Skip to content

[main][bugfix] Change seq_lens in dummy attn_metadata to max_query_len #4097

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 2 commits into from
Nov 12, 2025
Merged

[main][bugfix] Change seq_lens in dummy attn_metadata to max_query_len #4097

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
a30adc6
bugfix
Angazenn Nov 10, 2025
bf833ca
change to max_query_len
Angazenn Nov 11, 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
7 changes: 6 additions & 1 deletion vllm_ascend/worker/model_runner_v1.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2665,7 +2665,12 @@ def _build_dummy_attn_metadata(

attn_metadata = {}

seq_lens = self.model_config.max_model_len
# When force_attention == True, the model runs in capturing so we
# need seq_lens as max_model_len to get max workspace for attention op.
# However, when force_attention == False, the model might be running
# normal inference. If dp_size > 1, we only need dummy_run
# to execute a short attention with seq_lens as 1.
seq_lens = self.model_config.max_model_len if force_attention else 1
Copy link
Collaborator

@whx-sjtu whx-sjtu Nov 11, 2025

Choose a reason for hiding this comment

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

When force_attention == False, the dummy attn metadata is None meaning skipping attention part, so I don't quite understand why we need this.

Copy link
Collaborator Author

@Angazenn Angazenn Nov 11, 2025

Choose a reason for hiding this comment

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

The condition is force_attention or aclgraph_runtime_mode == CUDAGraphMode.FULL. When capturing, force_attention = (aclgraph_runtime_mode == CUDAGraphMode.FULL). However, when not capturing, force_attention in _dummy_run is always False. So _dummy_run in actual inference executes dummy attention when using full graph mode.

self.seq_lens_np[:num_reqs] = seq_lens
self.seq_lens_np[num_reqs:] = 0

Expand Down
Loading