Skip to content

[Spec] Cleanup idle stub and shape-check patterns #24881

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
hnyls2002 merged 3 commits into from
May 10, 2026
+16 −16
Merged

[Spec] Cleanup idle stub and shape-check patterns #24881

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
c429f5d
misc cleanup
hnyls2002 May 10, 2026
f6460f0
unify input_ids.numel -> input_ids.shape[0]
hnyls2002 May 10, 2026
479e4d6
doc EagleDraftInput batch-uniform fields
hnyls2002 May 10, 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
10 changes: 4 additions & 6 deletions python/sglang/srt/speculative/eagle_worker.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -515,12 +515,10 @@ def forward_batch_generation(self, batch: ScheduleBatch) -> GenerationBatchResul
batch.spec_info = next_draft_input
else:
# All reqs finished and dp_attention isn't forcing extend.
# Stash an empty EagleDraftInput so next iter's merge_batch
# short-circuits on None hidden_states (EagleVerifyInput
# has no merge_batch).
batch.spec_info = EagleDraftInput(
capture_hidden_mode=CaptureHiddenMode.LAST,
)
# Install an idle EagleDraftInput so next iter's scheduler
# ops (merge_batch / filter_batch) see well-typed empty
# tensors instead of None.
self._draft_preprocess_idle(batch)

set_time_batch(
batch.reqs, "set_spec_draft_extend_end_time", trace_only=True
Expand Down
10 changes: 4 additions & 6 deletions python/sglang/srt/speculative/multi_layer_eagle_worker.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -293,12 +293,10 @@ def forward_batch_generation(self, batch: ScheduleBatch) -> GenerationBatchResul
batch.spec_info = next_draft_input
else:
# All reqs finished and dp_attention isn't forcing extend.
# Stash an empty EagleDraftInput so next iter's merge_batch
# short-circuits on None hidden_states (EagleVerifyInput
# has no merge_batch).
batch.spec_info = EagleDraftInput(
capture_hidden_mode=CaptureHiddenMode.LAST,
)
# Install an idle EagleDraftInput so next iter's scheduler
# ops (merge_batch / filter_batch) see well-typed empty
# tensors instead of None.
self._draft_preprocess_idle(batch)

@gemini-code-assist gemini-code-assist Bot May 10, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

medium

The _draft_preprocess_idle method in MultiLayerEagleWorker (defined at line 363) delegates to EAGLEWorker._draft_preprocess_idle, which uses self.model_config.spec_hidden_size. However, MultiLayerEagleWorker appears to use self.model_config.hidden_size for its hidden states (as seen in line 683). This could lead to an AttributeError if spec_hidden_size is missing from the MTP model config, or a shape mismatch crash during merge_batch in the next iteration if the sizes differ. It would be safer to implement a local version of _draft_preprocess_idle that uses the correct hidden size logic for MTP models.


return GenerationBatchResult(
logits_output=logits_output,
Expand Down
Loading