Show HN：NanoRL——约 1,800 行实现面向 LLM 的强化学习训练
我可以编写的最小的异步 RL 训练器：一个循环在笔记本电脑上的 CartPole 上运行 REINFORCE，并在集群上运行异步 GRPO（例如 8xH100 训练器、8 个 vLLM 工作人员，在 k8s 上作为 [SkyPilot 作业组]( https://docs.skypilot.ai/en/latest/examples/job-groups.html ) 运行）。所有这些都没有 Ray 或 TRL 或 DeepSpeed 等，工作人员通过 stdlib HTTP 与训练器对话。
查看英文原文
Show HN: NanoRL – RL training for LLMs in ~1,800 lines
The smallest async RL trainer I could write: one loop that runs REINFORCE on CartPole on a laptop and async GRPO on a cluster (e.g. 8xH100 trainer, 8 vLLM workers, ran as a [SkyPilot job group]( https://docs.skypilot.ai/en/latest/examples/job-groups.html ) on k8s ). All without Ray or TRL or DeepSpeed etc., workers talk to the trainer over stdlib HTTP.
Show HN: NanoRL – RL training for LLMs in ~1,800 lines
The smallest async RL trainer I could write: one loop that runs REINFORCE on CartPole on a laptop and async GRPO on a cluster (e.g. 8xH100 trainer, 8 vLLM workers, ran as a [SkyPilot job group](<a href="https://docs.skypilot.ai/en/latest/examples/job-groups.html" rel="nofollow">https://docs.skypilot.ai/en/latest/examples/job-groups.html</a>) on k8s ).<p>All without Ray or TRL or DeepSpeed etc., workers talk to the trainer over stdlib HTTP.
- 来源
- Show HN · show-hn
- 证据角色
- project_and_comment_discovery
- 时间状态
- known
- 用户判决
- 未判