Skip to content

[docker] fix qwen3_vl visual module loading#1727

Open
ZHZisZZ wants to merge 1 commit intoTHUDM:mainfrom
ZHZisZZ:patch/qwen3vl
Open

[docker] fix qwen3_vl visual module loading#1727
ZHZisZZ wants to merge 1 commit intoTHUDM:mainfrom
ZHZisZZ:patch/qwen3vl

Conversation

@ZHZisZZ
Copy link
Copy Markdown

@ZHZisZZ ZHZisZZ commented Mar 15, 2026

Backport sgl-project/sglang#19333 to fix Qwen3-VL visual weight loading in sglang v0.5.9. The missing name mapping (model.visual. -> visual.) caused visual component weights to fail to load.

Backport sgl-project/sglang#19333 to fix Qwen3-VL visual weight loading
in sglang v0.5.9. The missing name mapping (model.visual. -> visual.)
caused visual component weights to fail to load.
yueming-yuan added a commit to sgl-project/sglang that referenced this pull request Mar 16, 2026
Add missing name mapping (model.visual. -> visual.) in
Qwen3VLForConditionalGeneration.load_weights to fix visual component
weights failing to load.

Cherry-picked from THUDM/slime#1727.
@ngothetai
Copy link
Copy Markdown

I'm also looking forward to this new image!

@ZHZisZZ
Copy link
Copy Markdown
Author

ZHZisZZ commented Mar 30, 2026

@zhuzilin
Hi, I double-checked and confirmed this is the reason multiple people have recently reported that Qwen3-VL training does not work in Slime. I’d appreciate it if you could take a look. Thanks!

@GentleZhu
Copy link
Copy Markdown

+1 on this issue

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants