[Roadmap] OpenRLHF Development Roadmap

[hijkzzz]

OpenRLHF-M Roadmap

Code Migration

• Complete LMM R1 code migration for the OpenRLHF organization. @TideDra
• Align migrated code with OpenRLHF’s training paradigm.
• Support Packing Samples @TideDra
• Support RingAttention @TideDra
• Code rebase regularly @TideDra

Finalize Scope of Supported Open-Source Models

• Determine supported open-source models:
  • QwenVL (one of the top SOTA models in China)
  • InternVL (one of the top SOTA models in China)
  • LLava (the first global open-source multimodal large model)
  • DeepSeekVL (one of the few MLLMs supporting MoE architecture)
  • MiniCPM-V (focused on edge-side applications with high demand)
  • Future models (may include new ones and unified framework models like Janus)