root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-11-11 13:30:35 +00:00
Code Issues Actions 11 Packages Projects Releases Wiki Activity

llama : add Command R Plus support (#6491)

Browse Source 
* Add Command R Plus GGUF

* Add Command R Plus GGUF

* Loading works up to LayerNorm2D

* Export new tensors in 1D so they are not quantized.

* Fix embedding layer based on Noeda's example

* Whitespace

* Add line

* Fix unexpected tokens on MPS. Re-add F16 fix. ((Noeda)

* dranger003: Fix block index overflow in CUDA dequantizing.

* Reverted blocked multiplication code as it still has issues and could affect other Llama arches

* export norms as f32

* fix overflow issues during quant and other cleanup

* Type convention

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* dranger003: Fix more int overflow during quant.

---------

Co-authored-by: S <seast@Ss-Mac-Studio.local>
Co-authored-by: S <s@example.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Carolinabanana 2024-04-09 09:16:13 +01:00 committed by GitHub
parent e11a8999b5
commit 5dc9dd7152
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
16 changed files with 358 additions and 318 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
convert-hf-to-gguf.py
View File

@ -160,7 +160,7 @@ class Model(ABC):
data = data.astype(np.float32)
# TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
if self.ftype == 1 and data_dtype == np.float16 and (n_dims == 1 or new_name.endswith("_norm.weight")):
data = data.astype(np.float32)
# if f16 desired, convert any float32 2-dim weight tensors to float16

6
ggml-cuda.cu
View File

@ -1225,7 +1225,7 @@ static void ggml_cuda_op_mul_mat_cublas(
// the main device has a larger memory buffer to hold the results from all GPUs
// ldc == nrows of the matrix that cuBLAS writes into
int ldc = id == ctx.device ? ne0 : row_diff;
int64_t ldc = id == ctx.device ? ne0 : row_diff;
const int compute_capability = ggml_cuda_info().devices[id].cc;
@ -1377,8 +1377,8 @@ static void ggml_cuda_op_mul_mat(
const int64_t ne0 = dst->ne[0];
const int64_t ne1 = dst->ne[1];
const int nb2 = dst->nb[2];
const int nb3 = dst->nb[3];
const int64_t nb2 = dst->nb[2];
const int64_t nb3 = dst->nb[3];
GGML_ASSERT(ggml_backend_buffer_is_cuda(dst->buffer));
GGML_ASSERT(ggml_backend_buffer_is_cuda(src1->buffer));

2
ggml-cuda/common.cuh
View File

@ -394,7 +394,7 @@ static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) {
// TODO: move to ggml-common.h
static const __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
typedef void (*dequantize_kernel_t)(const void * vx, const int ib, const int iqs, dfloat2 & v);
typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, dfloat2 & v);
//////////////////////

74
ggml-cuda/convert.cu
View File

@ -4,14 +4,14 @@
#define CUDA_Q8_0_NE_ALIGN 2048
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) {
const int i = 2*(blockDim.x*blockIdx.x + threadIdx.x);
static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k) {
const int64_t i = 2*(blockDim.x*blockIdx.x + threadIdx.x);
if (i >= k) {
return;
}
const int ib = i/qk; // block index
const int64_t 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;
@ -25,7 +25,7 @@ static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __
}
template <bool need_check>
static __global__ void dequantize_block_q8_0_f16(const void * __restrict__ vx, half * __restrict__ y, const int k) {
static __global__ void dequantize_block_q8_0_f16(const void * __restrict__ vx, half * __restrict__ y, const int64_t k) {
#if __CUDA_ARCH__ >= CC_PASCAL
constexpr int nint = CUDA_Q8_0_NE_ALIGN/sizeof(int) + WARP_SIZE;
@ -68,13 +68,13 @@ static __global__ void dequantize_block_q8_0_f16(const void * __restrict__ vx, h
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;
const int64_t 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;
const int64_t ib = 8*i + ir;
if (ib >= nb32) {
return;
}
@ -96,13 +96,13 @@ static __global__ void dequantize_block_q4_0(const void * __restrict__ vx, dst_t
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;
const int64_t 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;
const int64_t ib = 8*i + ir;
if (ib >= nb32) {
return;
}
@ -313,14 +313,14 @@ template<typename dst_t>
static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
const block_q6_K * x = (const block_q6_K *) vx;
const int i = blockIdx.x;
const int64_t i = blockIdx.x;
#if QK_K == 256
// 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;
const int64_t tid = threadIdx.x;
const int64_t ip = tid/32; // ip is 0 or 1
const int64_t il = tid - 32*ip; // 0...32
const int64_t is = 8*ip + il/16;
dst_t * y = yy + i*QK_K + 128*ip + il;
@ -337,9 +337,9 @@ static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t
#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
const int64_t tid = threadIdx.x;
const int64_t ip = tid/16; // 0 or 1
const int64_t il = tid - 16*ip; // 0...15
dst_t * y = yy + i*QK_K + 16*ip + il;
@ -571,12 +571,12 @@ static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst
#endif
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) {
static void dequantize_block_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, cudaStream_t stream) {
const int num_blocks = (k + 2*CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / (2*CUDA_DEQUANTIZE_BLOCK_SIZE);
dequantize_block<qk, qr, dequantize_kernel><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
}
static void dequantize_block_q8_0_f16_cuda(const void * __restrict__ vx, half * __restrict__ y, const int k, cudaStream_t stream) {
static void dequantize_block_q8_0_f16_cuda(const void * __restrict__ vx, half * __restrict__ y, const int64_t 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;
@ -588,7 +588,7 @@ static void dequantize_block_q8_0_f16_cuda(const void * __restrict__ vx, half *
}
template<typename dst_t>
static void dequantize_row_q2_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_q2_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
#if QK_K == 256
dequantize_block_q2_K<<<nb, 64, 0, stream>>>(vx, y);
@ -598,7 +598,7 @@ static void dequantize_row_q2_K_cuda(const void * vx, dst_t * y, const int k, cu
}
template<typename dst_t>
static void dequantize_row_q3_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_q3_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
#if QK_K == 256
dequantize_block_q3_K<<<nb, 64, 0, stream>>>(vx, y);
@ -608,27 +608,27 @@ static void dequantize_row_q3_K_cuda(const void * vx, dst_t * y, const int k, cu
}
template<typename dst_t>
static void dequantize_row_q4_0_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_q4_0_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
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>
static void dequantize_row_q4_1_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_q4_1_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb32 = k / 32;
const int nb = (k + 255) / 256;
dequantize_block_q4_1<<<nb, 32, 0, stream>>>(vx, y, nb32);
}
template<typename dst_t>
static void dequantize_row_q4_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_q4_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_q4_K<<<nb, 32, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_q5_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_q5_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
#if QK_K == 256
dequantize_block_q5_K<<<nb, 64, 0, stream>>>(vx, y);
@ -638,7 +638,7 @@ static void dequantize_row_q5_K_cuda(const void * vx, dst_t * y, const int k, cu
}
template<typename dst_t>
static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
#if QK_K == 256
dequantize_block_q6_K<<<nb, 64, 0, stream>>>(vx, y);
@ -648,55 +648,55 @@ static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int k, cu
}
template<typename dst_t>
static void dequantize_row_iq2_xxs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_iq2_xxs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_iq2_xxs<<<nb, 32, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_iq2_xs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_iq2_xs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_iq2_xs<<<nb, 32, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_iq2_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_iq2_s_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_iq2_s<<<nb, 32, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_iq3_xxs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_iq3_xxs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_iq3_xxs<<<nb, 32, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_iq3_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_iq3_s_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_iq3_s<<<nb, 32, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_iq1_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_iq1_s_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_iq1_s<<<nb, 32, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_iq4_nl_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_iq4_nl_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = (k + QK_K - 1) / QK_K;
dequantize_block_iq4_nl<<<nb, 32, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_iq1_m_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_iq1_m_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_iq1_m<<<nb, 32, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
const int nb = (k + QK_K - 1) / QK_K;
#if QK_K == 64
dequantize_block_iq4_nl<<<nb, 32, 0, stream>>>(vx, y);
@ -706,8 +706,8 @@ static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int k,
}
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;
static __global__ void convert_unary(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k) {
const int64_t i = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
if (i >= k) {
return;
@ -719,7 +719,7 @@ static __global__ void convert_unary(const void * __restrict__ vx, dst_t * __res
}
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) {
static void convert_unary_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t 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);
}

2
ggml-cuda/convert.cuh
View File

@ -3,7 +3,7 @@
#define CUDA_DEQUANTIZE_BLOCK_SIZE 256
template<typename T>
using to_t_cuda_t = void (*)(const void * __restrict__ x, T * __restrict__ y, int k, cudaStream_t stream);
using to_t_cuda_t = void (*)(const void * __restrict__ x, T * __restrict__ y, int64_t k, cudaStream_t stream);
typedef to_t_cuda_t<float> to_fp32_cuda_t;
typedef to_t_cuda_t<half> to_fp16_cuda_t;

10
ggml-cuda/dequantize.cuh
View File

@ -1,6 +1,6 @@
#include "common.cuh"
static __device__ __forceinline__ void dequantize_q4_0(const void * vx, const int ib, const int iqs, dfloat2 & v){
static __device__ __forceinline__ void dequantize_q4_0(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
const block_q4_0 * x = (const block_q4_0 *) vx;
const dfloat d = x[ib].d;
@ -19,7 +19,7 @@ static __device__ __forceinline__ void dequantize_q4_0(const void * vx, const in
#endif // GGML_CUDA_F16
}
static __device__ __forceinline__ void dequantize_q4_1(const void * vx, const int ib, const int iqs, dfloat2 & v){
static __device__ __forceinline__ void dequantize_q4_1(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
const block_q4_1 * x = (const block_q4_1 *) vx;
const dfloat d = __low2half(x[ib].dm);
@ -39,7 +39,7 @@ static __device__ __forceinline__ void dequantize_q4_1(const void * vx, const in
#endif // GGML_CUDA_F16
}
static __device__ __forceinline__ void dequantize_q5_0(const void * vx, const int ib, const int iqs, dfloat2 & v){
static __device__ __forceinline__ void dequantize_q5_0(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
const block_q5_0 * x = (const block_q5_0 *) vx;
const dfloat d = x[ib].d;
@ -62,7 +62,7 @@ static __device__ __forceinline__ void dequantize_q5_0(const void * vx, const in
#endif // GGML_CUDA_F16
}
static __device__ __forceinline__ void dequantize_q5_1(const void * vx, const int ib, const int iqs, dfloat2 & v){
static __device__ __forceinline__ void dequantize_q5_1(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
const block_q5_1 * x = (const block_q5_1 *) vx;
const dfloat d = __low2half(x[ib].dm);
@ -86,7 +86,7 @@ static __device__ __forceinline__ void dequantize_q5_1(const void * vx, const in
#endif // GGML_CUDA_F16
}
static __device__ __forceinline__ void dequantize_q8_0(const void * vx, const int ib, const int iqs, dfloat2 & v){
static __device__ __forceinline__ void dequantize_q8_0(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
const block_q8_0 * x = (const block_q8_0 *) vx;
const dfloat d = x[ib].d;

6
ggml-cuda/dmmv.cu
View File

@ -565,7 +565,7 @@ static __global__ void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx,
}
}
static __device__ void convert_f16(const void * vx, const int ib, const int iqs, dfloat2 & v){
static __device__ void convert_f16(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
const half * x = (const half *) vx;
// automatic half -> float type cast if dfloat == float
@ -577,7 +577,7 @@ template <int qk, int qr, dequantize_kernel_t dequantize_kernel>
static __global__ void dequantize_mul_mat_vec(const void * __restrict__ vx, const dfloat * __restrict__ y, float * __restrict__ dst, const int ncols, const int nrows) {
// qk = quantized weights per x block
// qr = number of quantized weights per data value in x block
const int row = blockIdx.x*blockDim.y + threadIdx.y;
const int64_t row = (int64_t)blockIdx.x*blockDim.y + threadIdx.y;
if (row >= nrows) {
return;
@ -598,7 +598,7 @@ static __global__ void dequantize_mul_mat_vec(const void * __restrict__ vx, cons
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 int64_t ib = ((int64_t)row*ncols + col)/qk; // x block index
const int iqs = (col%qk)/qr; // x quant index
const int iybs = col - col%qk; // y block start index

16
ggml-cuda/quantize.cu
View File

@ -1,20 +1,20 @@
#include "quantize.cuh"
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;
static __global__ void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, const int64_t kx, const int64_t kx_padded) {
const int64_t ix = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
if (ix >= kx_padded) {
return;
}
const int iy = blockDim.y*blockIdx.y + threadIdx.y;
const int64_t iy = (int64_t)blockDim.y*blockIdx.y + threadIdx.y;
const int i_padded = iy*kx_padded + ix;
const int64_t i_padded = (int64_t)iy*kx_padded + ix;
block_q8_1 * y = (block_q8_1 *) vy;
const int ib = i_padded / QK8_1; // block index
const int iqs = i_padded % QK8_1; // quant index
const int64_t ib = i_padded / QK8_1; // block index
const int64_t iqs = i_padded % QK8_1; // quant index
const float xi = ix < kx ? x[iy*kx + ix] : 0.0f;
float amax = fabsf(xi);
@ -36,8 +36,8 @@ static __global__ void quantize_q8_1(const float * __restrict__ x, void * __rest
reinterpret_cast<half&>(y[ib].ds.y) = sum;
}
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;
void quantize_row_q8_1_cuda(const float * x, void * vy, const int64_t kx, const int64_t ky, const int64_t kx_padded, cudaStream_t stream) {
const int64_t 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_QUANTIZE_BLOCK_SIZE, 1, 1);
quantize_q8_1<<<num_blocks, block_size, 0, stream>>>(x, vy, kx, kx_padded);

2
ggml-cuda/quantize.cuh
View File

@ -2,4 +2,4 @@
#define CUDA_QUANTIZE_BLOCK_SIZE 256
void quantize_row_q8_1_cuda(const float * x, void * vy, const int kx, const int ky, const int kx_padded, cudaStream_t stream);
void quantize_row_q8_1_cuda(const float * x, void * vy, const int64_t kx, const int64_t ky, const int64_t kx_padded, cudaStream_t stream);

310
ggml-quants.c
View File

File diff suppressed because it is too large Load Diff

148
ggml-quants.h
View File

@ -12,70 +12,70 @@ extern "C" {
#endif
// Quantization
void quantize_row_q4_0_reference(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int k);
void quantize_row_q4_1_reference(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int k);
void quantize_row_q5_0_reference(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int k);
void quantize_row_q5_1_reference(const float * GGML_RESTRICT x, block_q5_1 * GGML_RESTRICT y, int k);
void quantize_row_q8_0_reference(const float * GGML_RESTRICT x, block_q8_0 * GGML_RESTRICT y, int k);
void quantize_row_q8_1_reference(const float * GGML_RESTRICT x, block_q8_1 * GGML_RESTRICT y, int k);
void quantize_row_q4_0_reference(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t k);
void quantize_row_q4_1_reference(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_0_reference(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_1_reference(const float * GGML_RESTRICT x, block_q5_1 * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_0_reference(const float * GGML_RESTRICT x, block_q8_0 * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_1_reference(const float * GGML_RESTRICT x, block_q8_1 * GGML_RESTRICT y, int64_t k);
void quantize_row_q2_K_reference(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int k);
void quantize_row_q3_K_reference(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int k);
void quantize_row_q4_K_reference(const float * GGML_RESTRICT x, block_q4_K * GGML_RESTRICT y, int k);
void quantize_row_q5_K_reference(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int k);
void quantize_row_q6_K_reference(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int k);
void quantize_row_q8_K_reference(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int k);
void quantize_row_q2_K_reference(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int64_t k);
void quantize_row_q3_K_reference(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t k);
void quantize_row_q4_K_reference(const float * GGML_RESTRICT x, block_q4_K * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_K_reference(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int64_t k);
void quantize_row_q6_K_reference(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_K_reference(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int64_t k);
void quantize_row_iq3_xxs_reference(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int k);
void quantize_row_iq4_nl_reference (const float * GGML_RESTRICT x, block_iq4_nl * GGML_RESTRICT y, int k);
void quantize_row_iq4_xs_reference (const float * GGML_RESTRICT x, block_iq4_xs * GGML_RESTRICT y, int k);
void quantize_row_iq3_s_reference (const float * GGML_RESTRICT x, block_iq3_s * GGML_RESTRICT y, int k);
void quantize_row_iq2_s_reference (const float * GGML_RESTRICT x, block_iq2_s * GGML_RESTRICT y, int k);
void quantize_row_iq3_xxs_reference(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int64_t k);
void quantize_row_iq4_nl_reference (const float * GGML_RESTRICT x, block_iq4_nl * GGML_RESTRICT y, int64_t k);
void quantize_row_iq4_xs_reference (const float * GGML_RESTRICT x, block_iq4_xs * GGML_RESTRICT y, int64_t k);
void quantize_row_iq3_s_reference (const float * GGML_RESTRICT x, block_iq3_s * GGML_RESTRICT y, int64_t k);
void quantize_row_iq2_s_reference (const float * GGML_RESTRICT x, block_iq2_s * GGML_RESTRICT y, int64_t k);
void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_iq3_xxs(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
void quantize_row_iq3_s (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
void quantize_row_iq2_s (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
void quantize_row_iq3_xxs(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_iq3_s (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_iq2_s (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
// Dequantization
void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_q4_1(const block_q4_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_q5_0(const block_q5_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_q5_1(const block_q5_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_q8_0(const block_q8_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
//void dequantize_row_q8_1(const block_q8_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_q4_1(const block_q4_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_q5_0(const block_q5_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_q5_1(const block_q5_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_q8_0(const block_q8_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
//void dequantize_row_q8_1(const block_q8_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_q2_K(const block_q2_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_q3_K(const block_q3_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_q4_K(const block_q4_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_q5_K(const block_q5_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_q6_K(const block_q6_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_q8_K(const block_q8_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_q2_K(const block_q2_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_q3_K(const block_q3_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_q4_K(const block_q4_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_q5_K(const block_q5_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_q6_K(const block_q6_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_q8_K(const block_q8_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_iq2_xs (const block_iq2_xs * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_iq2_s (const block_iq2_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_iq1_s (const block_iq1_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_iq1_m (const block_iq1_m * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_iq4_nl (const block_iq4_nl * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_iq4_xs (const block_iq4_xs * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_iq3_s (const block_iq3_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_iq2_xs (const block_iq2_xs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_iq2_s (const block_iq2_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_iq1_s (const block_iq1_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_iq1_m (const block_iq1_m * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_iq4_nl (const block_iq4_nl * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_iq4_xs (const block_iq4_xs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_iq3_s (const block_iq3_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
// Dot product
void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
@ -101,26 +101,26 @@ void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const
void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
size_t quantize_iq2_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_iq2_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_iq2_s (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_iq3_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_iq1_s (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_iq1_m (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_iq4_nl (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_iq4_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_iq3_s (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_iq2_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_iq2_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_iq2_s (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_iq3_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_iq1_s (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_iq1_m (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_iq4_nl (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_iq4_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_iq3_s (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_q2_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_q3_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_q4_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_q5_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_q6_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_q4_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_q4_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_q2_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_q3_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_q4_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_q5_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_q6_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_q4_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_q4_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void iq2xs_init_impl(enum ggml_type type);
void iq2xs_free_impl(enum ggml_type type);

16
ggml.c
View File

@ -338,14 +338,14 @@ ggml_fp16_t ggml_fp32_to_fp16(float x) {
return GGML_FP32_TO_FP16(x);
}
void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, int n) {
for (int i = 0; i < n; i++) {
void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, int64_t n) {
for (int64_t i = 0; i < n; i++) {
y[i] = GGML_FP16_TO_FP32(x[i]);
}
}
void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, int n) {
int i = 0;
void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, int64_t n) {
int64_t i = 0;
#if defined(__F16C__)
for (; i + 7 < n; i += 8) {
__m256 x_vec = _mm256_loadu_ps(x + i);
@ -20331,11 +20331,11 @@ size_t ggml_quantize_chunk(
enum ggml_type type,
const float * src,
void * dst,
int start,
int nrows,
int n_per_row,
int64_t start,
int64_t nrows,
int64_t n_per_row,
const float * imatrix) {
const int n = nrows * n_per_row;
const int64_t n = (int64_t) nrows * n_per_row;
if (ggml_quantize_requires_imatrix(type)) {
GGML_ASSERT(imatrix != NULL);

14
ggml.h
View File

@ -332,8 +332,8 @@ extern "C" {
GGML_API float ggml_fp16_to_fp32(ggml_fp16_t x);
GGML_API ggml_fp16_t ggml_fp32_to_fp16(float x);
GGML_API void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, int n);
GGML_API void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, int n);
GGML_API void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, int64_t n);
GGML_API void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, int64_t n);
struct ggml_object;
struct ggml_context;
@ -2210,9 +2210,9 @@ extern "C" {
enum ggml_type type,
const float * src,
void * dst,
int start,
int nrows,
int n_per_row,
int64_t start,
int64_t nrows,
int64_t n_per_row,
const float * imatrix);
//
@ -2377,8 +2377,8 @@ extern "C" {
#else
#define GGML_RESTRICT restrict
#endif
typedef void (*ggml_to_float_t) (const void * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
typedef void (*ggml_to_float_t) (const void * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
typedef void (*ggml_vec_dot_t) (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, size_t bx,
const void * GGML_RESTRICT y, size_t by, int nrc);

2
gguf-py/gguf/constants.py
View File

@ -639,6 +639,8 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.ATTN_K_NORM,
MODEL_TENSOR.ATTN_Q_NORM,
],
# TODO
}

2
gguf-py/gguf/tensor_mapping.py
View File

@ -285,12 +285,14 @@ class TensorNameMap:
MODEL_TENSOR.ATTN_Q_NORM: (
"language_model.encoder.layers.{bid}.self_attention.q_layernorm",
"model.layers.{bid}.self_attn.q_layernorm", # persimmon
"model.layers.{bid}.self_attn.q_norm", # cohere
"transformer.blocks.{bid}.attn.q_ln", # sea-lion
),
MODEL_TENSOR.ATTN_K_NORM: (
"language_model.encoder.layers.{bid}.self_attention.k_layernorm",
"model.layers.{bid}.self_attn.k_layernorm", # persimmon
"model.layers.{bid}.self_attn.k_norm", # cohere
"transformer.blocks.{bid}.attn.k_ln", # sea-lion
),

64
llama.cpp
View File

@ -926,6 +926,8 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
{ LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
{ LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" },
{ LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" },
},
},
{
@ -5406,6 +5408,11 @@ static bool llm_load_tensors(
layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
if (n_layer >= 64){
layer.attn_q_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {hparams.n_embd_head_k, hparams.n_head});
layer.attn_k_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {hparams.n_embd_head_k, hparams.n_head_kv});
}
layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd});
layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa});
layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa});
@ -9454,6 +9461,31 @@ struct llm_build_context {
cb(Vcur, "Vcur", il);
}
if (model.layers[il].attn_q_norm) {
Qcur = ggml_view_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens,
ggml_element_size(Qcur) * n_embd_head,
ggml_element_size(Qcur) * n_embd_head * n_head,
0);
cb(Qcur, "Qcur", il);
Kcur = ggml_view_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens,
ggml_element_size(Kcur) * n_embd_head,
ggml_element_size(Kcur) * n_embd_head * n_head_kv,
0);
cb(Kcur, "Kcur", il);
Qcur = llm_build_norm(ctx0, Qcur, hparams,
model.layers[il].attn_q_norm,
NULL,
LLM_NORM, cb, il);
cb(Qcur, "Qcur", il);
Kcur = llm_build_norm(ctx0, Kcur, hparams,
model.layers[il].attn_k_norm,
NULL,
LLM_NORM, cb, il);
cb(Kcur, "Kcur", il);
}
Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
@ -13323,9 +13355,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
return new_type;
}
static size_t llama_tensor_quantize_internal(enum ggml_type new_type, const float * f32_data, void * new_data, const int chunk_size, int nrows, int n_per_row, const float * imatrix, std::vector<std::thread> & workers, const int nthread) {
static size_t llama_tensor_quantize_internal(enum ggml_type new_type, const float * f32_data, void * new_data, const int64_t chunk_size, int64_t nrows, int64_t n_per_row, const float * imatrix, std::vector<std::thread> & workers, const int nthread) {
std::mutex mutex;
int counter = 0;
int64_t counter = 0;
size_t new_size = 0;
if (nthread < 2) {
// single-thread
@ -13333,11 +13365,11 @@ static size_t llama_tensor_quantize_internal(enum ggml_type new_type, const floa
}
auto compute = [&mutex, &counter, &new_size, new_type, f32_data, new_data, chunk_size,
nrows, n_per_row, imatrix]() {
const int nrows_per_chunk = chunk_size / n_per_row;
const int64_t nrows_per_chunk = chunk_size / n_per_row;
size_t local_size = 0;
while (true) {
std::unique_lock<std::mutex> lock(mutex);
int first_row = counter; counter += nrows_per_chunk;
int64_t first_row = counter; counter += nrows_per_chunk;
if (first_row >= nrows) {
if (local_size > 0) {
new_size += local_size;
@ -13345,7 +13377,7 @@ static size_t llama_tensor_quantize_internal(enum ggml_type new_type, const floa
break;
}
lock.unlock();
const int this_nrow = std::min(nrows - first_row, nrows_per_chunk);
const int64_t this_nrow = std::min(nrows - first_row, nrows_per_chunk);
local_size += ggml_quantize_chunk(new_type, f32_data, new_data, first_row * n_per_row, this_nrow, n_per_row, imatrix);
}
};
@ -13539,6 +13571,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
// quantize only 2D and 3D tensors (experts)
quantize &= (ggml_n_dims(tensor) >= 2);
// do not quantize norm tensors
quantize &= name.find("_norm.weight") == std::string::npos;
quantize &= params->quantize_output_tensor || name != "output.weight";
quantize &= !params->only_copy;
@ -13585,7 +13621,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
new_size = ggml_nbytes(tensor);
LLAMA_LOG_INFO("size = %8.3f MB\n", ggml_nbytes(tensor)/1024.0/1024.0);
} else {
const size_t nelements = ggml_nelements(tensor);
const int64_t nelements = ggml_nelements(tensor);
const float * imatrix = nullptr;
if (imatrix_data) {
@ -13637,20 +13673,20 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
LLAMA_LOG_INFO("converting to %s .. ", ggml_type_name(new_type));
fflush(stdout);
if (work.size() < nelements * 4) {
if (work.size() < (size_t)nelements * 4) {
work.resize(nelements * 4); // upper bound on size
}
new_data = work.data();
const int n_per_row = tensor->ne[0];
const int nrows = tensor->ne[1];
const int64_t n_per_row = tensor->ne[0];
const int64_t nrows = tensor->ne[1];
static const int min_chunk_size = 32 * 512;
const int chunk_size = n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row);
static const int64_t min_chunk_size = 32 * 512;
const int64_t chunk_size = n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row);
const int nelements_matrix = tensor->ne[0] * tensor->ne[1];
const int nchunk = (nelements_matrix + chunk_size - 1)/chunk_size;
const int nthread_use = nthread > 1 ? std::max(1, std::min(nthread, nchunk)) : 1;
const int64_t nelements_matrix = tensor->ne[0] * tensor->ne[1];
const int64_t nchunk = (nelements_matrix + chunk_size - 1)/chunk_size;
const int64_t nthread_use = nthread > 1 ? std::max((int64_t)1, std::min((int64_t)nthread, nchunk)) : 1;
// quantize each expert separately since they have different importance matrices
new_size = 0;