vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Avoid using saved CUDA graph if scale changes and reset nodes/params on update

Browse source 
Fixes https://github.com/ggerganov/llama.cpp/issues/9451
This commit is contained in:
Alan Gray 2024-09-17 08:47:06 -07:00
parent 0d2ec43833
commit c6b3ea6595
2 changed files with 10 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

9
ggml/src/ggml-cuda.cu
View file

@ -2451,6 +2451,7 @@ static void set_ggml_graph_node_properties(ggml_tensor * node, ggml_graph_node_p
for (int i = 0; i < GGML_MAX_SRC; i++) {
graph_node_properties->src_address[i] = node->src[i] ? node->src[i]->data : nullptr;
}
memcpy(graph_node_properties->op_params, node->op_params, GGML_MAX_OP_PARAMS);
}
static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_graph_node_properties * graph_node_properties) {
@ -2482,6 +2483,12 @@ static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_gra
return false;
}
}
if (node->op == GGML_OP_SCALE &&
memcmp(graph_node_properties->op_params, node->op_params, GGML_MAX_OP_PARAMS) != 0) {
return false;
}
return true;
}
@ -2694,7 +2701,9 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
CUDA_CHECK(cudaGraphGetNodes(cuda_ctx->cuda_graph->graph, nullptr, &cuda_ctx->cuda_graph->num_nodes));
// Subsequent call with non-null argument gets nodes
cuda_ctx->cuda_graph->nodes.resize(cuda_ctx->cuda_graph->num_nodes);
cuda_ctx->cuda_graph->nodes.clear();
cuda_ctx->cuda_graph->params.resize(cuda_ctx->cuda_graph->num_nodes);
cuda_ctx->cuda_graph->params.clear();
if (cuda_ctx->cuda_graph->num_nodes > 0) {
CUDA_CHECK(cudaGraphGetNodes(cuda_ctx->cuda_graph->graph, cuda_ctx->cuda_graph->nodes.data(), &cuda_ctx->cuda_graph->num_nodes));

1
ggml/src/ggml-cuda/common.cuh
View file

@ -569,6 +569,7 @@ struct ggml_graph_node_properties {
int64_t ne[GGML_MAX_DIMS];
size_t nb[GGML_MAX_DIMS];
void * src_address[GGML_MAX_SRC];
int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
};
struct ggml_cuda_graph {