root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-12 19:50:17 +00:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

Disable CUDA graphs for old GPU arch and with env var

Browse Source
This commit is contained in:
Alan Gray 2024-04-23 06:27:08 -07:00
parent c2691d968a
commit df4719ec7e
1 changed files with 18 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
ggml-cuda.cu
View File

@ -2419,9 +2419,12 @@ struct ggml_cudaGraph {
int softmax_ne0 = 0; int softmax_ne0 = 0;
cudaGraphNode_t nodes[MAX_NODES_IN_CUDA_GRAPH]; cudaGraphNode_t nodes[MAX_NODES_IN_CUDA_GRAPH];
cudaKernelNodeParams params[MAX_NODES_IN_CUDA_GRAPH]; cudaKernelNodeParams params[MAX_NODES_IN_CUDA_GRAPH];
bool disableDueToGpuArch=false;
}; };
#endif #endif
const bool disableCudaGraphs = (getenv("LLAMACPP_DISABLE_CUDA_GRAPHS") != nullptr);
GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) { GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context; ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
@ -2437,8 +2440,21 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
// kernel parameters which need updated in the graph for each token // kernel parameters which need updated in the graph for each token
void* ggmlCudaCpyFn = nullptr; void* ggmlCudaCpyFn = nullptr;
if(ggml_backend_cuda_get_device_count() > 1){
useCudaGraph = false; // disable CUDA graphs for multi-gpu for now. TO DO investigate if(cudaGraph.count==0){
cudaDeviceProp prop;
int device;
cudaGetDevice(&device);
cudaGetDeviceProperties(&prop, device);
if (prop.major < 8){
cudaGraph.disableDueToGpuArch=true;
}
}
// Disable CUDA graphs in presence of env var or old GPU.
// Also disable for multi-gpu for now. TO DO investigate
if(disableCudaGraphs || cudaGraph.disableDueToGpuArch || ggml_backend_cuda_get_device_count() > 1){
useCudaGraph = false;
} }
if(useCudaGraph) { if(useCudaGraph) {