rllm-org/rllm

Provides automatic LLM call tracing through a decorator pattern, with a dual-backend training system supporting both distributed GPU training (verl) and single-machine setups (tinker). Implements a unified agent pipeline that transparently captures token IDs and logprobs via LiteLLM proxy, enabling reward computation and policy updates using algorithms like GRPO and REINFORCE without modifying agent code.

5,219 stars. Actively maintained with 104 commits in the last 30 days.