root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-25 02:44:36 +00:00
Code Issues Actions 9 Packages Projects Releases Wiki Activity

cuda : restore lost changes

Browse Source
This commit is contained in:
Georgi Gerganov 2023-12-07 14:17:39 +02:00
parent 213a4e2c32
commit c6b3d195b8
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
1 changed files with 3 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
ggml-cuda.cu
View File

@ -1,6 +1,7 @@
#include <algorithm>
#include <cstddef>
#include <cstdint>
#include <cinttypes>
#include <float.h>
#include <limits>
#include <stdint.h>
@ -8016,10 +8017,11 @@ static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1
#ifdef GGML_CUDA_FORCE_DMMV
const bool use_mul_mat_vec_q = false;
#else
const bool use_mul_mat_vec_q = min_compute_capability >= MIN_CC_DP4A && ggml_is_quantized(src0->type);
const bool use_mul_mat_vec_q = min_compute_capability >= MIN_CC_DP4A && ggml_is_quantized(src0->type) && ggml_nrows(src1) == 1;
#endif // GGML_CUDA_FORCE_DMMV
if (use_mul_mat_vec_q) {
// NOTE: this kernel does not support ggml_nrows(src1) > 1
ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_vec_q, true);
} else {
ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, false);