[Question] Gemma 4 support via HF wrapping approach?

Your Question

I'm interested in using Gemma 4 (dense 31B and 26B A4B MoE) with slime for RL training. I noticed SGLang already supports Gemma 4 (sgl-project/sglang#21952), so the rollout side should be covered.

For the training side, I'm considering using the black-box HF wrapping approach described in docs/en/advanced/arch-support-beyond-megatron.md, following the Qwen3Next plugin as a template.

A few questions:

1. Are there any existing plans to support Gemma 4?
2. Is the HF wrapping approach the right path here, or would you recommend a different direction?

What I've Tried

• I have read the example Qwen3Next support

Environment (if relevant)

• slime version:
• Python version:
• PyTorch version:
• CUDA/ROCm version:
• GPU type and count:
• OS:

Additional Context

No response

Pre-submission Checklist

• I have read the CONTRIBUTING.md and understand the collaboration scope.
• I have read the documentation and FAQ and my question is not answered there.
• I have searched for existing issues and my question has not been asked before.

Start by reading docs/en/advanced/arch-support-beyond-megatron.md and the existing Qwen3Next support example. Check how the proposed Hugging Face wrapping approach would apply to Gemma 4 and confirm the rollout-side assumptions against the referenced SGLang support. Done requires a maintainer decision on the recommended support path and scope.