Zhipu Open-Sources Slime RL Framework with Zero-Diff Train-Rollout Alignment

teortaxesTex · x · 2026-08-12

Slime, the reinforcement learning framework behind the training of the GLM series, has open-sourced its deterministic train–rollout alignment path.

This implementation tackles the notorious train-inference mismatch in LLM RL runs. It aligns Megatron training with SGLang rollout to an exceptionally high precision, achieving a 4096-token logprob MAE of 1.9e-7 and exact zero hidden-state diff across decoder layers 0–5. Developers dealing with train-inference mismatches in their RL pipelines are recommended to start with this framework.

Related event: Zhipu Open-Sources Slime RL Framework(2 posts)→

Original post →

More from coding & agent

coding & agent channel →