Checklist

• 1. If the issue you raised is not a feature but a question, please raise a discussion at https://github.com/sgl-project/sglang/discussions/new/choose Otherwise, it will be closed.
• 2. Please use English, otherwise it will be closed.

Motivation

Currently the implementation of vision model in mllama4 is imported from transformers, which may have bad performance. We could implement the modules using sglang's implementation of vision attention

Related resources

No response