Skip to content

[sgl] disable piecewise cuda graph when a model doesn't have layers #21565

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
ispobock merged 1 commit into from
Mar 29, 2026
Merged

[sgl] disable piecewise cuda graph when a model doesn't have layers #21565

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
8 changes: 8 additions & 0 deletions python/sglang/srt/model_executor/model_runner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2379,6 +2379,14 @@ def init_piecewise_cuda_graphs(self):
# Collect attention layers and moe layers from the model
self.model.model = resolve_language_model(self.model)
language_model = getattr(self.model, "language_model", self.model)

Copy link
Copy Markdown
Collaborator

@Oasis-Git Oasis-Git Mar 28, 2026

Choose a reason for hiding this comment

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

remove this line

# Some draft models (e.g. eagle3) don't have a standard 'layers' attribute
Copy link
Copy Markdown
Collaborator

@Oasis-Git Oasis-Git Mar 28, 2026

Choose a reason for hiding this comment

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

Could you move the change to original 2371 line?

Copy link
Copy Markdown
Contributor Author

@bixue2010 bixue2010 Mar 28, 2026

Choose a reason for hiding this comment

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

you mean Gemini's suggestion is wrong?

Copy link
Copy Markdown
Collaborator

@Oasis-Git Oasis-Git Mar 28, 2026

Choose a reason for hiding this comment

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

yes. the robustness problem will be fixed later

Copy link
Copy Markdown
Contributor Author

@bixue2010 bixue2010 Mar 28, 2026

Choose a reason for hiding this comment

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

@Oasis-Git , line 2371 doesn't have language_model yet though, and we need to use that to test language_model.model. we could move the resolve language_model before enable piecewise_cuda_graph check though, but want to confirm if this matches your motivation?

Copy link
Copy Markdown
Collaborator

@Oasis-Git Oasis-Git Mar 28, 2026

Choose a reason for hiding this comment

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

oh I see. sorry I made a mistake. what I hope to keep is that we should gather disable conditions for readability. Keep your first version. The disable conditions should be above the initialization of those layers.

Copy link
Copy Markdown
Contributor Author

@bixue2010 bixue2010 Mar 28, 2026

Choose a reason for hiding this comment

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

sounds good. updated. could you please take a look again?

if not hasattr(language_model.model, "layers"):
logger.warning(
"Disable piecewise CUDA graph because the model does not have a 'layers' attribute"
)
return

self.attention_layers = []
self.moe_layers = []
self.moe_fusions = []
Expand Down
Loading