slime-rl-training - LLM RL post-training using Megatron-LM and SGLang.
Provides guidance and workflows for LLM post-training with RL using the slime framework integrating Megatron-LM and SGLang.
Tags
Updated: 2026-09-24Capabilities
Typical Inputs
Typical Outputs
What this skill does
- Execute standard GRPO training
- Run asynchronous RL training
- Conduct multi-turn agentic training
- Configure Megatron and SGLang parameters
- Manage training data buffers
Inputs
- Model checkpoints
- JSONL format prompt data
- Model configuration scripts
- Custom generation functions
Outputs
- Saved model checkpoints
- TensorBoard logs and metrics
- Generated rollout samples
Requirements
- Docker or PyTorch environment
- NVIDIA GPUs with CUDA support
- torch >= 2.0.0
- transformers >= 4.40.0
- sglang-router >= 0.2.3
- Ray