THUDM / slime
A battle-tested RL post-training framework that unifies Megatron training and SGLang rollout into a single correctness-first pipeline for scalable LLM alignment.
星标趋势
AI 分析
项目摘要
slime is an LLM post-training framework for RL scaling that tightly integrates Megatron for training and SGLang for rollout, providing a unified dataflow for reinforcement learning pipelines. It is battle-tested behind multiple frontier GLM model releases and emphasizes correctness-first infrastructure with flexible data generation capabilities.
为什么值得关注
slime is the RL post-training framework behind GLM-5.2, GLM-5.1, GLM-5, and earlier GLM releases, validating it through complete production-grade training loops rather than isolated examples. Its opinionated design avoids abstraction bloat by natively passing through Megatron and SGLang arguments, making it both lightweight and deeply optimized for large-scale RL.
优势
- Battle-tested behind multiple frontier GLM model releases, validating full production-grade RL training loops
- Tight native integration of Megatron training and SGLang rollout without unnecessary abstraction layers
- Correctness-first design with explicit dataflow, separate rollout-only/train-only debugging paths, and first-class reproducibility
局限性
- Tightly coupled to Megatron + SGLang ecosystem, limiting flexibility for users on different stacks
- No Docker support, which may complicate deployment and reproducibility for some teams
使用场景
- Large-scale RL post-training of language models with custom reward functions and verifiers
- Training with complex data generation workflows including math, code, search, tools, and multi-agent systems
- Agentic workflow training with environment interaction and long-horizon rollouts