Interest in Contributing: Integrating FastGRPO's Speculative Decoding Innovations to Enhance SLIME

[yedaotian9]

Hi there,

We are the authors of FastGRPO: Accelerating Policy Optimization via Concurrency-aware Speculative Decoding and Online Draft Learning. In our work, we explore how speculative decoding can be leveraged to accelerate the rollout phase in GRPO-based reinforcement learning training.

We noticed that your project is also investigating similar directions in applying speculative decoding to RL workflows, which aligns closely with our research. We’re excited about the potential synergy and would love to contribute our ideas—particularly our concurrency-aware speculation framework—to help further improve the efficiency of your system.

Could you kindly let us know the best way to get involved as contributors? We’d be happy to collaborate and integrate our advancements to push the performance even further.

Looking forward to your thoughts!