{"id":31272,"date":"2026-07-11T13:47:35","date_gmt":"2026-07-11T05:47:35","guid":{"rendered":"http:\/\/139.9.1.231\/?p=31272"},"modified":"2026-07-11T13:47:37","modified_gmt":"2026-07-11T05:47:37","slug":"openrlhf-asr","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2026\/07\/11\/openrlhf-asr\/","title":{"rendered":"\u57fa\u4e8e OpenRLHF \u7684\u5927\u6a21\u578b\u5f3a\u5316\u8bad\u7ec3"},"content":{"rendered":"\n<ul><li><strong>Github\uff1a https:\/\/github.com\/OpenRLHF\/OpenRLHF<\/strong><\/li><li><strong>\u8bf4\u660e\u6587\u6863\uff1ahttps:\/\/www.aidoczh.com\/openrlhf\/<\/strong><\/li><li><a rel=\"noreferrer noopener\" href=\"https:\/\/github.com\/zhaochenyang20\/Awesome-ML-SYS-Tutorial\/blob\/main\/rlhf\/OpenRLHF\/readme.md\" target=\"_blank\"><strong>\u6d45\u6790\u4ee5 OpenRLHF \u4e3a\u4ee3\u8868\u7684 post-training \u7cfb\u7edf\u7684\u8ba1\u7b97\u6d41\u7a0b<\/strong><\/a><\/li><li><strong><a href=\"https:\/\/zhuanlan.zhihu.com\/p\/677607581\">\u56fe\u89e3\u5927\u6a21\u578bRLHF\u7cfb\u5217\u4e4b\uff1a\u4eba\u4eba\u90fd\u80fd\u770b\u61c2\u7684PPO\u539f\u7406\u4e0e\u6e90\u7801\u89e3\u8bfb<\/a><\/strong><\/li><li><strong><a href=\"https:\/\/zhuanlan.zhihu.com\/p\/12871616401\">\u56fe\u89e3OpenRLHF\u4e2d\u57fa\u4e8eRay\u7684\u5206\u5e03\u5f0f\u8bad\u7ec3\u6d41\u7a0b<\/a><\/strong><\/li><li><strong>ASR\u5f3a\u5316: Explore the Reinforcement Learning for the LLM based ASR and TTS system: <a href=\"https:\/\/arxiv.org\/pdf\/2509.18569v1\"><em>https:\/\/arxiv.org\/pdf\/2509.18569v1<\/em><\/a><\/strong><\/li><\/ul>\n\n\n\n<h2>OpenRLHF\u4ee3\u7801\u7ec6\u8282<\/h2>\n\n\n\n<p><a href=\"https:\/\/zhuanlan.zhihu.com\/p\/12871616401\"><strong>https:\/\/zhuanlan.zhihu.com\/p\/12871616401<\/strong><\/a><\/p>\n\n\n\n<h3>\u8bad\u7ec3\u5165\u53e3<\/h3>\n\n\n\n<figure class=\"wp-block-pullquote\"><blockquote><p><strong>ppo_ray<\/strong>\u76f8\u5173\u7684\u8bad\u7ec3\u5165\u53e3\u5728\uff1a<a rel=\"noreferrer noopener\" href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/cli\/train_ppo_ray.py\" target=\"_blank\">https:\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/cli\/train_ppo_ray.py<\/a>\u3002<\/p><\/blockquote><\/figure>\n\n\n\n<p>\u5728main\u4e2d\u6211\u4eec\u542f\u52a8\u4e86driver\u8fdb\u7a0b\uff0c\u5e76\u6267\u884c\u8bad\u7ec3\u51fd\u6570<code>train(args)<\/code>\uff0c\u8fd9\u91cc\u4e3b\u8981\u505a\u4e86\u5982\u4e0b\u51e0\u4ef6\u4e8b\uff1a<\/p>\n\n\n\n<ul><li><strong>\u5728ray\u96c6\u7fa4\u4e0a\u90e8\u7f72Actor\/Ref\/Critic\/RM\u5b9e\u4f8b<\/strong><\/li><li><strong>\u5728ray\u96c6\u7fa4\u4e0a\u90e8\u7f72vllm_engines\u5b9e\u4f8b<\/strong><\/li><li><strong>\u914d\u7f6eActor\u548cvllm_engines\u4e4b\u95f4\u7684\u901a\u8baf\uff0c\u7528\u4e8e\u4f20\u9012\u6743\u91cd<\/strong><\/li><li><strong>\u8bad\u7ec3Actor\u548cCritic\u6a21\u578b<\/strong><\/li><\/ul>\n\n\n\n<p>\u6211\u4eec\u4f9d\u6b21\u6765\u89e3\u8bfb\u8fd9\u51e0\u4e2a\u5173\u952e\u6b65\u9aa4\u3002<strong>\u540c\u65f6\u4e3a\u4e86\u5728\u8868\u8ff0\u4e0a\u6d88\u9664\u6b67\u4e49\uff0c\u6211\u4eec\u63a5\u4e0b\u6765\u8c08\u5230\u201cActor\u201d\u65f6\uff0c\u4f1a\u4f7f\u7528Ray-Actor\u548cPPO-Actor\u6765\u505a\u533a\u5206<\/strong>\uff0c\u4ece\u4e4b\u524d\u7684\u4ecb\u7ecd\u4e2d\u53ef\u77e5\uff0cRay-Actor\u662f\u6307\u90e8\u7f72\u5728Ray\u96c6\u7fa4\u4e2d\u7684\u8fdc\u7aefclass\uff0cPPO-Actor\/Ref\/Critic\/RM\u90fd\u5c5e\u4e8eRay-Actor\u3002<\/p>\n\n\n\n<h3>\u90e8\u7f72Actor\/Ref\/Critic\/RM\u5b9e\u4f8b<\/h3>\n\n\n\n<h4 id=\"h_12871616401_15\">\uff081\uff09\u975e\u5171\u540c\u90e8\u7f72<\/h4>\n\n\n\n<p>\u9488\u5bf9\u591a\u4e2anode\u7684\u60c5\u51b5\uff0c\u6211\u4eec\u4ee5PPO-Actor\u4e3a\u4f8b\uff0c\u770b\u4ee3\u7801\u662f\u5982\u4f55\u5c06\u5176\u90e8\u7f72\u5230Ray\u96c6\u7fa4\u4e0a\u7684\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/pica.zhimg.com\/v2-c46a2e47aa48f3c0a42eecc5003e28ee_1440w.jpg\" alt=\"\"\/><\/figure>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/pic3.zhimg.com\/v2-a7445701e230850618a1a055ad9a8cec_1440w.jpg\" alt=\"\"\/><\/figure>\n\n\n\n<ul><li><code>PPORayActorGroup<\/code>\uff1a<strong>\u521b\u5efa\u5728driver\u8fdb\u7a0b\u4e0a\uff0c\u53ef\u5c06\u5b83\u7406\u89e3\u6210\u4e00\u79cd<a href=\"https:\/\/zhida.zhihu.com\/search?content_id=251597453&amp;content_type=Article&amp;match_order=1&amp;q=%E9%83%A8%E7%BD%B2%E6%96%B9%E6%A1%88&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">\u90e8\u7f72\u65b9\u6848<\/a>\uff0c\u4e13\u95e8\u8d1f\u8d23\u90e8\u7f72PPO\u4e2d\u76844\u7c7b\u6a21\u578b<\/strong>\u3002<ul><li><code>PPORayActorGroup<\/code>\u4e2d\u7ef4\u62a4\u7740<code>self._actor_handlers<\/code>\uff0c\u5b83\u662f\u4e00\u4e2a<code>List[ray.actor.ActorHandle]<\/code>\uff0c\u5217\u8868\u4e2d\u6bcf\u4e2a\u5143\u7d20\u8868\u793a\u67d0\u4e2a\u8fdc\u7aefRay-Actor\u7684\u5f15\u7528\uff0c\u800c\u8fd9\u4e2a\u8fdc\u7aefRay-Actor\u53ef\u4ee5\u662fPPO-Actor\/Ref\/Critic\/RM\u5b9e\u4f8b\u3002\u5982\u524d\u6587\u6240\u8bf4\uff0c\u6211\u4eec\u53ef\u4ee5\u5728<a href=\"https:\/\/zhida.zhihu.com\/search?content_id=251597453&amp;content_type=Article&amp;match_order=3&amp;q=ray%E9%9B%86%E7%BE%A4&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">ray\u96c6\u7fa4<\/a>\u4e2d\u7684\u4efb\u4f55\u4f4d\u7f6e\u8c03\u7528\u8fd9\u4e2ahandler\uff0c\u6765\u5bf9\u76f8\u5e94\u7684\u8fdc\u7aefRay-Actor\u6267\u884c\u64cd\u4f5c\u3002<\/li><li>\u5728\u672c\u4f8b\u4e2d\uff0c\u6211\u4eec\u521b\u5efa\u4e864\u4e2aRay-Actor\uff081\u4e2amaster-actor\uff0c3\u4e2aworker_actor\uff09\u3002\u6bcf\u4e2aRay-Actor\u90fd\u8fd0\u884c\u5728\u4e00\u4e2aworker\u8fdb\u7a0b\u4e2d\u3002\u5728\u521b\u5efaRay-Actor\u7684\u540c\u65f6\uff0c\u6211\u4eec\u4e5f\u4f1a\u53bb\u4fee\u6539worker\u8fdb\u7a0b\u7684<a href=\"https:\/\/zhida.zhihu.com\/search?content_id=251597453&amp;content_type=Article&amp;match_order=1&amp;q=%E7%8E%AF%E5%A2%83%E5%8F%98%E9%87%8F&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">\u73af\u5883\u53d8\u91cf<\/a>\u3002\u540e\u7eed\u5f53\u6211\u4eec\u5728\u8fd9\u4e9bworker\u8fdb\u7a0b\u4e2d\u542f\u52a8ds_zero\u76f8\u5173\u7684\u5206\u5e03\u5f0f\u914d\u7f6e\u65f6\uff0cds\u4f1a\u8bfb\u53d6\u8fd9\u4e9b\u73af\u5883\u53d8\u91cf\u4fe1\u606f\uff0c\u8fd9\u6837\u6211\u4eec\u5c31\u77e5\u9053\u54ea\u4e9bRay-Actor\u540c\u65f6\u53c8\u6784\u6210ds\u4e2d\u7684\u6570\u636e\u5e76\u884c\u7ec4\u3002<\/li><li>\u4f7f\u7528<code>PPORayActorGroup<\/code>\u90e8\u7f72\u6a21\u578b\u5b9e\u4f8b\u7684\u4ee3\u7801\u5982\u4e0b\uff1a<\/li><\/ul><\/li><\/ul>\n\n\n\n<pre class=\"wp-block-code\"><code>model = PPORayActorGroup(\n        # \u4e3a\u90e8\u7f72\u8be5\u6a21\u578b\u7684\u5168\u90e8\u5b9e\u4f8b\uff0c\u6211\u4eec\u60f3\u7528\u591a\u5c11\u53f0node\uff0c\u4f8b\u5982\u672c\u4f8b\u4e2d\u4e3a2\n        args.actor_num_nodes,\n        # \u4e3a\u90e8\u7f72\u8be5\u6a21\u578b\u7684\u5168\u90e8\u5b9e\u4f8b\uff0c\u6211\u4eec\u6bcf\u53f0node\u4e0a\u60f3\u7528\u591a\u5c11gpu\uff0c\u4f8b\u5982\u672c\u4f8b\u4e2d\u4e3a2\n        args.actor_num_gpus_per_node,\n        # Actor\/Critic\/Reward\/ReferenceRayActor\n        ActorModelRayActor, \n        # pg\u53ef\u7406\u89e3\u4e3a\uff0c\u5728ray cluster\u4e2d\u9501\u5b9a\/\u9884\u7559\u4e00\u7247\u8d44\u6e90\uff0c\u7136\u540e\u53ea\u5728\u8fd9\u7247\u8d44\u6e90\u4e0a\u90e8\u7f72\u8be5\u6a21\u578b\u5168\u90e8\u5b9e\u4f8b\u3002\n        # \uff08pg\u7ef4\u62a4\u5728Head Node\u7684GCS\u4e0a\uff0c\u53c2\u89c13.3\uff09\n        # \u4f8b\u5982\u672c\u4f8b\u4e2d\uff0cpg\u9501\u5b9a\u7684\u8d44\u6e90\u4e3anode0 gpu0\/1, node1 gpu0\/1\uff0c\n        # \u6211\u4eec\u53ea\u5728\u4e0a\u9762\u90e8\u7f72ActorModelRayActor\u5168\u90e8\u5b9e\u4f8b\n        pg=pg,\n        # \u5f53\u6211\u4eec\u5728pg\u6307\u5411\u7684\u9884\u7559\u8d44\u6e90\u4e2d\u5206\u914d\u6a21\u578b\u5b9e\u4f8b\u65f6\uff0c\u518d\u8fdb\u4e00\u6b65\u6307\u5b9a\u6bcf\u4e2a\u5b9e\u4f8b\u5360\u636e\u4e00\u5f20gpu\u7684\u591a\u5c11\u90e8\u5206\n        # \u7b49\u4e8e1\u8bf4\u660e\u6bcf\u4e2a\u5b9e\u4f8b\u5360\u6ee1\u4e00\u5f20gpu\uff0c\u5373\u201c\u975e\u5171\u540c\u90e8\u7f72\u201d\n        # \u5c0f\u4e8e1\u8bf4\u660e\u6bcf\u4e2a\u5b9e\u4f8b\u53ea\u5360\u90e8\u5206gpu\uff0c\u5373\u201c\u5171\u540c\u90e8\u7f72\u201d\uff0c\u4f8b\u5982PPO-Actor\/Ref\u5171\u540c\u90e8\u7f72\u5728\u4e00\u5f20\u5361\u4e0a\n        num_gpus_per_actor=0.75 if pg else 1,\n    )\n\n<\/code><\/pre>\n\n\n\n<p><code>ActorModelRayActor<\/code>\uff1a<strong>\u521b\u5efa\u5728\u8fdc\u7aefworker\u8fdb\u7a0b\u4e0a\uff0c\u662fRay-Actor<\/strong>\u3002\u5b83\u5305\u542b\u4e86\u8bbe\u7f6eds_zero\u5206\u5e03\u5f0f\u73af\u5883\u3001\u52a0\u8f7d\u6a21\u578b\u6743\u91cd\u3001\u6570\u636e\u96c6\u51c6\u5907\u3001optimizer\/scheduler\u51c6\u5907\u3001\u8bad\u7ec3\u7b49\u4e00\u7cfb\u5217\u64cd\u4f5c\u3002<\/p>\n\n\n\n<h4 id=\"h_12871616401_16\">\u5171\u540c\u90e8\u7f72<\/h4>\n\n\n\n<p>\u9488\u5bf9\u4e0b\u56fe\u7684\u60c5\u51b5\uff0c\u6211\u4eec\u4ee5PPO-Actor\u4e3a\u4f8b\uff0c\u770b\u4ee3\u7801\u662f\u5982\u4f55\u5c06\u5176\u90e8\u7f72\u5230Ray\u96c6\u7fa4\u4e0a\u7684<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/pic2.zhimg.com\/v2-b91c1b4dd04d93e8b06674f47099304f_1440w.jpg\" alt=\"\"\/><\/figure>\n\n\n\n<ul><li><code>PPORayActorGroup<\/code>\uff1a\u5728driver\u8fdb\u7a0b\u4e0a\u521b\u5efa2\u4e2a<code>PPORayActorGroup<\/code>\uff0c\u5206\u522b\u7ba1\u7406PPO-Actor\uff0cPPO-Ref\u7684\u90e8\u7f72<\/li><li>\u4f7f\u7528<code>actor_model = PPORayActorGroup(..., pg = pg, num_gpus_per_actor=0.75)<\/code>\u521b\u5efaPPO-Actor\u90e8\u7f72\u65b9\u6848\u5b9e\u4f8b\uff1b\u4f7f\u7528<code>ref_model = PPORayActorGroup(..., pg = pg, num_gpus_per_actor=0.25)<\/code>\u521b\u5efaPPO-Ref\u90e8\u7f72\u65b9\u6848\u5b9e\u4f8b<\/li><li>\u8fd9\u91cc\uff0c\u4e24\u4e2a\u65b9\u6848\u5b9e\u4f8b\u4f7f\u7528\u7684pg\u90fd\u662f\u540c\u4e00\u4e2a\uff0c\u5373\u8fd9\u4e2apg\u90fd\u6307\u5411\u201c1\u53f0node\uff0c\u6bcf\u53f0node 8\u5f20\u5361\u201d\u8fd9\u7247\u9884\u7559\u597d\u7684\u8d44\u6e90\u3002<\/li><li><strong><code>num_gpus_per_actor = 0.75\/0.25<\/code>\u662f\u4e00\u79cd\u521b\u5efatrick<\/strong>\uff0c\u867d\u7136\u6211\u4eec\u7684\u6700\u7ec8\u76ee\u7684\u662f\u4e3a\u4e86\u8ba9PPO-Actor\u548cPPO-Ref\u5bf9\u534a\u5206\u4e00\u5f20\u5361\uff08\u5bf9\u534a=\u5171\u4eab\uff0c\u4e0d\u662f\u6307\u663e\u5b58\u4e0a\u5bf9\u534a\u5206\uff09\uff0c\u4f46\u662f\uff1a<ul><li>\u5047\u8bbe\u8bbe\u7f6e\u4e3a0.5\uff0c\u5f53\u6211\u4eec\u5b9e\u9645\u90e8\u7f72<code>ActorModelRayActor<\/code>\u65f6\uff0cRay\u5148\u5728\u5355\u5361\u4e0a\u90e8\u7f721\u4e2a<code>ActorModelRayActor<\/code>\u5b9e\u4f8b\uff0c\u5f53\u5b83\u51c6\u5907\u90e8\u7f72\u7b2c\u4e8c\u4e2a<code>ActorModelRayActor<\/code>\u5b9e\u4f8b\u65f6\uff0c\u5b83\u53d1\u73b0\u7531\u4e8e\u6bcf\u4e2a\u5b9e\u4f8b\u53ea\u53600.5\u5757\u5361\uff0c\u56e0\u6b64\u5b8c\u5168\u53ef\u4ee5\u628a\u7b2c2\u4e2a\u5b9e\u4f8b\u63a5\u7740\u7b2c1\u4e2a\u5b9e\u4f8b\u5728\u540c\u4e00\u5f20\u5361\u4e0a\u90e8\u7f72\uff0c\u8fd9\u6837\u5c31\u5bfc\u81f4\u6700\u7ec8\u65e0\u6cd5\u8ba9PPO-Actor\u548cPPO-Ref\u5171\u4eab\u4e00\u5f20\u5361<\/li><li>\u5047\u8bbe\u8bbe\u7f6e0.75\uff0c\u5f53\u6211\u4eec\u5728\u5355\u5361\u4e0a\u90e8\u7f72\u5b8c1\u4e2a<code>ActorModelRayActor<\/code>\u5b9e\u4f8b\u540e\uff0cray\u53d1\u73b0\u5355\u5361\u5269\u4e0b\u7684\u7a7a\u95f4\u4e0d\u8db3\u4ee5\u90e8\u7f72\u7b2c2\u4e2a<code>ActorModelRayActor<\/code>\u5b9e\u4f8b\uff0c\u6240\u4ee5\u5c31\u4f1a\u628a\u7b2c\u4e8c\u4e2a\u5b9e\u4f8b\u90e8\u7f72\u5230\u522b\u7684\u5361\u4e0a\uff0c\u8fd9\u6837\u6700\u7ec8\u5b9e\u73b0PPO-Actor\u548cPPO-Ref\u5171\u4eab\u4e00\u5f20\u5361<\/li><li>\u6240\u4ee5\uff0c\u8fd9\u4e2a\u8bbe\u7f6e\u662f\u4e3a\u4e86\u8fbe\u5230\u4e0d\u540c\u7c7b\u578b\u6a21\u578b\u7684\u5b9e\u4f8b\u5171\u4eab\u4e00\u5f20\u5361\u7684\u76ee\u7684\uff0c\u800c\u5e76\u975e\u771f\u6b63\u6307\u6a21\u578b\u5b9e\u9645\u5360\u636e\u7684\u5355\u5361\u663e\u5b58\u7a7a\u95f4\u3002<\/li><\/ul><\/li><li>\u6700\u540e\uff0c\u5728\u8fd9\u4e00\u6b65\u4e2d\uff0c\u6211\u4eec\u5bf9\u5168\u90e8<code>ActorModelRayActor<\/code>\u5171\u521b\u5efa8\u4e2aworker\u8fdb\u7a0b\uff0c\u5bf9\u5168\u90e8<code>RefenreceModelRayActor<\/code>\u5171\u521b\u5efa8\u4e2aworker\u8fdb\u7a0b\uff0c\u4e00\u5171\u521b\u5efa16\u4e2a\u5de5\u4f5c\u8fdb\u7a0b\u3002<\/li><\/ul>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>\u76f8\u5173\u4ee3\u7801\u4f9d\u7136\u5728\uff1a<a href=\"https:\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/trainer\/ray\/launcher.py#L143\">https:\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/trainer\/ray\/launcher.py#L143<\/a><\/p><\/blockquote>\n\n\n\n<h3>\u90e8\u7f72vllm_engines\u5b9e\u4f8b<\/h3>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/pic2.zhimg.com\/v2-9d6723b6a49bd58460e4cf2d4973dee5_1440w.jpg\" alt=\"\"\/><\/figure>\n\n\n\n<ul><li><code>create_vllm_engines<\/code>\uff1a\u5728driver\u7aef\uff0c\u6211\u4eec\u901a\u8fc7\u8fd0\u884c\u8be5\u51fd\u6570\u6765\u521b\u5efa<code>vllm_engines<\/code>\uff0c\u8fc7\u7a0b\u76f8\u4f3c\u4e8e4.2\u8282\u4e2d\u7684\u4ecb\u7ecd\uff0c\u4fe1\u606f\u90fd\u5728\u56fe\u4e2d\uff0c\u8fd9\u91cc\u4e0d\u8d58\u8ff0\u3002<\/li><li><code><a href=\"https:\/\/zhida.zhihu.com\/search?content_id=251597453&amp;content_type=Article&amp;match_order=1&amp;q=LLMRayActor&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">LLMRayActor<\/a><\/code>\uff1aworker\u7aefRay-Actor\uff0c\u5b83\u4e3b\u8981\u662f\u628avllm\u5b9e\u4f8b\u8fdb\u884c\u4e86\u4e00\u4e9b\u5305\u88c5\uff0c\u5305\u88c5\u7684\u76ee\u7684\u662f\u4e3a\u4e86\u8ba9ds_rank0\u548call vllm ranks\u95f4\u53ef\u4ee5\u8fdb\u884cPPO-Actor\u7684\u6743\u91cd\u901a\u8baf\uff08\u53c2\u89c12.1\uff083\uff09\uff09<\/li><li>\u5728\u4e0a\u9762\u7684\u4f8b\u5b50\u4e2d\uff0c\u6211\u4eec\u4f1a\u521b\u5efa4\u4e2aworker\u8fdb\u7a0b\uff08\u4e0d\u5360gpu\u8d44\u6e90\uff0c\u53ea\u5360<a href=\"https:\/\/zhida.zhihu.com\/search?content_id=251597453&amp;content_type=Article&amp;match_order=1&amp;q=cpu&amp;zhida_source=entity\" target=\"_blank\" rel=\"noreferrer noopener\">cpu<\/a>\u8d44\u6e90\uff09\uff0c\u7528\u4e8e\u8fd0\u884c\u7ba1\u74064\u4e2avllm_engine\u3002<strong>\u5728\u6bcf\u4e2aworker\u8fdb\u7a0b\u5185\uff0cvllm\u5b9e\u4f8b\u8fd8\u4f1a\u521b\u5efa\u5c5e\u4e8e\u81ea\u5df1\u7684worker\u8fdb\u7a0b\u505a\u5206\u5e03\u5f0f\u8fd0\u884c<\/strong>\uff08\u8fd9\u4e9bworker\u8fdb\u7a0b\u4f1a\u5b9e\u9645\u5360\u636egpu\u8d44\u6e90\uff09\u3002<\/li><\/ul>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>\u76f8\u5173\u4ee3\u7801\u53c2\u89c1\uff1a<br><a rel=\"noreferrer noopener\" href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/trainer\/ray\/vllm_engine.py\" target=\"_blank\">https:\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/trainer\/ray\/vllm_engine.py<\/a><\/p><p><br><a href=\"https:\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/trainer\/ray\/vllm_worker_wrap.py\">https:\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/trainer\/ray\/vllm_worker_wrap.py<\/a><\/p><\/blockquote>\n\n\n\n<h3>ds_rank0\u4e0evllm_ranks\u4e4b\u95f4\u7684\u901a\u8baf<\/h3>\n\n\n\n<p>PPO-Actor\u7684ds_rank0\u9700\u8981\u548call_vllm_ranks\u8fdb\u884c\u901a\u8baf\uff0c\u4f20\u9012\u6700\u65b0\u7684PPO-Actor\u6743\u91cd\uff0c\u4f8b\u5982\u4ee5\u4e0bds_rank0\u8981\u628a\u5b8c\u6574\u7684\u6743\u91cdbroadcast\u7ed916\u4e2avllm_ranks\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/picx.zhimg.com\/v2-6fe86bb652deb850279d513007e31079_1440w.jpg\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u6211\u4eec\u5206\u6210\u5982\u4e0b\u51e0\u6b65\u5b9e\u73b0\u8fd9\u4e2a\u76ee\u6807\uff1a<\/p>\n\n\n\n<h4 id=\"h_12871616401_19\"><strong>\uff081\uff09\u521b\u5efa\u901a\u4fe1\u7ec4<\/strong><\/h4>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/pic2.zhimg.com\/v2-9b59870d9f77273e7d89154b712b8ac5_1440w.jpg\" alt=\"\"\/><\/figure>\n\n\n\n<p><strong>Step1\uff1a<\/strong><\/p>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>\u4ee3\u7801\u6765\u81ea\uff1a<a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/trainer\/ray\/ppo_actor.py%23L58\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/trainer\/ray\/ppo_actor.py#L58<\/a><br><strong>\u8fd9\u6bb5\u4ee3\u7801\u6267\u884c\u5728PPO-Actor0\uff08ds_rank0\uff09\u6240\u5728\u7684worker\u8fdb\u7a0b\u4e2d\u3002\u8fd9\u4e2aworker\u8fdb\u7a0b\u5c06\u901a\u8fc7handler\u5f15\u7528\uff0c\u89e6\u53d1\u8fdc\u7aef\u6bcf\u4e2avllm_engine\u4e0a\u7684init_process_group\u64cd\u4f5c\uff0c\u5e76\u5c06ds_rank0\u7eb3\u5165\u901a\u8baf\u7ec4<\/strong><\/p><\/blockquote>\n\n\n\n<pre class=\"wp-block-code\"><code> # Create torch group with deepspeed rank 0 and all vllm ranks\n        # to update vllm engine's weights after each training stage.\n        #\n        # Say we have 3 vllm engines and eache of them has 4 GPUs,\n        # then the torch group is:\n        # &#91;    0,      1, 2, 3, 4,  5, 6, 7, 8,  9, 10, 11, 12]\n        # |ds rank 0 |  engine-0  |  engine-1  |   engine-2   |\n        #\n        # For ZeRO-1\/2:\n        #   1. Broadcast parameters from rank 0 to all vllm engines\n        # For ZeRO-3:\n        #   1. AllGather paramters to rank 0\n        #   2. Broadcast parameters from rank 0 to all vllm engines\n        if self.vllm_engines is not None and torch.distributed.get_rank() == 0:\n            ...\n            # world_size = num_of_all_vllm_ranks + 1 ds_rank0\n            world_size = vllm_num_engines * vllm_tensor_parallel_size + 1\n            ...\n            # =====================================================================\n            # \u904d\u5386\u6bcf\u4e2avllm_engines\uff0c\u5c06\u5176\u4e0b\u7684\u6bcf\u4e2avllm_rank\u6dfb\u52a0\u8fdb\u901a\u8baf\u7ec4\u4e2d\uff0c\u8fd9\u91cc\u53c8\u5206\u6210\u4e24\u6b65\uff1a\n            # 1. engine.init_process_group.remote(...)\uff1a\n            #    \u9996\u5148\uff0c\u89e6\u53d1\u8fdc\u7a0bvllm_engine\u7684init_process_group\u65b9\u6cd5\n            # 2. \u8fdc\u7a0bvllm_engine\u662f\u4e00\u4e2a\u5305\u88c5\u8fc7\u7684vllm\u5b9e\u4f8b\uff0c\u5b83\u7684init_process_group\n            #    \u65b9\u6cd5\u5c06\u8fdb\u4e00\u6b65\u89e6\u53d1\u8fd9\u4e2avllm\u5b9e\u4f8b\u4e0b\u7684\u5404\u4e2aworker\u8fdb\u7a0b\uff08\u89c14.4\u56fe\u4f8b\uff09\uff0c\n            #    \u6700\u7ec8\u662f\u5728\u8fd9\u4e9bworker\u8fdb\u7a0b\u4e0a\u6267\u884c\u201c\u5c06\u6bcf\u4e2avllm_rank\"\u6dfb\u52a0\u8fdbds_rank0\u901a\u8baf\u7ec4\u7684\u5de5\u4f5c\n            # =====================================================================\n            refs = &#91;\n                engine.init_process_group.remote(\n                    # ds_rank0\u6240\u5728node addr\n                    master_address, \n                    # ds_rank0\u6240\u5728node port\n                    master_port,\n                    # \u8be5vllm_engine\u7684\u7b2c\u4e00\u4e2arank\u5728\"ds_rank0 + all_vllm_ranks\u201c\u4e2d\u7684global_rank\uff0c\n                    # \u8be5\u503c\u5c06\u4f5c\u4e3a\u4e00\u4e2aoffset\uff0c\u4ee5\u8be5\u503c\u4e3a\u8d77\u70b9\uff0c\u53ef\u4ee5\u63a8\u7b97\u51fa\u8be5vllm_engine\u4e2d\u5176\u4f59vllm_rank\u7684global_rank\n                    i * vllm_tensor_parallel_size + 1, \n                    world_size,\n                    \"openrlhf\",\n                    backend=backend,\n                )\n                for i, engine in enumerate(self.vllm_engines)\n            ]\n            # =====================================================================\n            # \u5c06ds_rank0\u6dfb\u52a0\u8fdb\u901a\u8baf\u7ec4\u4e2d\n            # =====================================================================\n            self._model_update_group = init_process_group(\n                backend=backend,\n                init_method=f\"tcp:\/\/{master_address}:{master_port}\",\n                world_size=world_size,\n                rank=0,\n                group_name=\"openrlhf\",\n            )\n            # =====================================================================\n            # \u786e\u4fddall_vllm_ranks\u90fd\u5df2\u6dfb\u52a0\u8fdb\u901a\u8baf\u7ec4\u4e2d\n            # =====================================================================\n            ray.get(refs)<\/code><\/pre>\n\n\n\n<p><strong>Step2:<\/strong><\/p>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>\u4ee3\u7801\u6765\u81ea\uff1a<a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/trainer\/ray\/vllm_worker_wrap.py%23L11\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/trainer\/ray\/vllm_worker_wrap.py#L11<\/a><br><strong>\u8fd9\u6bb5\u4ee3\u7801\u5b9e\u9645\u8fd0\u884c\u5728\u6bcf\u4e2avllm_engine\uff08\u5373\u6bcf\u4e2a\u5305\u88c5\u540e\u7684vllm\u5b9e\u4f8b\uff09\u4e0b\u7684worker\u8fdb\u7a0b\u5185\u3002\u4f8b\u5982tp_size=2\uff0c\u90a3\u4e48\u6bcf\u4e2avllm\u5b9e\u4f8b\u4e0b\u5c31\u67092\u4e2aworker\u8fdb\u7a0b\uff0c\u8fd9\u4e24\u4e2aworker\u8fdb\u7a0b\u90fd\u4f1a\u8fd0\u884c\u8fd9\u6bb5\u4ee3\u7801<\/strong><\/p><\/blockquote>\n\n\n\n<pre class=\"wp-block-code\"><code>class WorkerWrap(Worker):\n    def init_process_group(self, master_address, master_port, rank_offset, world_size, group_name, backend=\"nccl\"):\n        \"\"\"Init torch process group for model weights update\"\"\"\n        assert torch.distributed.is_initialized(), f\"default torch process group must be initialized\"\n        assert group_name != \"\", f\"group name must not be empty\"\n        # =====================================================================\n        # torch.distributed.get_rank(): \u5728\u5f53\u524dvllm_engine\u5185\u90e8\u7684rank\uff0c\n        #                               \u4f8b\u5982\u5728tp_size = 2\u65f6\uff0c\u8fd9\u4e2a\u503c\u8981\u4e48\u662f0\uff0c\u8981\u4e48\u662f1\n        # rank_offset\uff1a\u5f53\u524dvllm_engine\u4e2d\u7684\u7b2c\u4e00\u4e2arank\u5728\u201cds_rank0 + all_vllm_ranks\"\u4e2d\u7684global_rank\n        # \u4e24\u8005\u76f8\u52a0\uff1a\u6700\u7ec8\u5f97\u5230\u5f53\u524drank\u5728\u201cds_rank0 + all_vllm_ranks\"\u4e2d\u7684global_rank\n        # =====================================================================\n        rank = torch.distributed.get_rank() + rank_offset\n        self._model_update_group = init_process_group(\n            backend=backend,\n            init_method=f\"tcp:\/\/{master_address}:{master_port}\",\n            world_size=world_size,\n            rank=rank,\n            group_name=group_name,\n        )\n        ...<\/code><\/pre>\n\n\n\n<h4 id=\"h_12871616401_20\"><strong>\uff082\uff09_broadcast_to_vllm<\/strong><\/h4>\n\n\n\n<p>\u6784\u5efa\u597d\u901a\u8baf\u7ec4\uff0c\u6211\u4eec\u5c31\u53ef\u4ee5\u4eceds_rank0\u5e7f\u64adPPO-Actor\u6743\u91cd\u5230all_vllm_ranks\u4e0a\u4e86\uff0c\u8fd9\u91cc\u4e5f\u5206\u6210\u4e24\u6b65\u3002<\/p>\n\n\n\n<p><strong>Step1\uff1aPPO-Actor ds_rank0\u53d1\u9001\u6743\u91cd<\/strong><\/p>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>\u4ee3\u7801\u5728\uff1a<a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/trainer\/ray\/ppo_actor.py%23L146\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/trainer\/ray\/ppo_actor.py#L146<\/a><br><strong>\u8fd9\u6bb5\u4ee3\u7801\u8fd0\u884c\u5728ds_rank0\u5bf9\u5e94\u7684worker\u8fdb\u7a0b\u4e2d<\/strong><\/p><\/blockquote>\n\n\n\n<pre class=\"wp-block-code\"><code>    def _broadcast_to_vllm(self):\n        # avoid OOM\n        torch.cuda.empty_cache()\n        model = self.actor.model.module\n        count, num_params = 0, len(list(model.named_parameters()))\n        for name, param in model.named_parameters():\n            count += 1  # empty_cache at last param\n\n            # Fire all vllm engines for broadcast\n            if torch.distributed.get_rank() == 0:\n                shape = param.shape if self.strategy.args.zero_stage != 3 else param.ds_shape\n                refs = &#91;\n                    # \u8fdc\u7aefvllm_engine\u7684\u6bcf\u4e2arank\u4e0a\uff0c\u521d\u59cb\u5316\u4e00\u4e2a\u5c3a\u5bf8\u4e3ashape\u7684empty weight\u5f20\u91cf\uff0c\n                    # \u7528\u4e8e\u63a5\u6536\u5e7f\u64ad\u800c\u6765\u7684\u6743\u91cd\n                    engine.update_weight.remote(name, dtype=param.dtype, shape=shape, empty_cache=count == num_params)\n                    for engine in self.vllm_engines\n                ]\n\n            # For ZeRO-3, allgather sharded parameter and broadcast to all vllm engines by rank 0\n            # ds_rank0\u53d1\u51fa\u6743\u91cd\uff08\u89c6\u662f\u5426\u4f7f\u7528zero3\u51b3\u5b9a\u5728\u53d1\u51fa\u524d\u662f\u5426\u8981\u505aall-gather\uff09\n            with deepspeed.zero.GatheredParameters(&#91;param], enabled=self.strategy.args.zero_stage == 3):\n                if torch.distributed.get_rank() == 0:\n                    torch.distributed.broadcast(param.data, 0, group=self._model_update_group)\n                    ray.get(refs) # \u786e\u4fdd\u6240\u6709vllm_ranks\u63a5\u6536\u6743\u91cd\u5b8c\u6bd5<\/code><\/pre>\n\n\n\n<p><strong>Step2: \u5404\u4e2avllm_ranks\u63a5\u6536\u6743\u91cd<\/strong><\/p>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>\u4ee3\u7801\u5728\uff1a<a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/trainer\/ray\/vllm_worker_wrap.py%23L29\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/trainer\/ray\/vllm_worker_wrap.py#L29<\/a><br><strong>\u4ee3\u7801\u8fd0\u884c\u5728\u6bcf\u4e2avllm_engine(\u5373\u6bcf\u4e2a\u5305\u88c5\u540e\u7684vllm\u5b9e\u4f8b)\u4e0b\u7684\u5404\u4e2aworker\u8fdb\u7a0b\u4e2d<\/strong>\u3002\u4f8b\u5982tp_size = 2\uff0c\u90a3\u4e48\u6bcf\u4e2avllm\u5b9e\u4f8b\u4e0b\u67092\u4e2aworker\u8fdb\u7a0b\uff0c\u8fd92\u4e2aworker\u8fdb\u7a0b\u90fd\u4f1a\u8fd0\u884c\u8fd9\u6bb5\u4ee3\u7801\u3002<\/p><\/blockquote>\n\n\n\n<pre class=\"wp-block-code\"><code> def update_weight(self, name, dtype, shape, empty_cache=False):\n        \"\"\"Broadcast weight to all vllm workers from source rank 0 (actor model)\"\"\"\n        if torch.distributed.get_rank() == 0:\n            print(f\"update weight: {name}, dtype: {dtype}, shape: {shape}\")\n\n        assert dtype == self.model_config.dtype, f\"mismatch dtype: src {dtype}, dst {self.model_config.dtype}\"\n        <em># \u521b\u5efa\u540c\u5c3a\u5bf8\u7a7a\u5f20\u91cf\u7528\u4e8e\u63a5\u6536ds_rank0\u5e7f\u64ad\u6765\u7684\u6743\u91cd<\/em>\n        weight = torch.empty(shape, dtype=dtype, device=\"cuda\")\n        <em># \u63a5\u6536\u6743\u91cd<\/em>\n        torch.distributed.broadcast(weight, 0, group=self._model_update_group)\n        <em># \u4f7f\u7528\u63a5\u6536\u5230\u7684\u6743\u91cd\u8fdb\u884c\u66f4\u65b0<\/em>\n        self.model_runner.model.load_weights(weights=&#91;(name, weight)])\n\n        del weight<\/code><\/pre>\n\n\n\n<h3>&nbsp;PPO-Actor\/Critic Training<\/h3>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/pic4.zhimg.com\/v2-033392dd9d6524b08efac6cc0362a30f_1440w.jpg\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u6b63\u59822.1\uff084\uff09\u4e2d\u6240\u8bf4\uff0c\u6211\u4eec\u5c06\u90e8\u7f72\u5728ray\u96c6\u7fa4\u4e0a\u7684PPO-Actor\/Ref\/Critic\/RM\u5b9e\u4f8b\u4eec\u8fdb\u884c\u5206\u7ec4\uff0c\u6bcf\u7ec4\u5206\u522b\u8d1f\u8d23\u4e00\u4efdmicro-batch\u7684\u8bad\u7ec3\uff0c<strong>\u4e0a\u56fe\u523b\u753b\u4e86\u67d0\u4e2a\u7ec4\u5185\u7684\u8bad\u7ec3\u6d41\u7a0b\u3002\u4e00\u7ec4\u5185\u7684\u8bad\u7ec3\u6d41\u7a0b\u53d1\u8d77\u81eaPPO-Actor\u5b9e\u4f8b\uff08fit\u65b9\u6cd5\uff09\uff0c\u6ce8\u610f\u4e0d\u540c\u989c\u8272\u7684worker0\u8868\u793a\u7684\u662f\u4e0d\u540c\u5de5\u4f5c\u8fdb\u7a0b\u3002<\/strong>\u5171\u5206\u6210\u5982\u4e0b\u6b65\u9aa4\u6267\u884c\u3002<\/p>\n\n\n\n<p><br><strong>Step1\uff1a\u53d1\u9001prompts\uff0c\u5e76\u4ecevllm_engine\u4e0a\u6536\u96c6(prompt, response)\u3002<\/strong><\/p>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>\u4ee3\u7801\u53c2\u89c1\uff1a<a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/trainer\/ppo_utils\/experience_maker.py%23L627\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/trainer\/ppo_utils\/experience_maker.py#L627<\/a><\/p><\/blockquote>\n\n\n\n<p><br><br><strong>Step2\uff1a\u4eceRef\/Reward\/Critic\u4e0a\u6536\u96c6\u5e76\u5904\u7406exps<\/strong>\u3002<\/p>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>\u4ee3\u7801\u53c2\u89c1\uff1a<a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/trainer\/ppo_utils\/experience_maker.py%23L492\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/trainer\/ppo_utils\/experience_maker.py#L492<\/a><\/p><\/blockquote>\n\n\n\n<p><br><br><strong>Step3: \u786e\u4fdd\u5c06\u5904\u7406\u540e\u7684exps\u4f20\u9001\u7ed9Critic\uff0c\u5e76\u884c\u6267\u884cActor\u548cCritic\u7684\u8bad\u7ec3<\/strong><\/p>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>\u5c06exps\u4f20\u9001\u7ed9Critic\uff1a<a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/trainer\/ppo_utils\/experience_maker.py%23L470\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/trainer\/ppo_utils\/experience_maker.py#L470<\/a><br>Actor\u8bad\u7ec3\uff1a<a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/trainer\/ray\/ppo_actor.py%23L125\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/trainer\/ray\/ppo_actor.py#L125<\/a><br>Critic\u8bad\u7ec3\uff1a<a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/trainer\/ray\/ppo_actor.py%23L122\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/trainer\/ray\/ppo_actor.py#L122<\/a><br>\u6211\u4eec\u5728Actor\u5b9e\u4f8b\u6240\u5728\u7684worker\u8fdb\u7a0b\u4e0a\u51fa\u53d1Actor\u548cCritic\u7684\u8bad\u7ec3\u3002\u4ee5\u4e0a\u4ee3\u7801\u53ea\u7ed9\u51fa\u4e86\u8bad\u7ec3\u5165\u53e3\uff0c\u66f4\u591a\u7ec6\u8282\u9700\u8981\u987a\u7740\u5165\u53e3\u53bb\u9605\u8bfb\u3002<\/p><\/blockquote>\n\n\n\n<p><br><br><strong>Step4\uff1avllm_engine\u6743\u91cd\u66f4\u65b0\u3002<\/strong><\/p>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>\u4ee3\u7801\u53c2\u89c1\uff1a<a href=\"https:\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/trainer\/ray\/ppo_actor.py#L130\">https:\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/trainer\/ray\/ppo_actor.py#L130<\/a><\/p><\/blockquote>\n\n\n\n<h2>RLHF-PPO\u7b97\u6cd5\u7ec6\u8282<\/h2>\n\n\n\n<p>\u6574\u4e2aRLHF-PPO\u8bad\u7ec3\u8fc7\u7a0b\u5927\u81f4\u5206\u62102\u6b65\uff1a<\/p>\n\n\n\n<ul><li><strong>Stage1\uff1a\u6536\u96c6exps<\/strong><\/li><li><strong>Stage2\uff1a\u4f7f\u7528\u6536\u96c6\u5230\u7684exps\u8ba1\u7b97actor_loss\u548ccritic_loss\uff0c\u7528\u4e8e\u8bad\u7ec3actor\u548ccritic<\/strong><\/li><\/ul>\n\n\n\n<blockquote class=\"wp-block-quote\"><p>\u5728OpenRLHF\u4e2d\u7684\u6838\u5fc3\u4ee3\u7801\u4e3a\uff1a<a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/trainer\/ppo_trainer.py%23L19\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/github.com\/OpenRLHF\/OpenRLHF\/blob\/bb46342711a203c457df2fbca5967fd0549557e0\/openrlhf\/trainer\/ppo_trainer.py#L19<\/a><\/p><\/blockquote>\n\n\n\n<p>\u4e0b\u9762\u6211\u4eec\u5206\u522b\u89e3\u8bfb\u8fd92\u4e2astage\u7684\u8fc7\u7a0b<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/pica.zhimg.com\/v2-34b1acd28f3434d3942dbe2538d60b02_1440w.jpg\" alt=\"\"\/><\/figure>\n\n\n\n<p>&nbsp;Stage2\uff1aTraining<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/pic3.zhimg.com\/v2-79624b7730863a933d1fe81caadd7504_1440w.jpg\" alt=\"\"\/><\/figure>\n\n\n\n<h2>ASR\u5927\u6a21\u578bGRPO\u8bad\u7ec3<\/h2>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/pic3.zhimg.com\/v2-0982cb25fae59c51673947e21521e194_1440w.jpg\" alt=\"\"\/><\/figure>\n\n\n\n<p>\u4ece\u4e00\u4e2a&nbsp;<strong>SFT\uff08\u76d1\u7763\u5fae\u8c03\uff09\u5f97\u5230\u7684 Qwen3-ASR\u8bed\u97f3\u8bc6\u522b\u6a21\u578b<\/strong>&nbsp;\u51fa\u53d1\uff0c\u7528&nbsp;<strong>Ray + DeepSpeed + vLLM \u7f3a\u7701\uff08\u6b64\u5904\u672a\u542f\u7528 vLLM\uff0c\u7531 actor \u81ea\u8eab&nbsp;<code>generate<\/code>\uff09<\/strong>&nbsp;\u7684\u65b9\u5f0f\u505a&nbsp;<strong>PPO\/GRPO \u5f3a\u5316\u5b66\u4e60<\/strong>\uff1a \u5bf9\u6bcf\u6761\u97f3\u9891\u91c7\u6837\u591a\u4e2a\u8f6c\u5199\u7ed3\u679c \u2192 \u7528\u4e00\u4e2a&nbsp;<strong>\u8fdc\u7a0b Python \u5956\u52b1\u51fd\u6570<\/strong>\uff08CER\u3001\u5173\u952e\u8bcd\u3001\u8bed\u8a00\u4e00\u81f4\u6027\u3001\u5e73\u6ed1\u5ea6\u7b49\u591a\u7ef4\u6253\u5206\uff09\u7ed9\u6bcf\u4e2a\u7ed3\u679c\u6253\u5206 \u2192 \u7528&nbsp;<strong>group_norm\uff08GRPO \u7ec4\u5185\u5f52\u4e00\u5316\uff09<\/strong>&nbsp;\u8ba1\u7b97\u4f18\u52bf \u2192 \u7528&nbsp;<strong>PPO \u7b56\u7565\u635f\u5931<\/strong>&nbsp;\u66f4\u65b0 actor\uff08<strong>\u51bb\u7ed3\u97f3\u9891 encoder\uff0c\u53ea\u8bad LLM\/adapter \u90e8\u5206<\/strong>\uff09\u2192 \u5468\u671f\u6027\u4fdd\u5b58 HuggingFace \u6743\u91cd\u3002<\/p>\n\n\n\n<p>\u6574\u4f53\u8c03\u7528\u94fe\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>run_train_v2_from_sft.sh                    # \u542f\u52a8\u811a\u672c\uff1a\u8d77 Ray \u96c6\u7fa4 + \u63d0\u4ea4 job\n  \u2514\u2500 openrlhf.cli.train_ppo_ray             # \u5165\u53e3\uff1a\u89e3\u6790\u53c2\u6570\u3001\u5efa Ray actor \u7ec4\u3001\u9a71\u52a8\u8bad\u7ec3\n       \u2514\u2500 ActorModelRayActor (ray\/ppo_actor.py)   # actor \u8fdb\u7a0b\uff1a\u5efa\u6a21\u578b\u3001\u6570\u636e\u3001\u4f18\u5316\u5668\n            \u2514\u2500 ActorPPOTrainer.fit \u2192 PPOTrainer.fit  # PPO \u4e3b\u5faa\u73af\n                 \u251c\u2500 RemoteExperienceMaker            # \u91c7\u6837 rollout + \u6253\u5206 + \u7b97\u4f18\u52bf\n                 \u2502    \u251c\u2500 actor.generate               # \u751f\u6210\u8f6c\u5199\uff08rollout\uff09\n                 \u2502    \u251c\u2500 reward_func (\u8fdc\u7a0b py)         # \u591a\u7ef4\u5956\u52b1\u6253\u5206\n                 \u2502    \u2514\u2500 group_norm \u4f18\u52bf               # GRPO \u7ec4\u5185\u5f52\u4e00\u5316\n                 \u2514\u2500 PPOTrainer.ppo_train              # \u7528 PolicyLossV3 \u66f4\u65b0 actor<\/code><\/pre>\n","protected":false},"excerpt":{"rendered":"<p>Github\uff1a https:\/\/github.com\/OpenRLHF\/OpenRLHF \u8bf4\u660e\u6587\u6863\uff1ahttps &hellip; <a href=\"http:\/\/139.9.1.231\/index.php\/2026\/07\/11\/openrlhf-asr\/\" class=\"more-link\">\u7ee7\u7eed\u9605\u8bfb<span class=\"screen-reader-text\">\u57fa\u4e8e OpenRLHF \u7684\u5927\u6a21\u578b\u5f3a\u5316\u8bad\u7ec3<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[42,4,38],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/31272"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=31272"}],"version-history":[{"count":33,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/31272\/revisions"}],"predecessor-version":[{"id":31357,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/31272\/revisions\/31357"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=31272"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=31272"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=31272"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}