{"full_name":"THUDM/slime","owner":"THUDM","name":"slime","shard":"1","url":"https://github.com/THUDM/slime","description":"slime is an LLM training framework for RL Scaling.","homepage":"https://thudm.github.io/slime","topics":[],"tag_pages":[],"language":"Python","license":"Apache-2.0","avatar":"https://avatars.githubusercontent.com/u/48590610?v=4","created_at":"2025-06-18T07:15:54Z","pushed_at":"2026-11-05T23:42:16Z","metrics":{"stars":8828,"forks":1364,"watchers":33,"issues":526,"prs":2002,"releases":11,"issues_open":240,"issues_closed":286,"prs_open":334,"prs_closed":283,"prs_merged":1385,"discussions":0},"growth":{"day":{"stars":6,"forks":0,"base_day":"2026-11-15"}},"history":{"dates":["2026-11-10","2026-11-11","2026-11-12","2026-11-13","2026-11-14","2026-11-15","2026-11-16"],"stars":[8810,8814,8816,8818,8820,8822,8828],"forks":[1355,1358,1361,1361,1364,1364,1364],"watchers":[null,33,33,33,33,33,33],"issues_open":[null,239,240,240,240,240,240],"prs_open":[null,324,331,331,333,334,334]},"boards":{},"same_owner":[{"full_name":"THUDM/ChatGLM-6B","name":"ChatGLM-6B","shard":"1","boards":{}},{"full_name":"THUDM/ChatGLM2-6B","name":"ChatGLM2-6B","shard":"1","boards":{}},{"full_name":"THUDM/ChatGLM3","name":"ChatGLM3","shard":"1","boards":{}}],"issues":[{"number":1820,"title":"[Bug] GLM5模型转megatron报错","comments":14,"created_at":"2026-04-09T03:54:28Z"},{"number":1673,"title":"[Bug] UpdateWeightFromDistributed fails for VLM models in non-colocate mode","comments":7,"created_at":"2026-03-07T00:18:36Z"},{"number":1026,"title":"ERROR about sglang_router","comments":11,"created_at":"2025-12-04T06:18:55Z"},{"number":795,"title":"[Bug] `Unexpected result nan` on 64x H100, Qwen3 235B","comments":11,"created_at":"2025-11-17T09:26:38Z"},{"number":1484,"title":"Slime + Megatron backend: training frequently hangs after SGLang CUDA graph capture & server startup (Uvicorn ready), no further progress","comments":10,"created_at":"2026-01-23T15:23:12Z"},{"number":1482,"title":"About GLM-4.7-Flash","comments":10,"created_at":"2026-01-23T10:18:41Z"},{"number":1109,"title":"AssertionError in _log_rollout_data when training qwen3-vl-8B with true_on_policy_mode","comments":10,"created_at":"2025-12-15T02:30:53Z"},{"number":1202,"title":"LoRA support GRPO training and examples","comments":9,"created_at":"2025-12-24T04:00:48Z"}],"discussions":[],"latest_release":{"name":"v0.4.0: A First Step Toward Greater Scale","tag":"v0.4.0","published_at":"2026-10-09T03:04:12Z","body":"OpenAI's [Navier–Stokes research effort](https://openai.com/index/navier-stokes-solution/) involved a group of roughly 10,000 concurrent agents and used approximately 130 billion output tokens during problem-solving. For us, it offered a concrete glimpse of how large agent workloads can become—and prompted us to think more ambitiously about the RL infrastructure that trains them.\r\n\r\nPreparing for that scale means giving rollout orchestration room to grow across nodes, making data flow durable, and preserving useful work when training fails. It also means focusing our development effort so that algorithm exploration and training at scale share a foundation we can keep improving.\r\n\r\nslime v0.4.0 is a first step in that direction. We are focusing on synchronous and fully asynchronous training, distributing rollout orchestration across the cluster, introducing persistent queues and tensor storage with straw, and enabling Megatron to recover independently of healthy SGLang serving resources. These changes lay the groundwork for higher concurrency and longer runs, with training correctness remaining central to slime's design.\r\n\r\n## One training lifecycle, two modes\r\n\r\nWe are focusing development on two training modes:\r\n\r\n- **Synchronous training** for exploring algorithmic limits and validating correctness.\r\n- **Fully asynchronous training** for keeping rollout and training resources busy at larger scale.\r\n\r\nBoth modes now use `train.py`, with asynchronous scheduling handled by the rollout implementation. We have removed the legacy one-step off-policy driver, `train_async.py`, giving both workflows a shared training lifecycle to maintain and improve. ([[#2391](https://github.com/THUDM/slime/pull/2391)](https://github.com/THUDM/slime/pull/2391))\r\n\r\nAlongside this simplification, we are expanding the tools available for stable training. v0.4.0 adds **score centering**, using a REINFORCE objective with support for top-p replay and the built-in importance-sampling corrections. Enabled with `--use-score-centering`, it provides another tool for addressing training–inference mismatch as we scale rollout and training. ([[#2405](https://github.com/THUDM/slime/pull/2405)](https://github.com/THUDM/slime/pull/2405), [[#2406](https://github.com/THUDM/slime/pull/2406)](https://github.com/THUDM/slime/pull/2406))\r\n\r\n## A distributed, persistent rollout pipeline\r\n\r\nIncreasing rollout capacity also requires orchestration to grow with the cluster. With a new storing backend **straw** enabled, fully asynchronous rollout now runs **one rollout worker per Ray node with CPU resources**, each managing its own generation pool and asyncio event loop. This distributes orchestration across the CPU resources already available in the cluster and reduces the bottleneck of a single process. Training batches can collect completed groups without waiting for every worker. ([[#2410](https://github.com/THUDM/slime/pull/2410)](https://github.com/THUDM/slime/pull/2410))\r\n\r\nThis distributed execution is backed by **[[straw](https://github.com/zhuzilin/straw)](https://github.com/zhuzilin/straw)**, a library providing durable queues and shared tensor storage on a filesystem. Rollout workers publish payloads to shared storage, and training ranks read their assigned data by reference. Bulk data between rollout and training moves off Ray’s object-store transport, while Ray continues to carry control messages and references.\r\n\r\nThe same data layer persists **prompt tasks, partial rollouts, completed samples, and converted training batches**. Coordinating this progress with model checkpoints establishes a consistent recovery boundary for asynchronous training and supports checkpoint rollback. Samples and tensors are packed into large append-only files to reduce small-file and shared-filesystem metadata overhead. ([[#2410](https://github.com/THUDM/slime/pull/2410)](https://github.com/THUDM/slime/pull/2410), [[#2427](https://github.com/THUDM/slime/pull/2427)](https://github.com/TH","is_prerelease":false},"readme":true,"readme_zh":true,"news":[],"stale":false,"snapshot_at":"2026-11-16T09:23:14Z","history_start":"2026-09-06"}