[RFC] Integrate VeOmni as the training engine in VERL

[A1waysBeenHere]

Feature request

This RFC proposes integrating VeOmni as a training engine backend for VERL.
The goal is to leverage VeOmni’s high-performance distributed training framework to enhance VERL’s scalability and efficiency in large-scale RLHF and post-training workflows.

Motivation

• Support FSDP2+EP, enabling VERL to run large MoE models easily on FSDP2 without relying on Megatron.
• Introduce GroupGemm ops for moe and integrat Liger-Kernel for higher performance.
• Scale any omni-models easily.

Your contribution

I am going to submitting a PR. #4072

[ji-huazhong]

cc @vermouth1992 @wuxibin89