Changelog

​

Version History

​

v0.1.0 - Redefining High-Performance RL Training Frameworks

​

Key Features

• Efficient training in various modes by connecting Megatron with SGLang
• Support for co-located and decoupled training and inference
• Dynamic batch sizing for optimal GPU utilization
• Data packing and variable-length processing enabled by default

• Arbitrary training data generation workflows
• Custom data generation interfaces
• Server-based generation engines
• Integration with SGLang for high-throughput rollout

• GLM series (GLM-5, GLM-4.7, GLM-4.6, GLM-4.5)
• Qwen3 series (Qwen3Next, Qwen3MoE, Qwen3)
• Qwen2.5 series
• DeepSeek V3 series (DeepSeek V3, V3.1, DeepSeek R1)
• Llama 3

• Speculative decoding in RL - Community collaboration for faster inference
• Low-precision training - fp8 rollout + bf16/fp8 training, int4 rollout + int4 QAT training
• Deterministic training - Reproducible training runs for research

​

Architecture

1. Training (Megatron) - Main training process that reads from the Data Buffer and synchronizes parameters to rollout
2. Rollout (SGLang + router) - Generates new data including rewards/verifier outputs
3. Data Buffer - Bridge module managing prompt initialization, custom data, and rollout generation

​

Community Projects

• P1 - Physics Olympiad reasoning models trained with RL
• RLVE - Scaling LM RL with adaptive verifiable environments
• TritonForge - Agentic RL training for kernel generation
• APRIL - Accelerating RL training with active partial rollouts
• qqr - Scaling open-ended agents with ArenaRL & MCP

​

Configuration

1. Megatron arguments - Full support for Megatron configuration
2. SGLang arguments - All SGLang arguments with --sglang- prefix
3. slime-specific arguments - Custom parameters for RL training workflows

​

Developer Tools

• Pre-commit hooks for code style consistency
• Comprehensive debugging guide
• Clear contribution guidelines focused on bug fixes and performance optimizations

​

Acknowledgements

• SGLang
• Megatron-LM
• mbridge
• OpenRLHF
• veRL
• Pai-Megatron-Patch

​

Citation

@misc{slime_github,
  author       = {Zilin Zhu and Chengxing Xie and Xin Lv and slime Contributors},
  title        = {slime: An LLM post-training framework for RL Scaling},
  year         = {2025},
  howpublished = {\url{https://github.com/THUDM/slime}},
  note         = {GitHub repository. Corresponding author: Xin Lv},
  urldate      = {2025-06-19}
}

​

Future Releases