Skip to content

[Bugfix][Model] Qwen3-Omni: move cu_seqlens to GPU before VIT attention #1063

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
mgehre-amd merged 1 commit into from
Jul 27, 2026
Merged

[Bugfix][Model] Qwen3-Omni: move cu_seqlens to GPU before VIT attention #1063

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
3 changes: 3 additions & 0 deletions vllm/model_executor/models/qwen3_omni_moe_thinker.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -991,6 +991,9 @@ def forward(
)
cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)

# Move cu_seqlens to GPU; grid_thw may be on CPU during profile_run
# and FA3 vit attention requires cu_seqlens on CUDA.
cu_seqlens = cu_seqlens.to(self.device, non_blocking=True)
hidden_states = hidden_states.unsqueeze(1)
rotary_pos_emb_cos = rotary_pos_emb_cos.to(hidden_states.device)
rotary_pos_emb_sin = rotary_pos_emb_sin.to(hidden_states.device)
Expand Down
Loading