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

ggml-cuda: reorder only relevant nodes (#17639)

Browse Source
This commit is contained in:
Aman Gupta
2025-12-02 12:36:31 +08:00
committed by GitHub
parent 7b6d745364
commit ed32089927
2 changed files with 62 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/common.cuh
+5 -2
View File
@@ -989,6 +989,10 @@ struct ggml_cuda_concurrent_event {
int n_streams = 0;
std::unordered_map<const ggml_tensor *, int> stream_mapping;
// Original order of nodes in this concurrent region (before interleaving)
// Used to restore grouping for fusion within streams
std::vector<const ggml_tensor *> original_order;
const ggml_tensor * join_node;
ggml_cuda_concurrent_event() = default;
@@ -1011,6 +1015,7 @@ struct ggml_cuda_concurrent_event {
, fork_event(other.fork_event)
, n_streams(other.n_streams)
, stream_mapping(std::move(other.stream_mapping))
, original_order(std::move(other.original_order))
, join_node(other.join_node) {
other.fork_event = nullptr;
}
@@ -1121,11 +1126,9 @@ struct ggml_cuda_concurrent_event {
};
struct ggml_cuda_stream_context {
std::vector<const ggml_tensor *> original_nodes;
std::unordered_map<const ggml_tensor *, ggml_cuda_concurrent_event> concurrent_events;
void reset() {
original_nodes.clear();
concurrent_events.clear();
}
};