Slime

RL post-training for LLMs with Megatron and SGLang.

By Orchestra Research · Official Docs · Date unknown

slime is an LLM post-training framework from Tsinghua's THUDM team, powering GLM-4.5, GLM-4.6, and GLM-4.7. It connects Megatron-LM for training with SGLang for high-throughput rollout generation.

Platforms: linux, macos.

hermes skills install official/mlops/slime

• #skills
• #optional
• #mlops
• #reinforcement learning
• #megatron-lm
• #sglang
• #grpo