Skip to content

Fix: audio encoder uses full attention instead of windowed on non-FA2 backends #103

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

Open
Dvad wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Fix: audio encoder uses full attention instead of windowed on non-FA2 backends #103

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
4 changes: 4 additions & 0 deletions qwen_asr/core/transformers_backend/modeling_qwen3_asr.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -724,11 +724,15 @@ def forward(
if remainder != 0:
cu_chunk_lens += [remainder]
cu_seqlens = torch.tensor(cu_chunk_lens, device=aftercnn_lens.device).cumsum(-1, dtype=torch.int32)
# Build block-diagonal attention mask for non-FA2 backends (SDPA, eager).
# FA2 uses cu_seqlens natively; _prepare_attention_mask returns None for FA2.
attention_mask = self._prepare_attention_mask(hidden_states, cu_seqlens)

for encoder_layer in self.layers:
layer_outputs = encoder_layer(
hidden_states,
cu_seqlens,
attention_mask=attention_mask,
)

hidden_states = layer_outputs[0]
Expand Down