Recommendations for running GRPO on Llama Scout

[mizadri]

Hello everyone!

I am a little bit confused about which backend to use to run GRPO on LLama 4 Scout, I have been using FSDP with GPT-oss with VLLM, but I have seen the megatron backend has good support for MoEs. Given that I need to configure my docker container on the specific VLLM/SGLang frameworks, can someone give me a few pointers of which is the recommended setup to get the best performance for LLama 4 Scout?

Thanks a lot for the great framework and the help!