[{"data":1,"prerenderedAt":4},["ShallowReactive",2],{"readme:slime":3},"\u003Ch1>slime\u003C\u002Fh1>\n\u003Cp>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FTHUDM\u002Fslime\u002Fblob\u002FHEAD\u002FREADME_zh.md\" rel=\"nofollow ugc noopener\">中文版\u003C\u002Fa>\u003C\u002Fp>\n\u003Cp>\u003Ca href=\"https:\u002F\u002Fthudm.github.io\u002Fslime\u002F\" rel=\"nofollow ugc noopener\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fdocs-latest-brightgreen.svg?style=flat\" alt=\"Documentation\" \u002F>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fgithub.com\u002FTHUDM\u002Fslime\u002Fpull\u002F2053\u002Fchecks\" rel=\"nofollow ugc noopener\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Factions\u002Fworkflow\u002Fstatus\u002FTHUDM\u002Fslime\u002Fpr-test.yml?branch=zilin%2Fci-dont-merge&event=pull_request&label=CI&logo=github\" alt=\"CI\" \u002F>\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fdeepwiki.com\u002FTHUDM\u002Fslime\" rel=\"nofollow ugc noopener\">\u003Cimg src=\"https:\u002F\u002Fdeepwiki.com\u002Fbadge.svg\" alt=\"Ask DeepWiki\" \u002F>\u003C\u002Fa>\u003C\u002Fp>\n\u003Cp>\u003Cstrong>slime\u003C\u002Fstrong> is an LLM post-training framework for RL scaling, providing two core capabilities:\u003C\u002Fp>\n\u003Col>\n\u003Cli>\u003Cstrong>High-Performance Training\u003C\u002Fstrong>: Supports efficient training in various modes by connecting Megatron with SGLang;\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Flexible Data Generation\u003C\u002Fstrong>: Enables arbitrary training data generation workflows through custom data generation interfaces and server-based engines.\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>slime's design goal is to make these two capabilities reinforce each other without turning the system into a heavy stack of disconnected trainers, rollout services, and agent frameworks. Megatron training, SGLang rollout, custom data generation, reward computation, verifier feedback, and environment interaction all flow through the same training \u002F rollout \u002F Data Buffer path.\u003C\u002Fp>\n\u003Cp>This makes slime one of the most battle-tested open RL post-training frameworks: small enough to understand and extend, but validated through complete training loops behind SOTA-level model releases.\u003C\u002Fp>\n\u003Ch2>Why This Design Matters\u003C\u002Fh2>\n\u003Cul>\n\u003Cli>\u003Cstrong>Battle-tested by frontier model training\u003C\u002Fstrong>: slime is the RL framework behind \u003Ca href=\"https:\u002F\u002Fz.ai\u002Fblog\u002Fglm-5.2\" rel=\"nofollow ugc noopener\">GLM-5.2\u003C\u002Fa>, \u003Ca href=\"https:\u002F\u002Fz.ai\u002Fblog\u002Fglm-5.1\" rel=\"nofollow ugc noopener\">GLM-5.1\u003C\u002Fa>, \u003Ca href=\"https:\u002F\u002Fz.ai\u002Fblog\u002Fglm-5\" rel=\"nofollow ugc noopener\">GLM-5\u003C\u002Fa>, \u003Ca href=\"https:\u002F\u002Fz.ai\u002Fblog\u002Fglm-4.7\" rel=\"nofollow ugc noopener\">GLM-4.7\u003C\u002Fa>, \u003Ca href=\"https:\u002F\u002Fz.ai\u002Fblog\u002Fglm-4.6\" rel=\"nofollow ugc noopener\">GLM-4.6\u003C\u002Fa>, and \u003Ca href=\"https:\u002F\u002Fz.ai\u002Fblog\u002Fglm-4.5\" rel=\"nofollow ugc noopener\">GLM-4.5\u003C\u002Fa>. This validates the full post-training loop, not only isolated examples.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Correctness-first infrastructure\u003C\u002Fstrong>: RL bugs are often silent. slime keeps the dataflow explicit, supports separate rollout-only and train-only debugging paths, and documents reproducibility, fault tolerance, tracing, profiling, and CI as first-class engineering concerns.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Native by design\u003C\u002Fstrong>: slime passes Megatron arguments through directly and exposes installed SGLang arguments with a \u003Ccode>--sglang-\u003C\u002Fcode> prefix. New upstream training and serving optimizations can be used without adding another abstraction layer inside slime.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Maximum data-generation freedom\u003C\u002Fstrong>: math, code, search, tools, sandboxes, verifiers, environments, multi-agent systems, and long-horizon agentic workflows plug in as data generation or reward workflows. They do not fork the training kernel.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Lightweight and opinionated\u003C\u002Fstrong>: slime focuses deeply on the Megatron + SGLang path used for large-scale RL. By choosing one rollout backend, slime can use SGLang-specific capabilities directly instead of flattening multiple inference engines into a lowest-common-denominator abstraction.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2>Production Validation\u003C\u002Fh2>\n\u003Cp>slime has been exercised by the complete workflow needed for release-grade model post-training: large-scale training, high-throughput rollout, weight synchronization, reward\u002Fverifier data, checkpointing, debugging, and long-running stability.\u003C\u002Fp>\n\u003Cp>Beyond the GLM family, slime also supports:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Qwen series: Qwen3.6, Qwen3.5, Qwen3Next, Qwen3MoE, Qwen3, Qwen2.5;\u003C\u002Fli>\n\u003Cli>DeepSeek V3 series: DeepSeek V3, V3.1, DeepSeek R1;\u003C\u002Fli>\n\u003Cli>Llama 3.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2>Native Engine Pass-Through and SGLang Deployment\u003C\u002Fh2>\n\u003Cp>slime is not just a framework that can call an inference backend. It keeps the Megatron and SGLang control surfaces close to the upstream engines while adding the RL dataflow around them:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>native SGLang argument pass-through: every argument supported by the installed SGLang can be used by adding the \u003Ccode>--sglang-\u003C\u002Fcode> prefix, such as passing \u003Ccode>--mem-fraction-static\u003C\u002Fcode> as \u003Ccode>--sglang-mem-fraction-static\u003C\u002Fcode>;\u003C\u002Fli>\n\u003Cli>native Megatron argument pass-through: slime reads Megatron arguments directly, so Megatron-side parallelism, optimizer, checkpointing, and model options remain available without wrapper code;\u003C\u002Fli>\n\u003Cli>[SGLang Config](docs\u002Fen\u002Fadvanced\u002Fsglan\u003C\u002Fli>\n\u003C\u002Ful>\n",1784240408134]