CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-11 10:36:54 +00:00
Code Issues Packages Projects Releases Wiki Activity

CUDA: Fix bug in topk-moe for gpt-oss (#16821)

Browse Source 
* CUDA: Fix bug in topk-moe for gpt-oss

When using ggml_can_fuse_subgraph, the output nodes which are passed are wrong. This causes `test-backend-ops` to still fuse ndoes (because the nodes are not used elsewhere in the graph),
but it actually doesn't fuse in the actual gpt-oss

* fix for qwen3 too

* change ifndef to ifdef
This commit is contained in:
Aman Gupta
2025-10-29 15:55:06 +08:00
committed by GitHub
parent 338074c383
commit 9a3ea685b9
1 changed files with 13 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
ggml/src/ggml-cuda/ggml-cuda.cu
View File

@@ -2978,7 +2978,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
ggml_cuda_topk_moe_ops(/*with_norm=*/false, /*delayed_softmax=*/true);
if (ops.size() == topk_moe_ops_with_norm.size() &&
ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 3, node_idx + 8 })) {
ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 3, node_idx + 9 })) {
ggml_tensor * softmax = cgraph->nodes[node_idx];
ggml_tensor * weights = cgraph->nodes[node_idx + 9];
@@ -2997,7 +2997,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
}
if (ops.size() == topk_moe_ops_delayed_softmax.size() &&
ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 2, node_idx + 5 })) {
ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 1, node_idx + 5 })) {
ggml_tensor * softmax = cgraph->nodes[node_idx + 4];
ggml_tensor * weights = cgraph->nodes[node_idx + 5];
@@ -3118,9 +3118,20 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
// With the use of CUDA graphs, the execution will be performed by the graph launch.
if (!use_cuda_graph || cuda_graph_update_required) {
[[maybe_unused]] int prev_i = 0;
for (int i = 0; i < cgraph->n_nodes; i++) {
ggml_tensor * node = cgraph->nodes[i];
#ifdef GGML_CUDA_DEBUG
const int nodes_fused = i - prev_i - 1;
prev_i = i;
if (nodes_fused > 0) {
GGML_LOG_INFO("nodes_fused: %d\n", nodes_fused);
}
#endif
if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
continue;
}