Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

cuda : Enable CUDA Graph usage for Nemotron Nano v2 (NemotronH) (#16328)

Browse Source 
* Fix Nemotron Nano v2 9B not executing as CUDA Graph on NVIDIA GPUs

* fix to ensure test-backend-ops check passes
This commit is contained in:
anavp-nvidia
2025-09-30 11:13:22 +03:00
committed by GitHub
parent 35fb82497e
commit a014310374
3 changed files with 20 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/cpy.cu
+12 -2
View File
@@ -329,7 +329,11 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
} else
#endif // GGML_USE_MUSA && GGML_MUSA_MUDNN_COPY
{
CUDA_CHECK(cudaMemcpyAsync(src1_ddc, src0_ddc, ggml_nbytes(src0), cudaMemcpyDeviceToDevice, main_stream));
if (src0->type == GGML_TYPE_F32) {
ggml_cpy_flt_cuda<float, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
} else {
CUDA_CHECK(cudaMemcpyAsync(src1_ddc, src0_ddc, ggml_nbytes(src0), cudaMemcpyDeviceToDevice, main_stream));
}
}
} else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
ggml_cpy_flt_cuda<float, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
@@ -400,7 +404,13 @@ void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
if (src0->type == src1->type && ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
return nullptr;
// Prioritize CUDA graph compatibility over direct memory copy optimization.
// Using copy kernels here maintains graph indirection support, preventing performance regression from disabled CUDA graphs.
if (src0->type == GGML_TYPE_F32) {
return (void*) cpy_flt<cpy_1_flt<float, float>>;
} else {
return nullptr;
}
} else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
return (void*) cpy_flt<cpy_1_flt<float, float>>;
} else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {