Skip to content

ggml-cuda: check for srcs outside the cgraph #18583

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
am17an merged 2 commits into from
Jan 5, 2026
Merged

ggml-cuda: check for srcs outside the cgraph #18583

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
71ed0f9
ggml-cuda: check for srcs outside the cgraph
am17an Jan 4, 2026
261b186
review: use leafs instead
am17an Jan 5, 2026
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
1 change: 1 addition & 0 deletions ggml/src/ggml-cuda/common.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1065,6 +1065,7 @@ struct ggml_cuda_graph {
int number_consecutive_updates = 0;
bool cuda_graphs_enabled = false;
std::vector<ggml_graph_node_properties> ggml_graph_properties;
std::vector<ggml_graph_node_properties> extraneous_srcs_properties;
Copy link
Member

@ggerganov ggerganov Jan 5, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is no longer used

#endif
};

Expand Down
16 changes: 14 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 @@ -2973,15 +2973,16 @@ static bool is_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx,
}

// Check if the graph size has changed
if (cuda_ctx->cuda_graph->ggml_graph_properties.size() != (size_t)cgraph->n_nodes) {
if (cuda_ctx->cuda_graph->ggml_graph_properties.size() != (size_t)cgraph->n_nodes + cgraph->n_leafs) {
cuda_graph_update_required = true;
cuda_ctx->cuda_graph->ggml_graph_properties.resize(cgraph->n_nodes);
cuda_ctx->cuda_graph->ggml_graph_properties.resize(cgraph->n_nodes + cgraph->n_leafs);
}

// Loop over nodes in GGML graph to determine if CUDA graph update is required
// and store properties to allow this comparison for the next token
for (int i = 0; i < cgraph->n_nodes; i++) {
bool has_matching_properties = true;

if (!cuda_graph_update_required) {
has_matching_properties = ggml_graph_node_has_matching_properties(cgraph->nodes[i], &cuda_ctx->cuda_graph->ggml_graph_properties[i]);
}
Expand All @@ -2991,6 +2992,17 @@ static bool is_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx,
set_ggml_graph_node_properties(cgraph->nodes[i], &cuda_ctx->cuda_graph->ggml_graph_properties[i]);
Copy link
Collaborator

@ORippler ORippler Jan 5, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can't we move this setting inside the if condition as well? (see comment below)

}

for (int i = 0; i < cgraph->n_leafs; i++) {
bool has_matching_properties = true;
if (!cuda_graph_update_required) {
Copy link
Contributor

@JohannesGaessler JohannesGaessler Jan 4, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What is the purpose of this check?

Copy link
Contributor Author

@am17an am17an Jan 5, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

it skips calling ggml_graph_node_has_matching_properties? Or did you mean something else

Copy link
Contributor

@JohannesGaessler JohannesGaessler Jan 5, 2026 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I meant to ask whether it is relevant for program correctness but based on your answer I guess not.

has_matching_properties = ggml_graph_node_has_matching_properties(cgraph->leafs[i], &cuda_ctx->cuda_graph->ggml_graph_properties[cgraph->n_nodes + i]);
}
if (!has_matching_properties) {
cuda_graph_update_required = true;
}
set_ggml_graph_node_properties(cgraph->leafs[i], &cuda_ctx->cuda_graph->ggml_graph_properties[cgraph->n_nodes + i]);
Comment on lines +3001 to +3003
Copy link
Collaborator

@ORippler ORippler Jan 5, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
cuda_graph_update_required = true;
}
set_ggml_graph_node_properties(cgraph->leafs[i], &cuda_ctx->cuda_graph->ggml_graph_properties[cgraph->n_nodes + i]);
cuda_graph_update_required = true;
set_ggml_graph_node_properties(cgraph->leafs[i], &cuda_ctx->cuda_graph->ggml_graph_properties[cgraph->n_nodes + i]);
}

Shouldn't we have to update the node properties only if they are not matching? Or do we set something with set_ggml_graph_node_properties that we don't check for in ggml_graph_node_has_matching_properties

Copy link
Contributor Author

@am17an am17an Jan 5, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm not sure, just following the existing pattern here. Will address in the refactor

}

return cuda_graph_update_required;
}

Expand Down