ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-30 19:31:48 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix PPL increase caused by mmq_id

Browse Source
This commit is contained in:
Iwan Kawrakow
2025-11-07 13:56:24 +02:00
parent 9d0b834405
commit e49cfff302
1 changed files with 4 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
ggml/src/ggml-cuda/mmq_id_common.cuh
View File

@@ -3960,7 +3960,10 @@ template <ggml_type type, int mmq_x>
static void launch_mul_mat_q_id(ggml_backend_cuda_context & ctx, const mmq_args_id & args, cudaStream_t stream) { static void launch_mul_mat_q_id(ggml_backend_cuda_context & ctx, const mmq_args_id & args, cudaStream_t stream) {
const int id = ggml_cuda_get_device(); const int id = ggml_cuda_get_device();
const int cc = ggml_cuda_info().devices[id].cc; const int cc = ggml_cuda_info().devices[id].cc;
const int nsm = ggml_cuda_info().devices[id].nsm; const int nsm_max = ggml_cuda_info().devices[id].nsm;
int nsm = 1;
//while (nsm*2 <= nsm_max) nsm *= 2;
while (nsm < nsm_max) nsm *= 2;
const int warp_size = ggml_cuda_get_physical_warp_size_host(); //ggml_cuda_info().devices[id].warp_size; const int warp_size = ggml_cuda_get_physical_warp_size_host(); //ggml_cuda_info().devices[id].warp_size;
const int nwarps = mmq_get_nwarps_host(cc, warp_size); const int nwarps = mmq_get_nwarps_host(cc, warp_size);
const int mmq_y = get_mmq_y_host(cc); const int mmq_y = get_mmq_y_host(cc);