THUDM/slime

slime is an LLM post-training framework for RL Scaling.

What it solves

slime is a high-performance LLM post-training framework designed for RL scaling. It solves the problem of disconnected and heavy stacks of trainers, rollout services, and agent frameworks by unifying the training, rollout, and data generation paths into a single, streamlined flow.

How it works

slime connects Megatron-LM for training and SGLang for rollouts. It uses a "native pass-through" design, allowing users to use Megatron and SGLang arguments directly without adding abstraction layers. The architecture consists of three main modules:

  • Training (Megatron): Handles the main training process and synchronizes parameters to the rollout module.
  • Rollout (SGLang + router): Generates new data and computes rewards/verifier outputs.
  • Data Buffer: Acts as a bridge managing prompt initialization and rollout generation methods.

It supports various advanced deployment configurations, such as PD disaggregation (prefill/decode separation), delta weight synchronization for efficient updates, and external rollout engines for disaggregated serving.

Who it’s for

Researchers and engineers working on large-scale RL post-training for LLMs, specifically those using the Megatron-SGLang stack. It is designed for those needing a battle-tested framework that is lightweight enough to be extended for agentic RL, omni-modal training, or verifiable environments.

Highlights

  • Battle-tested: Used to train the GLM family of models (GLM-4.5 through GLM-5.3).
  • SGLang-Native: Deeply optimized for SGLang to leverage specific serving, routing, and caching capabilities.
  • Flexible Data Generation: Supports custom interfaces for multi-agent systems, tool use, and long-horizon agentic workflows without forking the training kernel.
  • Production-Ready: Includes built-in support for fault tolerance, tracing, profiling, and comprehensive CI/CD for RL infrastructure.
  • Broad Model Support: Compatible with Qwen, DeepSeek V3/R1, and Llama 3 series.

Related

  • Project
    vllm-project/vime

    Vime is an open‑source framework that combines Megatron‑LM training with vLLM inference to enable high‑performance reinforcement‑learning fine‑tuning of large language models. It offers a unified data‑flow loop, flexible custom generation hooks, and support for many modern LLMs (Qwen, DeepSeek, LLaMA 3).

  • Dispatch
    vime RL Framework Release

    vLLM introduces vime, an open-source RL post-training framework that integrates Megatron training with vLLM inference to provide a stable and efficient pipeline for LLM reinforcement learning.

  • Dispatch
    GLM-5.2: Built for Long-Horizon Tasks

    GLM-5.2, released by Z.AI on Hugging Face, introduces a solid 1M-token context, improved architecture via IndexShare and MTP enhancements, effort-level control, and strong open-source performance on long-horizon coding benchmarks.

  • Project
    alibaba/ROLL

    ROLL is an open‑source library for reinforcement‑learning training of large language models at scale. It provides Ray‑based distributed execution, supports many RL algorithms (PPO, GRPO, Reinforce++, etc.), works with Megatron‑Core, vLLM, SGLang, and LoRA, and offers pipelines for RL‑based fine‑tuning (RLVR) and multi‑turn agentic RL. The project includes extensive docs, hardware‑agnostic support (NVIDIA, AMD, Ascend), and built‑in observability.

  • Project
    PrimeIntellect-ai/prime-rl

    A framework for large-scale asynchronous reinforcement learning designed to train frontier MoE and dense models on up to 1000+ GPUs.