Skip to content

fix qwen3_next #58

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
Jintao-Huang merged 1 commit into from
May 4, 2026
Merged

fix qwen3_next #58

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
2 changes: 1 addition & 1 deletion src/mcore_bridge/model/gpts/qwen3_next.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -497,7 +497,7 @@ def forward(self, hidden_states: torch.Tensor, **kwargs):
hidden_states = new_hidden_states
else:
hidden_states = hidden_states.transpose(0, 1)
attention_mask = resolve_hf_attention_mask(kwargs)
attention_mask = resolve_gdn_attention_mask(kwargs)
Copy link
Copy Markdown

@gemini-code-assist gemini-code-assist Bot May 4, 2026

Choose a reason for hiding this comment

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

high

The function resolve_gdn_attention_mask (called here and defined at line 62) has a potential logic error and runtime risk:

  1. Type Safety: At line 70, it uses the bitwise NOT operator ~ on attention_mask. If the mask is an integer tensor (common for HuggingFace masks), ~1 results in -2, which will cause the sum(...) > 0 check to behave unexpectedly. It should be explicitly cast to boolean: (~attention_mask.to(torch.bool)).
  2. Dimensionality: The function assumes a 4D input by using dim=(1, 2). If kwargs.get('attention_mask') returns a 2D tensor (standard HuggingFace format), this will raise an IndexError.

Since this function is now being used to resolve the attention mask, these issues should be addressed in its implementation to ensure robustness.

res = super().forward(hidden_states=hidden_states, attention_mask=attention_mask)
if thd_format:
res = res[attention_mask][:, None]
Expand Down
Loading