{"id":1789,"date":"2026-07-15T12:26:11","date_gmt":"2026-07-15T04:26:11","guid":{"rendered":"https:\/\/www.izhuhn.cn\/?p=1789"},"modified":"2026-07-15T12:26:11","modified_gmt":"2026-07-15T04:26:11","slug":"rlinf-%e9%80%82%e9%85%8d-openpi-sft-%e4%b8%8e-pi0-5-%e6%a8%a1%e5%9e%8b%e5%a4%9a%e5%8d%a1%e8%ae%ad%e7%bb%83%e7%ae%80%e8%ae%b0","status":"publish","type":"post","link":"https:\/\/www.izhuhn.cn\/index.php\/2026\/07\/15\/rlinf-%e9%80%82%e9%85%8d-openpi-sft-%e4%b8%8e-pi0-5-%e6%a8%a1%e5%9e%8b%e5%a4%9a%e5%8d%a1%e8%ae%ad%e7%bb%83%e7%ae%80%e8%ae%b0\/","title":{"rendered":"RLinf \u9002\u914d OpenPI SFT \u4e0e pi0.5 \u6a21\u578b\u591a\u5361\u8bad\u7ec3\u7b80\u8bb0"},"content":{"rendered":"\n<p class=\"wp-block-paragraph\">\u6700\u8fd1\u4f7f\u7528 RLinf \u6846\u67b6\u9002\u914d\u4e86 pi0.5 \u6a21\u578b\u7684 SFT\uff08Supervised Fine-Tuning\uff09\u8bad\u7ec3\uff0c\u5728 Ascend 910B3 NPU \u96c6\u7fa4\u4e0a\u8dd1\u901a\u3002\u8fc7\u7a0b\u4e2d\u9047\u5230\u51e0\u4e2a\u5751\uff0c\u8bb0\u5f55\u4e00\u4e0b\u4fee\u6539\u70b9\u548c\u6392\u67e5\u601d\u8def\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e00\u3001\u80cc\u666f<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">RLinf \u4e4b\u524d\u5df2\u7ecf\u6709 RECAP \u7b97\u6cd5\u7684\u9002\u914d\uff08\u57fa\u4e8e pi0 \u6a21\u578b\uff09\uff0c\u8fd9\u6b21\u9700\u8981\u65b0\u589e pi0.5 \u6a21\u578b\u7684 SFT \u8bad\u7ec3\u652f\u6301\u3002\u6a21\u578b\u6743\u91cd\u5df2\u63d0\u524d\u4e0b\u8f7d\u5230 <code>\/data\/pi05_base<\/code>\uff0c\u6570\u636e\u96c6\u4f7f\u7528 RECAP \u683c\u5f0f\u7684 Libero10 \u5b50\u96c6\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5bb9\u5668\u57fa\u4e8e\u4e4b\u524d RECAP \u9002\u914d\u7684\u955c\u50cf\uff08<code>rlinf_train:cann83-py311-v3<\/code>\uff09\uff0cAscend CANN \u73af\u5883\u548c torch_npu \u5df2\u5c31\u7eea\u3002\u76f4\u63a5\u8dd1 <code>bash examples\/sft\/run_vla_sft.sh libero_sft_openpi<\/code> \u4f1a\u62a5\u4e24\u4e2a\u9519\u8bef\uff0c\u9010\u4e2a\u6392\u67e5\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e8c\u3001\u95ee\u9898\u6392\u67e5\u4e0e\u4fee\u590d<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">2.1 is_lora \u914d\u7f6e\u7f3a\u5931<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u9996\u6b21\u8fd0\u884c\u76f4\u63a5\u62a5 <code>ConfigAttributeError: Key 'is_lora' is not in struct<\/code>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u539f\u56e0\u662f <code>libero_sft_openpi.yaml<\/code> \u4f7f\u7528\u4e86\u5185\u8054\u7684 model \u914d\u7f6e\uff08\u6ca1\u6709\u901a\u8fc7 defaults \u5f15\u5165 <code>model\/pi0_5<\/code> \u6a21\u677f\uff09\uff0c\u5185\u8054\u914d\u7f6e\u91cc\u6ca1\u6709 <code>is_lora<\/code> \u548c <code>lora_rank<\/code> \u5b57\u6bb5\u3002\u800c <code>rlinf\/models\/__init__.py<\/code> \u7b2c 268 \u884c\u76f4\u63a5 <code>cfg.is_lora<\/code> \u8bbf\u95ee\uff0cOmegaConf \u7684 struct \u6a21\u5f0f\u4e0b\u4f1a\u629b\u51fa\u5f02\u5e38\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u4fee\u590d\uff1a<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>YAML \u914d\u7f6e\u4e2d\u6dfb\u52a0 <code>is_lora: False<\/code> \u548c <code>lora_rank: 32<\/code><\/li>\n<li>\u540c\u65f6\u5c06\u4ee3\u7801\u4e2d\u7684 <code>cfg.is_lora<\/code> \u6539\u4e3a <code>cfg.get(\"is_lora\", False)<\/code>\uff0c\u9632\u5fa1\u6027\u5904\u7406<\/li>\n<\/ul>\n\n\n\n<pre class=\"wp-block-code\"><code># examples\/sft\/config\/libero_sft_openpi.yaml \u7247\u6bb5\n  model:\n    model_type: \"openpi\"\n    precision: null\n    model_path: \"\/data\/pi05_base\"\n    is_lora: False\n    lora_rank: 32\n    num_action_chunks: 4\n    add_value_head: False\n    openpi:\n      config_name: \"pi05_libero\"\n      detach_critic_input: True\n<\/code><\/pre>\n\n\n\n<pre class=\"wp-block-code\"><code># rlinf\/models\/__init__.py \u7b2c 268 \u884c\n-    if cfg.is_lora:\n+    if cfg.get(\"is_lora\", False):\n<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">2.2 JAX CUDA \u521d\u59cb\u5316\u5931\u8d25<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u4fee\u590d is_lora \u540e\uff0c\u6570\u636e\u52a0\u8f7d\u9636\u6bb5\u53c8\u62a5\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>RuntimeError: Unable to initialize backend 'cuda': FAILED_PRECONDITION: No visible GPU devices.\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u670d\u52a1\u5668\u53ea\u6709 Ascend NPU\uff0c\u6ca1\u6709 CUDA GPU\u3002JAX \u9ed8\u8ba4\u5c1d\u8bd5\u521d\u59cb\u5316 CUDA \u540e\u7aef\uff0c\u4f46 <code>openpi\/training\/data_loader.py<\/code> \u7b2c 412 \u884c\u4ec5\u4ec5\u8c03\u7528 <code>jax.process_count()<\/code> \u6765\u5224\u65ad\u662f\u5426\u591a\u8fdb\u7a0b\u573a\u666f\u2014\u2014\u8fd9\u4e2a\u8c03\u7528\u5b8c\u5168\u53ef\u4ee5\u7528 CPU \u540e\u7aef\u6ee1\u8db3\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u4fee\u590d\uff1a<\/strong>\u5728\u542f\u52a8\u811a\u672c\u4e2d\u6dfb\u52a0\u73af\u5883\u53d8\u91cf\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code># examples\/sft\/run_vla_sft.sh\nexport JAX_PLATFORMS=cpu\n<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">2.3 Ascend \u81ea\u52a8\u8fc1\u79fb\uff08transfer_to_npu\uff09<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 Ascend NPU \u4e0a\u8dd1 PyTorch \u4ee3\u7801\u65f6\uff0c<code>torch_npu<\/code> \u7684 <code>transfer_to_npu<\/code> \u6a21\u5757\u4f1a\u81ea\u52a8\u5c06 <code>torch.cuda.*<\/code> \u8c03\u7528\u66ff\u6362\u4e3a <code>torch.npu.*<\/code>\u3002\u8fd9\u662f\u7531 <code>fsdp_sft_worker.py<\/code> \u9876\u5c42\u7684 <code>import torch_npu<\/code> \u89e6\u53d1\u7684\u3002\u65e5\u5fd7\u4e2d\u4f1a\u770b\u5230\u5982\u4e0b\u63d0\u793a\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>The torch.Tensor.cuda and torch.nn.Module.cuda are replaced with torch.Tensor.npu and torch.nn.Module.npu now..\nThe backend in torch.distributed.init_process_group set to hccl now..\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6b64\u529f\u80fd\u4e3a Ascend CANN \u5185\u7f6e\uff0c\u65e0\u9700\u989d\u5916\u914d\u7f6e\u3002\u5df2\u7ecf\u5305\u542b\u5728\u5bb9\u5668\u955c\u50cf\u7684 torch_npu \u5305\u4e2d\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e09\u3001\u591a\u5361\u914d\u7f6e<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u9ed8\u8ba4\u914d\u7f6e\u53ea\u4f7f\u7528 1 \u5f20 NPU\uff08<code>actor,env,rollout: 0-0<\/code>\uff09\uff0c\u9700\u8981\u6539\u4e3a 8 \u5361\u5168\u91cf\u8bad\u7ec3\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code># examples\/sft\/config\/libero_sft_openpi.yaml\ncluster:\n  num_nodes: 1\n  component_placement:\n    actor,env,rollout: 0-7    # \u4f7f\u7528\u5168\u90e8 8 \u5f20 Ascend 910B3\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">RLinf \u7684 <code>FlexiblePlacementStrategy<\/code> \u4f1a\u6839\u636e <code>component_placement<\/code> \u81ea\u52a8\u5206\u914d GPU \u8303\u56f4\u3002FSDP \u7684 <code>full_shard<\/code> \u7b56\u7565\u4f1a\u81ea\u52a8\u8de8 8 \u5361\u5206\u7247\u6a21\u578b\u53c2\u6570\u3002\u5355\u5361\u9a8c\u8bc1\u65f6\u6bcf\u6b65\u7ea6 62 \u79d2\uff08pi0.5 \u7ea6 14GB \u53c2\u6570\u91cf\uff09\uff0c8 \u5361\u540e\u6bcf\u6b65\u5185\u5b58\u5360\u7528\u964d\u4f4e\u3001\u541e\u5410\u91cf\u5e94\u6709\u663e\u8457\u63d0\u5347\u3002<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table><thead><tr><th>\u53c2\u6570<\/th><th>\u5355\u5361\uff081\u00d7910B3\uff09<\/th><th>8\u5361\uff088\u00d7910B3\uff09<\/th><\/tr><\/thead><tbody><tr><td>component_placement<\/td><td>actor,env,rollout: 0-0<\/td><td>actor,env,rollout: 0-7<\/td><\/tr><tr><td>micro_batch_size<\/td><td>4<\/td><td>4<\/td><\/tr><tr><td>global_batch_size<\/td><td>128<\/td><td>128<\/td><\/tr><tr><td>FSDP sharding<\/td><td>NO_SHARD\uff08world=1\uff09<\/td><td>FULL_SHARD<\/td><\/tr><tr><td>\u6bcf\u6b65\u65f6\u95f4\uff08\u5b9e\u6d4b\uff09<\/td><td>~62s<\/td><td>\u5f85\u9a8c\u8bc1<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\">\u56db\u3001\u9a8c\u8bc1\u7ed3\u679c<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u4fee\u590d\u540e\u5728\u5355\u5f20 NPU \u4e0a\u9a8c\u8bc1\u901a\u8fc7\uff0c\u8bad\u7ec3\u65e5\u5fd7\u8f93\u51fa\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>Global Step: 1\/30000 [01:01&lt;514:18:22, 61.72s\/it\ntime\/step=61.7, train\/grad_norm=0.533\ntrain\/learning_rate=2.5e-8, train\/loss=0.0633\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">NPU 0 \u8d44\u6e90\u4f7f\u7528\u60c5\u51b5\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>AICore \u5229\u7528\u7387\uff1a33%<\/li>\n<li>HBM \u5360\u7528\uff1a65494\/65536 MB\uff08\u6a21\u578b\u52a0\u8f7d\uff09<\/li>\n<li>\u529f\u8017\uff1a190.6W<\/li>\n<li>NPU \u8fdb\u7a0b\uff1a<code>rayFSDPVlaSft<\/code> \u5360\u7528 62GB<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e94\u3001Patch \u4f7f\u7528<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u6240\u6709\u4fee\u6539\u6c47\u603b\u4e3a\u4e00\u4efd patch \u6587\u4ef6\uff0c\u53ef\u5728\u5176\u4ed6\u8282\u70b9\u76f4\u63a5\u5e94\u7528\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code># \u5728 RLinf \u4ed3\u5e93\u6839\u76ee\u5f55\u5e94\u7528\ncd \/workspace\/RLinf\ngit apply rlinf-openpi-sft-pi05-8card.patch\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">Patch \u5305\u542b\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>examples\/sft\/config\/libero_sft_openpi.yaml<\/code>\uff1a\u6dfb\u52a0 is_lora + 8\u5361 placement<\/li>\n<li><code>examples\/sft\/run_vla_sft.sh<\/code>\uff1a\u6dfb\u52a0 JAX_PLATFORMS=cpu<\/li>\n<li><code>rlinf\/models\/__init__.py<\/code>\uff1acfg.is_lora \u2192 cfg.get \u9632\u5fa1<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u5e94\u7528\u540e\u542f\u52a8\u8bad\u7ec3\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>cd \/workspace\/RLinf\nsource .venv\/bin\/activate\nbash examples\/sft\/run_vla_sft.sh libero_sft_openpi\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b80\u8bb0\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6700\u8fd1\u4f7f\u7528 RLinf \u6846\u67b6\u9002\u914d\u4e86 pi0.5 \u6a21\u578b\u7684 SFT\uff08Supervised Fine-Tuning\uff09\u8bad\u7ec3\uff0c\u5728 Ascend &#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[1],"tags":[54,53,45,56,43,55],"class_list":["post-1789","post","type-post","status-publish","format-standard","hentry","category-uncategorized","tag-ascend-npu","tag-openpi","tag-pi0-5","tag-rl","tag-rlinf","tag-sft"],"jetpack_featured_media_url":"","jetpack-related-posts":[],"jetpack_sharing_enabled":true,"_links":{"self":[{"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/posts\/1789","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/comments?post=1789"}],"version-history":[{"count":1,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/posts\/1789\/revisions"}],"predecessor-version":[{"id":1790,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/posts\/1789\/revisions\/1790"}],"wp:attachment":[{"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/media?parent=1789"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/categories?post=1789"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/tags?post=1789"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}