Skip to content

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

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Oct 29, 2025
Merged

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

Changes from 2 commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
3449842
CUDA: Fix bug in topk-moe for gpt-oss
am17an Oct 28, 2025
91cd070
fix for qwen3 too
am17an Oct 28, 2025
5ea78fb
change ifndef to ifdef
am17an Oct 29, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 13 additions & 2 deletions ggml/src/ggml-cuda/ggml-cuda.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2974,7 +2974,7 @@ 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];

Expand All @@ -2993,7 +2993,7 @@ 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];

Expand Down Expand Up @@ -3114,9 +3114,20 @@ 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];


#ifndef 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;
}
Expand Down
Loading