happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

add a case havign memory access violation

This commit is contained in:
bssrdf 2025-11-10 15:33:32 -05:00
parent 89103a856c
commit 5e1352cb60
2 changed files with 8 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
ggml/src/ggml-cuda/conv3d-implicit.cu
View File

@ -1082,7 +1082,6 @@ static void conv3d_implicit_cuda_f16(ggml_backend_cuda_context & ctx, const floa
cudaFuncAttributeMaxDynamicSharedMemorySize, 65536); // set shared memory limit to 64KB which is maximum for sm_75
dim3 gridDim(BlocksN, BlocksM);
dim3 blockDim(ThreadsN, ThreadsM);
conv3d_implicit_kernel<float, BM_dim, BN_dim, BK_dim,
WM_dim, WN_dim, WK_dim, 0, NumThreads>
<<<gridDim, blockDim, shmem_bytes, st>>>(X_H, K_H, Y_D, P);

15
tests/test-conv3d.cpp
View File

@ -350,7 +350,7 @@ int main(void)
{
ggml_time_init();
std::vector<std::tuple<int, int, int, int, int, int, int, int>> configs = {
std::make_tuple(1,2,16,32,4,3,3,3),
// std::make_tuple(1,2,16,32,4,3,3,3),
// std::make_tuple(320,1280,26,38,8,3,3,3),
// std::make_tuple(1280,1280,26,38,8,3,3,3),
// std::make_tuple(320,1280,52,76,8,3,3,3),
@ -358,12 +358,13 @@ int main(void)
// std::make_tuple(320,1280,104,152,8,3,3,3),
// std::make_tuple(1280,1280,104,152,8,3,3,3),
// std::make_tuple(320,1280,208,304,4,3,3,3),
std::make_tuple(1024,2048,30,52,3,3,3,3),
std::make_tuple(1024,2048,52,76,4,3,3,3),
std::make_tuple(1024,2048,52,76,6,3,3,3),
std::make_tuple(48,3072,64,64,9,2,2,1),
std::make_tuple(48,3072,64,64,17,2,2,1),
std::make_tuple(48,3072,64,64,33,2,2,1),
// std::make_tuple(1024,2048,30,52,3,3,3,3),
// std::make_tuple(1024,2048,52,76,4,3,3,3),
// std::make_tuple(1024,2048,52,76,6,3,3,3),
// std::make_tuple(48,3072,64,64,9,2,2,1),
// std::make_tuple(48,3072,64,64,17,2,2,1),
// std::make_tuple(48,3072,64,64,33,2,2,1),
std::make_tuple(320,320,104,158,8,3,3,3),
};
int k = 0;