse and fix a merge error"
This reverts commit 2bb2ff1748513591ad45b175a75ed1d8089d84c8.
* Update gpt_params_parse and fix a merge error take 2
* examples : add persistent chat (#1495)
* examples : add persistent chat
* examples : fix whitespace
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* tests : add missing header
* ggml : use F16 instead of F32 in Q4_0, Q4_1, Q8_0 (#1508)
* ggml : use F16 instead of F32 in Q4_0, Q4_1 and Q8_0
* llama : bump LLAMA_FILE_VERSION to 3
* cuda : update Q4 and Q8 dequantize kernels
* ggml : fix AVX dot products
* readme : update performance table + hot topics
* ggml : fix scalar implementation of Q4_1 dot
* llama : fix compile warnings in llama_set_state_data()
* llama : fix name shadowing and C4146 (#1526)
* Fix name shadowing and C4146
* Fix if macros not using defined when required
* Update llama-util.h
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
* Update llama-util.h
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
* Code style
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Fix for mingw (#1462)
* llama : add llama_init_backend() API (close #1527)
* feature : add blis and other BLAS implementation support (#1502)
* feature: add blis support
* feature: allow all BLA_VENDOR to be assigned in cmake arguments. align with whisper.cpp pr 927
* fix: version detection for BLA_SIZEOF_INTEGER, recover min version of cmake
* Fix typo in INTEGER
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Revert "feature : add blis and other BLAS implementation support (#1502)"
This reverts commit 07e9ace0f9da424d82e75df969642522880feb92.
* GPU weights not in RAM, direct loading with cuFile
* llama : code style fixes + progress print fix
* ggml : ggml_mul better broadcast support
* cmake : workarounds for cufile when CMake version < 3.25
* gg rebase fixup
* Loop in llama.cpp, fixed progress callback
* Attempt clang-tidy fix
* llama : fix vram size computation
* Add forgotten fclose()
---------
Co-authored-by: András Salamon <ott2@users.noreply.github.com>
Co-authored-by: Ilya Kurdyukov <59548320+ilyakurdyukov@users.noreply.github.com>
Co-authored-by: Tom Jobbins <784313+TheBloke@users.noreply.github.com>
Co-authored-by: rankaiyx <rankaiyx@rankaiyx.com>
Co-authored-by: Stephan Walter <stephan@walter.name>
Co-authored-by: DannyDaemonic <DannyDaemonic@gmail.com>
Co-authored-by: Erik Scholz <Green-Sky@users.noreply.github.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: David Kennedy <dakennedyd@gmail.com>
Co-authored-by: Jason McCartney <jmac@theroot.org>
Co-authored-by: Evan Jones <evan.q.jones@gmail.com>
Co-authored-by: Maxime <672982+maximegmd@users.noreply.github.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Zenix <zenixls2@gmail.com>
2023-05-20 12:19:28 +00:00
|
|
|
}
|
2023-08-18 10:44:58 +00:00
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL int ggml_cuda_get_device_count() {
|
2023-08-18 10:44:58 +00:00
|
|
|
int device_count;
|
2023-12-07 20:26:54 +00:00
|
|
|
if (cudaGetDeviceCount(&device_count) != cudaSuccess) {
|
|
|
|
return 0;
|
|
|
|
}
|
2023-08-18 10:44:58 +00:00
|
|
|
return device_count;
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL void ggml_cuda_get_device_description(int device, char * description, size_t description_size) {
|
2023-08-18 10:44:58 +00:00
|
|
|
cudaDeviceProp prop;
|
|
|
|
CUDA_CHECK(cudaGetDeviceProperties(&prop, device));
|
|
|
|
snprintf(description, description_size, "%s", prop.name);
|
|
|
|
}
|
2023-10-08 17:19:14 +00:00
|
|
|
|
|
|
|
////////////////////////////////////////////////////////////////////////////////
|
|
|
|
|
|
|
|
// backend interface
|
|
|
|
|
|
|
|
#define UNUSED GGML_UNUSED
|
|
|
|
|
2024-01-12 19:07:38 +00:00
|
|
|
struct ggml_backend_cuda_context {
|
|
|
|
int device;
|
|
|
|
std::string name;
|
|
|
|
};
|
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
// cuda buffer
|
2023-10-08 17:19:14 +00:00
|
|
|
|
2024-01-12 19:07:38 +00:00
|
|
|
struct ggml_backend_cuda_buffer_context {
|
2023-12-07 20:26:54 +00:00
|
|
|
int device;
|
|
|
|
void * dev_ptr = nullptr;
|
2023-10-08 17:19:14 +00:00
|
|
|
ggml_tensor_extra_gpu * temp_tensor_extras = nullptr;
|
|
|
|
size_t temp_tensor_extra_index = 0;
|
2024-01-12 19:07:38 +00:00
|
|
|
std::string name;
|
2023-10-08 17:19:14 +00:00
|
|
|
|
2024-01-12 19:07:38 +00:00
|
|
|
ggml_backend_cuda_buffer_context(int device, void * dev_ptr) :
|
|
|
|
device(device), dev_ptr(dev_ptr),
|
|
|
|
name(GGML_CUDA_NAME + std::to_string(device)) {
|
|
|
|
}
|
2023-12-07 20:26:54 +00:00
|
|
|
|
2024-01-12 19:07:38 +00:00
|
|
|
~ggml_backend_cuda_buffer_context() {
|
2023-10-08 17:19:14 +00:00
|
|
|
delete[] temp_tensor_extras;
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_tensor_extra_gpu * ggml_cuda_alloc_temp_tensor_extra() {
|
2024-01-12 19:07:38 +00:00
|
|
|
// TODO: remove GGML_CUDA_MAX_NODES, allocate dynamically and reuse in backend_buffer_reset
|
2023-10-08 17:19:14 +00:00
|
|
|
if (temp_tensor_extras == nullptr) {
|
2023-11-15 12:58:13 +00:00
|
|
|
temp_tensor_extras = new ggml_tensor_extra_gpu[GGML_CUDA_MAX_NODES];
|
2023-10-08 17:19:14 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
size_t alloc_index = temp_tensor_extra_index;
|
2023-11-15 12:58:13 +00:00
|
|
|
temp_tensor_extra_index = (temp_tensor_extra_index + 1) % GGML_CUDA_MAX_NODES;
|
2023-10-08 17:19:14 +00:00
|
|
|
ggml_tensor_extra_gpu * extra = &temp_tensor_extras[alloc_index];
|
|
|
|
memset(extra, 0, sizeof(*extra));
|
|
|
|
|
|
|
|
return extra;
|
|
|
|
}
|
|
|
|
};
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static const char * ggml_backend_cuda_buffer_get_name(ggml_backend_buffer_t buffer) {
|
2024-01-12 19:07:38 +00:00
|
|
|
ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
|
|
|
|
return ctx->name.c_str();
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static bool ggml_backend_buffer_is_cuda(ggml_backend_buffer_t buffer) {
|
2024-01-12 19:07:38 +00:00
|
|
|
return buffer->iface.get_name == ggml_backend_cuda_buffer_get_name;
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static void ggml_backend_cuda_buffer_free_buffer(ggml_backend_buffer_t buffer) {
|
2024-01-12 19:07:38 +00:00
|
|
|
ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
|
2023-12-07 20:26:54 +00:00
|
|
|
CUDA_CHECK(cudaFree(ctx->dev_ptr));
|
2023-10-08 17:19:14 +00:00
|
|
|
delete ctx;
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static void * ggml_backend_cuda_buffer_get_base(ggml_backend_buffer_t buffer) {
|
2024-01-12 19:07:38 +00:00
|
|
|
ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
|
2023-12-07 20:26:54 +00:00
|
|
|
return ctx->dev_ptr;
|
2023-10-08 17:19:14 +00:00
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
|
2024-01-12 19:07:38 +00:00
|
|
|
ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
|
2023-10-08 17:19:14 +00:00
|
|
|
|
|
|
|
if (tensor->view_src != NULL && tensor->view_offs == 0) {
|
2023-12-21 20:07:46 +00:00
|
|
|
assert(tensor->view_src->buffer->buft == buffer->buft);
|
2023-10-08 17:19:14 +00:00
|
|
|
tensor->backend = tensor->view_src->backend;
|
|
|
|
tensor->extra = tensor->view_src->extra;
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_tensor_extra_gpu * extra = ctx->ggml_cuda_alloc_temp_tensor_extra();
|
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
extra->data_device[ctx->device] = tensor->data;
|
2023-10-08 17:19:14 +00:00
|
|
|
|
2024-02-25 10:09:09 +00:00
|
|
|
tensor->backend = GGML_BACKEND_TYPE_GPU;
|
2023-10-08 17:19:14 +00:00
|
|
|
tensor->extra = extra;
|
|
|
|
|
|
|
|
if (ggml_is_quantized(tensor->type)) {
|
|
|
|
// initialize padding to 0 to avoid possible NaN values
|
2024-01-26 17:59:43 +00:00
|
|
|
size_t original_size = ggml_nbytes(tensor);
|
2023-12-07 20:26:54 +00:00
|
|
|
size_t padded_size = ggml_backend_buft_get_alloc_size(buffer->buft, tensor);
|
2023-10-08 17:19:14 +00:00
|
|
|
|
|
|
|
if (padded_size > original_size && tensor->view_src == nullptr) {
|
2024-01-26 17:59:43 +00:00
|
|
|
CUDA_CHECK(cudaMemset((char *)tensor->data + original_size, 0, padded_size - original_size));
|
2023-10-08 17:19:14 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static void ggml_backend_cuda_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
|
2024-02-25 10:09:09 +00:00
|
|
|
GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
|
2023-12-07 20:26:54 +00:00
|
|
|
|
2024-01-12 19:07:38 +00:00
|
|
|
ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
|
2023-12-07 20:26:54 +00:00
|
|
|
|
2023-12-21 20:07:46 +00:00
|
|
|
ggml_cuda_set_device(ctx->device);
|
|
|
|
CUDA_CHECK(cudaDeviceSynchronize());
|
|
|
|
CUDA_CHECK(cudaMemcpy((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice));
|
2024-01-05 15:00:00 +00:00
|
|
|
CUDA_CHECK(cudaDeviceSynchronize());
|
2023-12-07 20:26:54 +00:00
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static void ggml_backend_cuda_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
|
2024-02-25 10:09:09 +00:00
|
|
|
GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
|
2023-12-07 20:26:54 +00:00
|
|
|
|
2024-01-12 19:07:38 +00:00
|
|
|
ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
|
2023-12-21 20:07:46 +00:00
|
|
|
|
|
|
|
ggml_cuda_set_device(ctx->device);
|
|
|
|
CUDA_CHECK(cudaDeviceSynchronize());
|
2023-12-07 20:26:54 +00:00
|
|
|
CUDA_CHECK(cudaMemcpy(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost));
|
2024-01-12 19:07:38 +00:00
|
|
|
CUDA_CHECK(cudaDeviceSynchronize());
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static bool ggml_backend_cuda_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
|
2024-01-12 19:07:38 +00:00
|
|
|
if (ggml_backend_buffer_is_cuda(src->buffer)) {
|
|
|
|
ggml_backend_cuda_buffer_context * src_ctx = (ggml_backend_cuda_buffer_context *)src->buffer->context;
|
|
|
|
ggml_backend_cuda_buffer_context * dst_ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
|
|
|
|
|
|
|
|
ggml_cuda_set_device(src_ctx->device);
|
|
|
|
CUDA_CHECK(cudaDeviceSynchronize());
|
|
|
|
ggml_cuda_set_device(dst_ctx->device);
|
|
|
|
CUDA_CHECK(cudaDeviceSynchronize());
|
|
|
|
CUDA_CHECK(cudaMemcpy((char *)dst->data, (const char *)src->data, ggml_nbytes(src), cudaMemcpyDeviceToDevice));
|
|
|
|
CUDA_CHECK(cudaDeviceSynchronize());
|
|
|
|
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
return false;
|
2023-12-21 20:07:46 +00:00
|
|
|
}
|
2023-12-07 20:26:54 +00:00
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static void ggml_backend_cuda_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
|
2024-01-12 19:07:38 +00:00
|
|
|
ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
|
2023-12-21 20:07:46 +00:00
|
|
|
|
|
|
|
ggml_cuda_set_device(ctx->device);
|
|
|
|
CUDA_CHECK(cudaDeviceSynchronize());
|
|
|
|
CUDA_CHECK(cudaMemset(ctx->dev_ptr, value, buffer->size));
|
2024-01-12 19:07:38 +00:00
|
|
|
CUDA_CHECK(cudaDeviceSynchronize());
|
2023-12-07 20:26:54 +00:00
|
|
|
}
|
|
|
|
|
2024-01-12 19:07:38 +00:00
|
|
|
static ggml_backend_buffer_i ggml_backend_cuda_buffer_interface = {
|
|
|
|
/* .get_name = */ ggml_backend_cuda_buffer_get_name,
|
2023-12-07 20:26:54 +00:00
|
|
|
/* .free_buffer = */ ggml_backend_cuda_buffer_free_buffer,
|
|
|
|
/* .get_base = */ ggml_backend_cuda_buffer_get_base,
|
|
|
|
/* .init_tensor = */ ggml_backend_cuda_buffer_init_tensor,
|
|
|
|
/* .set_tensor = */ ggml_backend_cuda_buffer_set_tensor,
|
|
|
|
/* .get_tensor = */ ggml_backend_cuda_buffer_get_tensor,
|
2024-01-12 19:07:38 +00:00
|
|
|
/* .cpy_tensor = */ ggml_backend_cuda_buffer_cpy_tensor,
|
2023-12-21 20:07:46 +00:00
|
|
|
/* .clear = */ ggml_backend_cuda_buffer_clear,
|
2024-01-12 19:07:38 +00:00
|
|
|
/* .reset = */ NULL,
|
2023-10-08 17:19:14 +00:00
|
|
|
};
|
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
// cuda buffer type
|
2024-01-12 19:07:38 +00:00
|
|
|
struct ggml_backend_cuda_buffer_type_context {
|
|
|
|
int device;
|
|
|
|
std::string name;
|
|
|
|
};
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static const char * ggml_backend_cuda_buffer_type_name(ggml_backend_buffer_type_t buft) {
|
2024-01-12 19:07:38 +00:00
|
|
|
ggml_backend_cuda_buffer_type_context * ctx = (ggml_backend_cuda_buffer_type_context *)buft->context;
|
|
|
|
|
|
|
|
return ctx->name.c_str();
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
|
2024-01-12 19:07:38 +00:00
|
|
|
ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)buft->context;
|
2023-10-08 17:19:14 +00:00
|
|
|
|
2024-01-12 19:07:38 +00:00
|
|
|
ggml_cuda_set_device(buft_ctx->device);
|
2023-11-13 12:16:23 +00:00
|
|
|
|
|
|
|
size = std::max(size, (size_t)1); // cudaMalloc returns null for size 0
|
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
void * dev_ptr;
|
2024-01-12 19:07:38 +00:00
|
|
|
cudaError_t err = cudaMalloc(&dev_ptr, size);
|
|
|
|
if (err != cudaSuccess) {
|
|
|
|
fprintf(stderr, "%s: allocating %.2f MiB on device %d: cudaMalloc failed: %s\n", __func__, size/1024.0/1024.0, buft_ctx->device, cudaGetErrorString(err));
|
|
|
|
return nullptr;
|
|
|
|
}
|
2023-11-13 12:16:23 +00:00
|
|
|
|
2024-01-12 19:07:38 +00:00
|
|
|
ggml_backend_cuda_buffer_context * ctx = new ggml_backend_cuda_buffer_context(buft_ctx->device, dev_ptr);
|
2023-12-07 20:26:54 +00:00
|
|
|
|
2024-01-12 19:07:38 +00:00
|
|
|
return ggml_backend_buffer_init(buft, ggml_backend_cuda_buffer_interface, ctx, size);
|
2023-10-08 17:19:14 +00:00
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
|
2023-10-08 17:19:14 +00:00
|
|
|
return 128;
|
2023-12-07 20:26:54 +00:00
|
|
|
|
|
|
|
UNUSED(buft);
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
|
2024-01-26 17:59:43 +00:00
|
|
|
size_t size = ggml_nbytes(tensor);
|
2023-12-07 20:26:54 +00:00
|
|
|
int64_t ne0 = tensor->ne[0];
|
|
|
|
|
|
|
|
if (ggml_is_quantized(tensor->type)) {
|
|
|
|
if (ne0 % MATRIX_ROW_PADDING != 0) {
|
2023-12-14 19:05:21 +00:00
|
|
|
size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
|
2023-12-07 20:26:54 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return size;
|
|
|
|
|
|
|
|
UNUSED(buft);
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static bool ggml_backend_cuda_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
|
2024-01-12 19:07:38 +00:00
|
|
|
if (!ggml_backend_is_cuda(backend)) {
|
|
|
|
return false;
|
|
|
|
}
|
2023-12-07 20:26:54 +00:00
|
|
|
|
2024-01-12 19:07:38 +00:00
|
|
|
ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)buft->context;
|
|
|
|
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
|
|
|
|
|
|
|
|
return buft_ctx->device == cuda_ctx->device;
|
2023-12-07 20:26:54 +00:00
|
|
|
}
|
|
|
|
|
2023-12-21 20:07:46 +00:00
|
|
|
static ggml_backend_buffer_type_i ggml_backend_cuda_buffer_type_interface = {
|
2024-01-12 19:07:38 +00:00
|
|
|
/* .get_name = */ ggml_backend_cuda_buffer_type_name,
|
2023-12-07 20:26:54 +00:00
|
|
|
/* .alloc_buffer = */ ggml_backend_cuda_buffer_type_alloc_buffer,
|
|
|
|
/* .get_alignment = */ ggml_backend_cuda_buffer_type_get_alignment,
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
/* .get_max_size = */ NULL, // defaults to SIZE_MAX
|
2023-12-07 20:26:54 +00:00
|
|
|
/* .get_alloc_size = */ ggml_backend_cuda_buffer_type_get_alloc_size,
|
|
|
|
/* .supports_backend = */ ggml_backend_cuda_buffer_type_supports_backend,
|
2024-01-12 19:07:38 +00:00
|
|
|
/* .is_host = */ NULL,
|
2023-12-07 20:26:54 +00:00
|
|
|
};
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device) {
|
2024-01-12 19:07:38 +00:00
|
|
|
// FIXME: this is not thread safe
|
|
|
|
if (device >= ggml_backend_cuda_get_device_count()) {
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
static ggml_backend_buffer_type ggml_backend_cuda_buffer_types[GGML_CUDA_MAX_DEVICES];
|
2023-12-21 20:07:46 +00:00
|
|
|
|
|
|
|
static bool ggml_backend_cuda_buffer_type_initialized = false;
|
|
|
|
|
|
|
|
if (!ggml_backend_cuda_buffer_type_initialized) {
|
2023-12-07 20:26:54 +00:00
|
|
|
for (int i = 0; i < GGML_CUDA_MAX_DEVICES; i++) {
|
2023-12-21 20:07:46 +00:00
|
|
|
ggml_backend_cuda_buffer_types[i] = {
|
|
|
|
/* .iface = */ ggml_backend_cuda_buffer_type_interface,
|
2024-01-12 19:07:38 +00:00
|
|
|
/* .context = */ new ggml_backend_cuda_buffer_type_context{i, GGML_CUDA_NAME + std::to_string(i)},
|
2023-12-07 20:26:54 +00:00
|
|
|
};
|
|
|
|
}
|
2023-12-21 20:07:46 +00:00
|
|
|
ggml_backend_cuda_buffer_type_initialized = true;
|
2023-12-07 20:26:54 +00:00
|
|
|
}
|
|
|
|
|
2023-12-21 20:07:46 +00:00
|
|
|
return &ggml_backend_cuda_buffer_types[device];
|
2023-12-07 20:26:54 +00:00
|
|
|
}
|
|
|
|
|
2024-01-12 19:07:38 +00:00
|
|
|
// cuda split buffer
|
|
|
|
|
|
|
|
struct ggml_backend_cuda_split_buffer_context {
|
|
|
|
~ggml_backend_cuda_split_buffer_context() {
|
|
|
|
for (ggml_tensor_extra_gpu * extra : tensor_extras) {
|
|
|
|
for (int id = 0; id < g_device_count; ++id) {
|
|
|
|
for (int64_t is = 0; is < MAX_STREAMS; ++is) {
|
|
|
|
if (extra->events[id][is] != nullptr) {
|
|
|
|
CUDA_CHECK(cudaEventDestroy(extra->events[id][is]));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (extra->data_device[id] != nullptr) {
|
|
|
|
CUDA_CHECK(cudaFree(extra->data_device[id]));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
delete extra;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
std::vector<ggml_tensor_extra_gpu *> tensor_extras;
|
|
|
|
};
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static const char * ggml_backend_cuda_split_buffer_get_name(ggml_backend_buffer_t buffer) {
|
2024-01-12 19:07:38 +00:00
|
|
|
return GGML_CUDA_NAME "_Split";
|
|
|
|
|
|
|
|
UNUSED(buffer);
|
|
|
|
}
|
|
|
|
|
2024-02-19 22:40:26 +00:00
|
|
|
static bool ggml_backend_buffer_is_cuda_split(ggml_backend_buffer_t buffer) {
|
|
|
|
return buffer->iface.get_name == ggml_backend_cuda_split_buffer_get_name;
|
|
|
|
UNUSED(ggml_backend_buffer_is_cuda_split); // only used in debug builds currently, avoid unused function warning in release builds
|
|
|
|
}
|
2024-01-12 19:07:38 +00:00
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static void ggml_backend_cuda_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
|
2024-01-12 19:07:38 +00:00
|
|
|
ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context;
|
|
|
|
delete ctx;
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static void * ggml_backend_cuda_split_buffer_get_base(ggml_backend_buffer_t buffer) {
|
2024-01-12 19:07:38 +00:00
|
|
|
// the pointers are stored in the tensor extras, this is just a dummy address and never dereferenced
|
|
|
|
return (void *)0x1000;
|
|
|
|
|
|
|
|
UNUSED(buffer);
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static void ggml_backend_cuda_split_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
|
2024-01-12 19:07:38 +00:00
|
|
|
GGML_ASSERT(tensor->view_src == nullptr); // views of split tensors are not supported
|
|
|
|
|
|
|
|
ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context;
|
|
|
|
ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *)buffer->buft->context;
|
|
|
|
|
|
|
|
const int64_t ne0 = tensor->ne[0];
|
|
|
|
|
|
|
|
ggml_tensor_extra_gpu * extra = new ggml_tensor_extra_gpu{};
|
|
|
|
|
|
|
|
ctx->tensor_extras.push_back(extra);
|
|
|
|
|
|
|
|
for (int id = 0; id < g_device_count; ++id) {
|
|
|
|
int64_t row_low, row_high;
|
|
|
|
get_row_split(&row_low, &row_high, tensor, buft_ctx->tensor_split, id);
|
|
|
|
|
|
|
|
int64_t nrows_split = row_high - row_low;
|
|
|
|
if (nrows_split == 0) {
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
size_t size = ggml_nbytes_split(tensor, nrows_split);
|
|
|
|
const size_t original_size = size;
|
|
|
|
|
|
|
|
// pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
|
|
|
|
if (ne0 % MATRIX_ROW_PADDING != 0) {
|
|
|
|
size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
|
|
|
|
}
|
|
|
|
|
|
|
|
// FIXME: do not crash if cudaMalloc fails
|
|
|
|
// currently, init_tensor cannot fail, it needs to be fixed in ggml-backend first
|
|
|
|
ggml_cuda_set_device(id);
|
|
|
|
char * buf;
|
|
|
|
CUDA_CHECK(cudaMalloc(&buf, size));
|
|
|
|
|
|
|
|
// set padding to 0 to avoid possible NaN values
|
|
|
|
if (size > original_size) {
|
|
|
|
CUDA_CHECK(cudaMemset(buf + original_size, 0, size - original_size));
|
|
|
|
}
|
|
|
|
|
|
|
|
extra->data_device[id] = buf;
|
|
|
|
|
|
|
|
for (int64_t is = 0; is < MAX_STREAMS; ++is) {
|
|
|
|
CUDA_CHECK(cudaEventCreateWithFlags(&extra->events[id][is], cudaEventDisableTiming));
|
|
|
|
}
|
|
|
|
}
|
2024-02-25 10:09:09 +00:00
|
|
|
tensor->backend = GGML_BACKEND_TYPE_GPU_SPLIT;
|
2024-01-12 19:07:38 +00:00
|
|
|
tensor->extra = extra;
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static void ggml_backend_cuda_split_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
|
2024-01-12 19:07:38 +00:00
|
|
|
// split tensors must always be set in their entirety at once
|
|
|
|
GGML_ASSERT(offset == 0);
|
|
|
|
GGML_ASSERT(size == ggml_nbytes(tensor));
|
|
|
|
|
|
|
|
ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *)buffer->buft->context;
|
|
|
|
|
|
|
|
const int64_t ne0 = tensor->ne[0];
|
|
|
|
const size_t nb1 = tensor->nb[1];
|
|
|
|
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *)tensor->extra;
|
|
|
|
|
|
|
|
for (int id = 0; id < g_device_count; ++id) {
|
|
|
|
int64_t row_low, row_high;
|
|
|
|
get_row_split(&row_low, &row_high, tensor, buft_ctx->tensor_split, id);
|
|
|
|
|
|
|
|
int64_t nrows_split = row_high - row_low;
|
|
|
|
if (nrows_split == 0) {
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
const size_t offset_split = row_low*nb1;
|
|
|
|
size_t size = ggml_nbytes_split(tensor, nrows_split);
|
|
|
|
const size_t original_size = size;
|
|
|
|
|
|
|
|
// pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
|
|
|
|
if (ne0 % MATRIX_ROW_PADDING != 0) {
|
|
|
|
size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
|
|
|
|
}
|
|
|
|
|
|
|
|
const char * buf_host = (const char *)data + offset_split;
|
|
|
|
CUDA_CHECK(cudaMemcpy(extra->data_device[id], buf_host, original_size, cudaMemcpyHostToDevice));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static void ggml_backend_cuda_split_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
|
2024-01-12 19:07:38 +00:00
|
|
|
// split tensors must always be set in their entirety at once
|
|
|
|
GGML_ASSERT(offset == 0);
|
|
|
|
GGML_ASSERT(size == ggml_nbytes(tensor));
|
|
|
|
|
|
|
|
ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *)buffer->buft->context;
|
|
|
|
|
|
|
|
const int64_t ne0 = tensor->ne[0];
|
|
|
|
const size_t nb1 = tensor->nb[1];
|
|
|
|
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *)tensor->extra;
|
|
|
|
|
|
|
|
for (int id = 0; id < g_device_count; ++id) {
|
|
|
|
int64_t row_low, row_high;
|
|
|
|
get_row_split(&row_low, &row_high, tensor, buft_ctx->tensor_split, id);
|
|
|
|
|
|
|
|
int64_t nrows_split = row_high - row_low;
|
|
|
|
if (nrows_split == 0) {
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
const size_t offset_split = row_low*nb1;
|
|
|
|
size_t size = ggml_nbytes_split(tensor, nrows_split);
|
|
|
|
const size_t original_size = size;
|
|
|
|
|
|
|
|
// pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
|
|
|
|
if (ne0 % MATRIX_ROW_PADDING != 0) {
|
|
|
|
size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
|
|
|
|
}
|
|
|
|
|
|
|
|
char * buf_host = (char *)data + offset_split;
|
|
|
|
CUDA_CHECK(cudaMemcpy(buf_host, extra->data_device[id], original_size, cudaMemcpyDeviceToHost));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static void ggml_backend_cuda_split_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
|
2024-01-12 19:07:38 +00:00
|
|
|
UNUSED(buffer);
|
|
|
|
UNUSED(value);
|
|
|
|
}
|
|
|
|
|
|
|
|
static struct ggml_backend_buffer_i ggml_backend_cuda_split_buffer_interface = {
|
|
|
|
/* .get_name = */ ggml_backend_cuda_split_buffer_get_name,
|
|
|
|
/* .free_buffer = */ ggml_backend_cuda_split_buffer_free_buffer,
|
|
|
|
/* .get_base = */ ggml_backend_cuda_split_buffer_get_base,
|
|
|
|
/* .init_tensor = */ ggml_backend_cuda_split_buffer_init_tensor,
|
|
|
|
/* .set_tensor = */ ggml_backend_cuda_split_buffer_set_tensor,
|
|
|
|
/* .get_tensor = */ ggml_backend_cuda_split_buffer_get_tensor,
|
|
|
|
/* .cpy_tensor = */ NULL,
|
|
|
|
/* .clear = */ ggml_backend_cuda_split_buffer_clear,
|
|
|
|
/* .reset = */ NULL,
|
|
|
|
};
|
|
|
|
|
|
|
|
// cuda split buffer type
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static const char * ggml_backend_cuda_split_buffer_type_name(ggml_backend_buffer_type_t buft) {
|
2024-01-12 19:07:38 +00:00
|
|
|
return GGML_CUDA_NAME "_Split";
|
|
|
|
|
|
|
|
UNUSED(buft);
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_split_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
|
2024-01-12 19:07:38 +00:00
|
|
|
// since we don't know the exact split after rounding, we cannot allocate the device buffers at this point
|
|
|
|
// instead, we allocate them for each tensor separately in init_tensor
|
|
|
|
// however, the size still represents the maximum cumulative size of all the device buffers after the tensors are allocated,
|
|
|
|
// as returned by get_alloc_size. this limit is enforced during tensor allocation by ggml-alloc, so it must be correct.
|
|
|
|
ggml_backend_cuda_split_buffer_context * ctx = new ggml_backend_cuda_split_buffer_context();
|
|
|
|
|
|
|
|
return ggml_backend_buffer_init(buft, ggml_backend_cuda_split_buffer_interface, ctx, size);
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static size_t ggml_backend_cuda_split_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
|
2024-01-12 19:07:38 +00:00
|
|
|
return 128;
|
|
|
|
|
|
|
|
UNUSED(buft);
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static size_t ggml_backend_cuda_split_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
|
2024-01-12 19:07:38 +00:00
|
|
|
ggml_backend_cuda_split_buffer_type_context * ctx = (ggml_backend_cuda_split_buffer_type_context *)buft->context;
|
|
|
|
|
|
|
|
size_t total_size = 0;
|
|
|
|
|
|
|
|
const int64_t ne0 = tensor->ne[0];
|
|
|
|
|
|
|
|
for (int id = 0; id < g_device_count; ++id) {
|
|
|
|
int64_t row_low, row_high;
|
|
|
|
get_row_split(&row_low, &row_high, tensor, ctx->tensor_split, id);
|
|
|
|
|
|
|
|
int64_t nrows_split = row_high - row_low;
|
|
|
|
if (nrows_split == 0) {
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
total_size += ggml_nbytes_split(tensor, nrows_split);
|
|
|
|
|
|
|
|
// pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
|
|
|
|
if (ne0 % MATRIX_ROW_PADDING != 0) {
|
|
|
|
total_size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return total_size;
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static bool ggml_backend_cuda_split_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
|
2024-01-12 19:07:38 +00:00
|
|
|
return ggml_backend_is_cuda(backend);
|
|
|
|
|
|
|
|
UNUSED(buft);
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static bool ggml_backend_cuda_split_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
|
2024-01-12 19:07:38 +00:00
|
|
|
return false;
|
|
|
|
|
|
|
|
UNUSED(buft);
|
|
|
|
}
|
|
|
|
|
|
|
|
static ggml_backend_buffer_type_i ggml_backend_cuda_split_buffer_type_interface = {
|
|
|
|
/* .get_name = */ ggml_backend_cuda_split_buffer_type_name,
|
|
|
|
/* .alloc_buffer = */ ggml_backend_cuda_split_buffer_type_alloc_buffer,
|
|
|
|
/* .get_alignment = */ ggml_backend_cuda_split_buffer_type_get_alignment,
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
/* .get_max_size = */ NULL, // defaults to SIZE_MAX
|
2024-01-12 19:07:38 +00:00
|
|
|
/* .get_alloc_size = */ ggml_backend_cuda_split_buffer_type_get_alloc_size,
|
|
|
|
/* .supports_backend = */ ggml_backend_cuda_split_buffer_type_supports_backend,
|
|
|
|
/* .is_host = */ ggml_backend_cuda_split_buffer_type_is_host,
|
|
|
|
};
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split) {
|
2024-01-12 19:07:38 +00:00
|
|
|
// FIXME: this is not thread safe
|
|
|
|
static std::map<std::array<float, GGML_CUDA_MAX_DEVICES>, struct ggml_backend_buffer_type> buft_map;
|
|
|
|
|
|
|
|
std::array<float, GGML_CUDA_MAX_DEVICES> tensor_split_arr = {};
|
|
|
|
|
|
|
|
bool all_zero = tensor_split == nullptr || std::all_of(tensor_split, tensor_split + GGML_CUDA_MAX_DEVICES, [](float x) { return x == 0.0f; });
|
|
|
|
if (all_zero) {
|
|
|
|
tensor_split_arr = g_default_tensor_split;
|
|
|
|
} else {
|
|
|
|
float split_sum = 0.0f;
|
|
|
|
for (int i = 0; i < g_device_count; ++i) {
|
|
|
|
tensor_split_arr[i] = split_sum;
|
|
|
|
split_sum += tensor_split[i];
|
|
|
|
}
|
|
|
|
for (int i = 0; i < g_device_count; ++i) {
|
|
|
|
tensor_split_arr[i] /= split_sum;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
auto it = buft_map.find(tensor_split_arr);
|
|
|
|
if (it != buft_map.end()) {
|
|
|
|
return &it->second;
|
|
|
|
}
|
|
|
|
|
|
|
|
struct ggml_backend_buffer_type buft {
|
|
|
|
/* .iface = */ ggml_backend_cuda_split_buffer_type_interface,
|
|
|
|
/* .context = */ new ggml_backend_cuda_split_buffer_type_context{tensor_split_arr},
|
|
|
|
};
|
|
|
|
|
|
|
|
auto result = buft_map.emplace(tensor_split_arr, buft);
|
|
|
|
return &result.first->second;
|
|
|
|
}
|
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
// host buffer type
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static const char * ggml_backend_cuda_host_buffer_type_name(ggml_backend_buffer_type_t buft) {
|
2024-01-12 19:07:38 +00:00
|
|
|
return GGML_CUDA_NAME "_Host";
|
|
|
|
|
|
|
|
UNUSED(buft);
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static const char * ggml_backend_cuda_host_buffer_name(ggml_backend_buffer_t buffer) {
|
2024-01-12 19:07:38 +00:00
|
|
|
return GGML_CUDA_NAME "_Host";
|
|
|
|
|
|
|
|
UNUSED(buffer);
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
|
2023-12-23 15:10:51 +00:00
|
|
|
ggml_cuda_host_free(buffer->context);
|
2023-12-07 20:26:54 +00:00
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
|
2023-12-23 15:10:51 +00:00
|
|
|
void * ptr = ggml_cuda_host_malloc(size);
|
2023-12-24 13:34:22 +00:00
|
|
|
|
2023-12-23 15:10:51 +00:00
|
|
|
if (ptr == nullptr) {
|
2023-12-24 13:34:22 +00:00
|
|
|
// fallback to cpu buffer
|
|
|
|
return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
|
2023-12-23 15:10:51 +00:00
|
|
|
}
|
2023-12-07 20:26:54 +00:00
|
|
|
|
|
|
|
ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
|
|
|
|
buffer->buft = buft;
|
2024-01-12 19:07:38 +00:00
|
|
|
buffer->iface.get_name = ggml_backend_cuda_host_buffer_name;
|
2023-12-07 20:26:54 +00:00
|
|
|
buffer->iface.free_buffer = ggml_backend_cuda_host_buffer_free_buffer;
|
|
|
|
|
|
|
|
return buffer;
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type() {
|
2023-12-21 20:07:46 +00:00
|
|
|
static struct ggml_backend_buffer_type ggml_backend_cuda_buffer_type_host = {
|
|
|
|
/* .iface = */ {
|
2024-01-12 19:07:38 +00:00
|
|
|
/* .get_name = */ ggml_backend_cuda_host_buffer_type_name,
|
2023-12-21 20:07:46 +00:00
|
|
|
/* .alloc_buffer = */ ggml_backend_cuda_host_buffer_type_alloc_buffer,
|
|
|
|
/* .get_alignment = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
/* .get_max_size = */ NULL, // defaults to SIZE_MAX
|
2023-12-21 20:07:46 +00:00
|
|
|
/* .get_alloc_size = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
|
|
|
|
/* .supports_backend = */ ggml_backend_cpu_buffer_type()->iface.supports_backend,
|
|
|
|
/* .is_host = */ ggml_backend_cpu_buffer_type()->iface.is_host,
|
|
|
|
},
|
2023-12-07 20:26:54 +00:00
|
|
|
/* .context = */ nullptr,
|
|
|
|
};
|
|
|
|
|
2023-12-21 20:07:46 +00:00
|
|
|
return &ggml_backend_cuda_buffer_type_host;
|
2023-12-07 20:26:54 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// backend
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static const char * ggml_backend_cuda_name(ggml_backend_t backend) {
|
2024-01-12 19:07:38 +00:00
|
|
|
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
|
2023-12-07 20:26:54 +00:00
|
|
|
|
2024-01-12 19:07:38 +00:00
|
|
|
return cuda_ctx->name.c_str();
|
2023-10-08 17:19:14 +00:00
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static void ggml_backend_cuda_free(ggml_backend_t backend) {
|
2024-01-12 19:07:38 +00:00
|
|
|
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
|
2023-12-07 20:26:54 +00:00
|
|
|
|
|
|
|
delete cuda_ctx;
|
|
|
|
delete backend;
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static ggml_backend_buffer_type_t ggml_backend_cuda_get_default_buffer_type(ggml_backend_t backend) {
|
2024-01-12 19:07:38 +00:00
|
|
|
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
|
2023-12-07 20:26:54 +00:00
|
|
|
|
|
|
|
return ggml_backend_cuda_buffer_type(cuda_ctx->device);
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static void ggml_backend_cuda_set_tensor_async(ggml_backend_t backend, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
|
2024-01-12 19:07:38 +00:00
|
|
|
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
|
2023-12-07 20:26:54 +00:00
|
|
|
|
|
|
|
GGML_ASSERT(tensor->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");
|
2024-02-25 10:09:09 +00:00
|
|
|
GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
|
2023-10-08 17:19:14 +00:00
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
CUDA_CHECK(cudaMemcpyAsync((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice, g_cudaStreams[cuda_ctx->device][0]));
|
2023-10-08 17:19:14 +00:00
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static void ggml_backend_cuda_get_tensor_async(ggml_backend_t backend, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
|
2024-01-12 19:07:38 +00:00
|
|
|
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
|
2023-12-07 20:26:54 +00:00
|
|
|
|
|
|
|
GGML_ASSERT(tensor->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");
|
2024-02-25 10:09:09 +00:00
|
|
|
GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
|
2023-10-08 17:19:14 +00:00
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
CUDA_CHECK(cudaMemcpyAsync(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost, g_cudaStreams[cuda_ctx->device][0]));
|
2023-10-08 17:19:14 +00:00
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend, const ggml_tensor * src, ggml_tensor * dst) {
|
2024-01-12 19:07:38 +00:00
|
|
|
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
|
2023-10-08 17:19:14 +00:00
|
|
|
|
2024-01-12 19:07:38 +00:00
|
|
|
if (dst->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && ggml_backend_buffer_is_cuda(src->buffer)) {
|
|
|
|
CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, g_cudaStreams[cuda_ctx->device][0]));
|
|
|
|
return true;
|
|
|
|
}
|
2023-10-08 17:19:14 +00:00
|
|
|
|
2024-01-12 19:07:38 +00:00
|
|
|
return false;
|
2023-10-08 17:19:14 +00:00
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
|
2024-01-12 19:07:38 +00:00
|
|
|
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
|
2023-10-08 17:19:14 +00:00
|
|
|
|
2024-01-12 19:07:38 +00:00
|
|
|
CUDA_CHECK(cudaStreamSynchronize(g_cudaStreams[cuda_ctx->device][0]));
|
2023-10-08 17:19:14 +00:00
|
|
|
|
|
|
|
UNUSED(backend);
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static bool ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
|
2024-01-12 19:07:38 +00:00
|
|
|
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
|
2023-12-07 20:26:54 +00:00
|
|
|
|
|
|
|
ggml_cuda_set_main_device(cuda_ctx->device);
|
2023-10-08 17:19:14 +00:00
|
|
|
|
|
|
|
ggml_compute_params params = {};
|
2024-02-25 10:09:09 +00:00
|
|
|
params.type = GGML_TASK_TYPE_COMPUTE;
|
2023-10-08 17:19:14 +00:00
|
|
|
params.ith = 0;
|
|
|
|
for (int i = 0; i < cgraph->n_nodes; i++) {
|
|
|
|
ggml_tensor * node = cgraph->nodes[i];
|
|
|
|
|
2024-01-12 19:07:38 +00:00
|
|
|
if (node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
|
2023-11-13 12:16:23 +00:00
|
|
|
continue;
|
2024-01-12 19:07:38 +00:00
|
|
|
}
|
2023-12-07 20:26:54 +00:00
|
|
|
|
2024-01-12 19:07:38 +00:00
|
|
|
#ifndef NDEBUG
|
2024-02-25 10:09:09 +00:00
|
|
|
assert(node->backend == GGML_BACKEND_TYPE_GPU || node->backend == GGML_BACKEND_TYPE_GPU_SPLIT);
|
2023-12-07 20:26:54 +00:00
|
|
|
assert(node->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device));
|
|
|
|
assert(node->extra != nullptr);
|
|
|
|
|
2023-10-08 17:19:14 +00:00
|
|
|
for (int j = 0; j < GGML_MAX_SRC; j++) {
|
|
|
|
if (node->src[j] != nullptr) {
|
2024-02-25 10:09:09 +00:00
|
|
|
assert(node->src[j]->backend == GGML_BACKEND_TYPE_GPU || node->src[j]->backend == GGML_BACKEND_TYPE_GPU_SPLIT);
|
2024-02-19 22:40:26 +00:00
|
|
|
assert(node->src[j]->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) || ggml_backend_buffer_is_cuda_split(node->src[j]->buffer));
|
2023-12-07 20:26:54 +00:00
|
|
|
assert(node->src[j]->extra != nullptr);
|
2023-10-08 17:19:14 +00:00
|
|
|
}
|
|
|
|
}
|
2024-01-12 19:07:38 +00:00
|
|
|
#endif
|
2023-10-08 17:19:14 +00:00
|
|
|
|
|
|
|
bool ok = ggml_cuda_compute_forward(¶ms, node);
|
|
|
|
if (!ok) {
|
|
|
|
fprintf(stderr, "%s: error: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
|
|
|
|
}
|
|
|
|
GGML_ASSERT(ok);
|
|
|
|
}
|
|
|
|
|
2024-01-03 13:39:43 +00:00
|
|
|
return true;
|
2023-10-08 17:19:14 +00:00
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
|
2023-12-07 20:26:54 +00:00
|
|
|
switch (op->op) {
|
|
|
|
case GGML_OP_UNARY:
|
|
|
|
switch (ggml_get_unary_op(op)) {
|
|
|
|
case GGML_UNARY_OP_GELU:
|
|
|
|
case GGML_UNARY_OP_SILU:
|
|
|
|
case GGML_UNARY_OP_RELU:
|
2024-01-31 13:10:15 +00:00
|
|
|
case GGML_UNARY_OP_HARDSIGMOID:
|
|
|
|
case GGML_UNARY_OP_HARDSWISH:
|
2023-12-13 19:54:54 +00:00
|
|
|
case GGML_UNARY_OP_GELU_QUICK:
|
|
|
|
case GGML_UNARY_OP_TANH:
|
2023-12-07 20:26:54 +00:00
|
|
|
return true;
|
|
|
|
default:
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case GGML_OP_MUL_MAT:
|
|
|
|
case GGML_OP_MUL_MAT_ID:
|
|
|
|
{
|
|
|
|
struct ggml_tensor * a;
|
|
|
|
struct ggml_tensor * b;
|
|
|
|
if (op->op == GGML_OP_MUL_MAT) {
|
|
|
|
a = op->src[0];
|
|
|
|
b = op->src[1];
|
|
|
|
} else {
|
|
|
|
a = op->src[2];
|
|
|
|
b = op->src[1];
|
|
|
|
}
|
|
|
|
if (a->ne[3] != b->ne[3]) {
|
|
|
|
return false;
|
|
|
|
}
|
2024-01-17 16:54:56 +00:00
|
|
|
ggml_type a_type = a->type;
|
2024-02-21 09:39:52 +00:00
|
|
|
if (a_type == GGML_TYPE_IQ2_XXS || a_type == GGML_TYPE_IQ2_XS || a_type == GGML_TYPE_IQ3_XXS ||
|
2024-02-26 16:28:38 +00:00
|
|
|
a_type == GGML_TYPE_IQ1_S || a_type == GGML_TYPE_IQ4_NL || a_type == GGML_TYPE_IQ3_S ||
|
2024-02-27 14:34:24 +00:00
|
|
|
a_type == GGML_TYPE_IQ2_S || a_type == GGML_TYPE_IQ4_XS) {
|
2024-01-17 16:54:56 +00:00
|
|
|
if (b->ne[1] == 1 && ggml_nrows(b) > 1) {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
2023-12-07 20:26:54 +00:00
|
|
|
return true;
|
|
|
|
} break;
|
2023-12-13 12:04:25 +00:00
|
|
|
case GGML_OP_GET_ROWS:
|
|
|
|
{
|
|
|
|
switch (op->src[0]->type) {
|
|
|
|
case GGML_TYPE_F16:
|
|
|
|
case GGML_TYPE_F32:
|
|
|
|
case GGML_TYPE_Q4_0:
|
|
|
|
case GGML_TYPE_Q4_1:
|
|
|
|
case GGML_TYPE_Q5_0:
|
|
|
|
case GGML_TYPE_Q5_1:
|
|
|
|
case GGML_TYPE_Q8_0:
|
|
|
|
return true;
|
|
|
|
default:
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
} break;
|
|
|
|
case GGML_OP_CPY:
|
|
|
|
{
|
|
|
|
ggml_type src0_type = op->src[0]->type;
|
|
|
|
ggml_type src1_type = op->src[1]->type;
|
|
|
|
if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q8_0) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q4_0) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q4_1) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
|
|
|
|
return true;
|
|
|
|
}
|
2024-01-29 12:37:33 +00:00
|
|
|
if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
|
|
|
|
return true;
|
|
|
|
}
|
2023-12-13 12:04:25 +00:00
|
|
|
return false;
|
|
|
|
} break;
|
2024-01-03 11:01:44 +00:00
|
|
|
case GGML_OP_DUP:
|
|
|
|
case GGML_OP_REPEAT:
|
|
|
|
case GGML_OP_CONCAT:
|
|
|
|
{
|
|
|
|
ggml_type src0_type = op->src[0]->type;
|
2024-01-03 12:18:46 +00:00
|
|
|
return src0_type != GGML_TYPE_I32 && src0_type != GGML_TYPE_I16;
|
2024-01-03 11:01:44 +00:00
|
|
|
} break;
|
2023-12-07 20:26:54 +00:00
|
|
|
case GGML_OP_NONE:
|
|
|
|
case GGML_OP_RESHAPE:
|
|
|
|
case GGML_OP_VIEW:
|
|
|
|
case GGML_OP_PERMUTE:
|
|
|
|
case GGML_OP_TRANSPOSE:
|
|
|
|
case GGML_OP_NORM:
|
|
|
|
case GGML_OP_ADD:
|
|
|
|
case GGML_OP_MUL:
|
|
|
|
case GGML_OP_DIV:
|
|
|
|
case GGML_OP_RMS_NORM:
|
|
|
|
case GGML_OP_SCALE:
|
|
|
|
case GGML_OP_SQR:
|
|
|
|
case GGML_OP_CLAMP:
|
|
|
|
case GGML_OP_CONT:
|
|
|
|
case GGML_OP_DIAG_MASK_INF:
|
|
|
|
case GGML_OP_SOFT_MAX:
|
|
|
|
case GGML_OP_ROPE:
|
|
|
|
case GGML_OP_ALIBI:
|
|
|
|
case GGML_OP_IM2COL:
|
2024-01-31 13:10:15 +00:00
|
|
|
case GGML_OP_POOL_2D:
|
2023-12-07 20:26:54 +00:00
|
|
|
case GGML_OP_SUM_ROWS:
|
|
|
|
case GGML_OP_ARGSORT:
|
2023-12-13 19:54:54 +00:00
|
|
|
case GGML_OP_ACC:
|
|
|
|
case GGML_OP_GROUP_NORM:
|
|
|
|
case GGML_OP_UPSCALE:
|
|
|
|
case GGML_OP_PAD:
|
|
|
|
case GGML_OP_LEAKY_RELU:
|
2023-12-07 20:26:54 +00:00
|
|
|
return true;
|
|
|
|
default:
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
UNUSED(backend);
|
|
|
|
}
|
|
|
|
|
2024-01-12 19:07:38 +00:00
|
|
|
static ggml_backend_i ggml_backend_cuda_interface = {
|
2023-12-07 20:26:54 +00:00
|
|
|
/* .get_name = */ ggml_backend_cuda_name,
|
|
|
|
/* .free = */ ggml_backend_cuda_free,
|
|
|
|
/* .get_default_buffer_type = */ ggml_backend_cuda_get_default_buffer_type,
|
|
|
|
/* .set_tensor_async = */ ggml_backend_cuda_set_tensor_async,
|
|
|
|
/* .get_tensor_async = */ ggml_backend_cuda_get_tensor_async,
|
2024-01-12 19:07:38 +00:00
|
|
|
/* .cpy_tensor_async = */ ggml_backend_cuda_cpy_tensor_async,
|
2023-12-07 20:26:54 +00:00
|
|
|
/* .synchronize = */ ggml_backend_cuda_synchronize,
|
2024-01-12 19:07:38 +00:00
|
|
|
/* .graph_plan_create = */ NULL,
|
|
|
|
/* .graph_plan_free = */ NULL,
|
|
|
|
/* .graph_plan_compute = */ NULL,
|
2023-12-07 20:26:54 +00:00
|
|
|
/* .graph_compute = */ ggml_backend_cuda_graph_compute,
|
|
|
|
/* .supports_op = */ ggml_backend_cuda_supports_op,
|
2023-10-08 17:19:14 +00:00
|
|
|
};
|
|
|
|
|
2024-02-24 16:27:36 +00:00
|
|
|
static ggml_guid_t ggml_backend_cuda_guid() {
|
|
|
|
static ggml_guid guid = { 0x2c, 0xdd, 0xe8, 0x1c, 0x65, 0xb3, 0x65, 0x73, 0x6a, 0x12, 0x88, 0x61, 0x1c, 0xc9, 0xdc, 0x25 };
|
|
|
|
return &guid;
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device) {
|
2023-10-08 17:19:14 +00:00
|
|
|
ggml_init_cublas(); // TODO: remove from ggml.c
|
|
|
|
|
2023-12-07 20:26:54 +00:00
|
|
|
if (device < 0 || device >= ggml_cuda_get_device_count()) {
|
|
|
|
fprintf(stderr, "%s: error: invalid device %d\n", __func__, device);
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
// not strictly necessary, but it may reduce the overhead of the first graph_compute
|
|
|
|
ggml_cuda_set_main_device(device);
|
|
|
|
|
2024-01-12 19:07:38 +00:00
|
|
|
ggml_backend_cuda_context * ctx = new ggml_backend_cuda_context {
|
|
|
|
/* .device = */ device,
|
|
|
|
/* .name = */ GGML_CUDA_NAME + std::to_string(device),
|
2023-12-07 20:26:54 +00:00
|
|
|
};
|
2023-10-08 17:19:14 +00:00
|
|
|
|
|
|
|
ggml_backend_t cuda_backend = new ggml_backend {
|
2024-02-24 16:27:36 +00:00
|
|
|
/* .guid = */ ggml_backend_cuda_guid(),
|
2024-01-12 19:07:38 +00:00
|
|
|
/* .interface = */ ggml_backend_cuda_interface,
|
2023-10-08 17:19:14 +00:00
|
|
|
/* .context = */ ctx
|
|
|
|
};
|
|
|
|
|
|
|
|
return cuda_backend;
|
|
|
|
}
|
2023-12-07 20:26:54 +00:00
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL bool ggml_backend_is_cuda(ggml_backend_t backend) {
|
2024-02-24 16:27:36 +00:00
|
|
|
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cuda_guid());
|
2024-01-12 19:07:38 +00:00
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL int ggml_backend_cuda_get_device_count() {
|
2024-01-12 19:07:38 +00:00
|
|
|
return ggml_cuda_get_device_count();
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size) {
|
2024-01-12 19:07:38 +00:00
|
|
|
ggml_cuda_get_device_description(device, description, description_size);
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total) {
|
2024-01-12 19:07:38 +00:00
|
|
|
ggml_cuda_set_device(device);
|
|
|
|
|
|
|
|
CUDA_CHECK(cudaMemGetInfo(free, total));
|
2023-12-07 20:26:54 +00:00
|
|
|
}
|
|
|
|
|
2024-01-12 19:07:38 +00:00
|
|
|
// backend registry
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL static ggml_backend_t ggml_backend_reg_cuda_init(const char * params, void * user_data) {
|
2023-12-07 20:26:54 +00:00
|
|
|
ggml_backend_t cuda_backend = ggml_backend_cuda_init((int) (intptr_t) user_data);
|
|
|
|
return cuda_backend;
|
|
|
|
|
|
|
|
UNUSED(params);
|
|
|
|
}
|
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
extern "C" GGML_CALL int ggml_backend_cuda_reg_devices();
|
2023-12-13 12:04:25 +00:00
|
|
|
|
2024-01-16 11:16:33 +00:00
|
|
|
GGML_CALL int ggml_backend_cuda_reg_devices() {
|
2023-12-07 20:26:54 +00:00
|
|
|
int device_count = ggml_cuda_get_device_count();
|
|
|
|
//int device_count = 1; // DEBUG: some tools require delaying CUDA initialization
|
|
|
|
for (int i = 0; i < device_count; i++) {
|
|
|
|
char name[128];
|
|
|
|
snprintf(name, sizeof(name), "%s%d", GGML_CUDA_NAME, i);
|
|
|
|
ggml_backend_register(name, ggml_backend_reg_cuda_init, ggml_backend_cuda_buffer_type(i), (void *) (intptr_t) i);
|
|
|
|
}
|
|
|
|
return device_count;
|
|
|
|
}
|