vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

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
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
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);