[Question] Guidance on integrating MiMo-V2-Flash with VeRL (v0.6.1/v0.7.0): Should I use the mbridge/mcore path?

[Eisenhower]

Description I am currently using verl versions 0.6.1 and 0.7.0. My goal is to integrate the MiMo-V2-Flash model for RL training.

Questions

Should I use the mbridge (Megatron-Core integration) scheme for this model?

Do you have any specific suggestions or best practices for integrating a model with this architecture?

Context

VeRL Version: 0.6.1 / 0.7.0

Target Model: MiMo-V2-Flash

Current Plan: Exploring the mbridge / Megatron-Core path.

Any advice would be appreciated. Thanks!
@ISEEKYAN

[ISEEKYAN]

Mow the bridge path is a default path of megatron in verl, and the legacy path without bridge is being deprecated.

Either mbridge and megatron-bridge is ok for your need, but the megatron-bridge is been refactored lately(ETA for 2 weeks), so if you want a very fast integration, it is better to try expand mbridge now.