slime is an LLM post-training framework for reinforcement learning scaling that integrates Megatron training with SGLang inference.
The framework addresses the challenge of building production-grade RL post-training systems by unifying high-performance training, flexible data generation, and inference into a single coherent dataflow. Rather than assembling separate trainers, rollout services, and data generation components, slime routes Megatron training, SGLang rollout, custom data generation, reward computation, verifier feedback, and environment interaction through the same training and data buffer pipeline. This design keeps the system lightweight and understandable while enabling complex workflows including math, code, search, tools, sandboxes, verifiers, and multi-agent agentic systems without forking the training kernel.
The tool suits teams building large-scale RL post-training systems who want a battle-tested foundation rather than assembling infrastructure from scratch. It is particularly appropriate for organizations working with Megatron-scale training and SGLang-based serving, as the framework passes arguments directly to both engines without introducing abstraction layers that would obscure upstream optimizations. The project explicitly targets the Megatron plus SGLang path for large-scale RL, making it a focused choice rather than a lowest-common-denominator solution attempting to support multiple inference backends simultaneously. Teams needing to integrate custom data generation workflows, reward models, or environment interaction will find the extensible data generation interfaces designed for this purpose.
Development activity shows sustained investment in production infrastructure. The framework has been exercised through complete post-training workflows for release-grade models, including large-scale training, high-throughput rollout, weight synchronization, checkpointing, debugging, and long-running stability. Support extends across multiple model families including Qwen, DeepSeek, and Llama series, indicating ongoing maintenance and compatibility work. The project maintains explicit focus on correctness-first engineering, with separate rollout-only and train-only debugging paths, reproducibility documentation, fault tolerance, tracing, and profiling treated as first-class concerns rather than afterthoughts.
