tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-08-17 21:51:27 -04:00
Code Issues Packages Projects Releases Wiki Activity

CUDA: revise q8_1 data layout for mul_mat_q (#7824)

Browse Source
This commit is contained in:
Johannes Gäßler
2024-06-09 09:42:25 +02:00
committed by GitHub
parent 2decf57bc6
commit 42b53d192f
5 changed files with 282 additions and 151 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
ggml-cuda/quantize.cuh
View File

@@ -1,5 +1,20 @@
#pragma once
#include "common.cuh"
#include "mmq.cuh"
#include <cstdint>
#define CUDA_QUANTIZE_BLOCK_SIZE 256
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);
typedef void (*quantize_cuda_t)(
const float * x, void * vy, const int64_t kx0, const int64_t kx1, const int64_t channels, const int64_t kx0_padded,
const ggml_type type_x, cudaStream_t stream);
void quantize_row_q8_1_cuda(
const float * x, void * vy, const int64_t kx0, const int64_t kx1, const int64_t channels, const int64_t kx0_padded,
const ggml_type type_x, cudaStream_t stream);
void quantize_mmq_q8_1_cuda(
const float * x, void * vy, const int64_t kx0, const int64_t kx1, const int64_t channels, const int64_t kx0_padded,
const ggml_type type_x, cudaStream_t stream);