slime — SGLang-native RL Post-Training Framework (behind GLM)
open-source framework · source date 2026-03-29 · added 2026-08-07 15:52:43 · updated 2026-08-07 15:52:43 · Open original blog
1
Problems / challenges / motivations
- RL post-training couples a training engine to a rollout engine, and synchronous coupling leaves accelerators idle.
- Agentic rollouts (tool calls, sandboxes, multi-agent loops) are slow and variable, which makes the coupling worse.
- Agentic RL is often bolted on as a separate framework rather than treated as data generation.
2
Key ideas
- Connects Megatron (training) with SGLang (rollout) through a decoupled, asynchronous training/rollout/Data-Buffer path.
- Agentic workflows — tool use, sandbox execution, verifier rewards, multi-agent loops — plug in as data generation rather than as a parallel system.
- It is the RL infrastructure behind the entire GLM family, from GLM-4.5 through GLM-5.x.
3
Why it matters for AI engineering
- slime is the most-cited open-source Chinese agentic-RL training stack, described as battle-tested by frontier model training and reused across labs (TritonForge, Ascend ports, Relax).
- It is the clearest articulation of agentic RL as infrastructure, and the training-side counterpart to "agent = model + harness."
- The decoupled-async design is the reusable idea, independent of any model trained with it.
- Caveat: adoption is real, but capability claims for models trained on it (GLM SWE-bench and Terminal-Bench rates) are vendor-reported. As open infrastructure it carries no single validated benchmark of its own.
Comments
No comments yet.