{"id":1777,"date":"2026-06-30T11:09:13","date_gmt":"2026-06-30T03:09:13","guid":{"rendered":"https:\/\/www.izhuhn.cn\/?p=1777"},"modified":"2026-07-01T12:56:34","modified_gmt":"2026-07-01T04:56:34","slug":"rlinf-recap-ascend-910b","status":"publish","type":"post","link":"https:\/\/www.izhuhn.cn\/index.php\/2026\/06\/30\/rlinf-recap-ascend-910b\/","title":{"rendered":"RLinf RECAP (pi0.5) \u6607\u817e910B\u9002\u914d\u7b80\u8bb0"},"content":{"rendered":"

\u73af\u5883\u6982\u89c8<\/h2>\n\n
\u7ec4\u4ef6<\/th>\u7248\u672c\/\u8bf4\u660e<\/th><\/tr><\/thead>
\u786c\u4ef6<\/td>8 \u00d7 Ascend 910B (aarch64)<\/td><\/tr>
\u57fa\u7840\u955c\u50cf<\/td>`rlinf\/rlinf:agentic-rlinf0.2-libero-cann9.0`<\/td><\/tr>
Python<\/td>3.11\uff08\u81ea\u5efa .venv\uff09<\/td><\/tr>
PyTorch<\/td>2.6.0+cpu + torch-npu 2.6.0<\/td><\/tr>
CANN<\/td>9.0.0<\/td><\/tr>
transformers<\/td>4.57.x<\/td><\/tr>
Ray<\/td>2.x\uff08\u5206\u5e03\u5f0f\u8c03\u5ea6\uff09<\/td><\/tr>
\u6570\u636e\u96c6<\/td>RECAP-Libero10-Task0-48succ-Data<\/td><\/tr>
\u5206\u5e03\u5f0f<\/td>8 \u5361 FSDP\uff08hccl \u901a\u4fe1\u540e\u7aef\uff09<\/td><\/tr><\/tbody><\/table><\/figure>\n\n

\u90e8\u7f72\u6b65\u9aa4<\/h2>\n\n

1. \u62c9\u53d6 Docker \u955c\u50cf<\/h3>\n\n

RLinf \u5b98\u65b9\u63d0\u4f9b\u4e86\u9884\u7f6e CANN 9.0 \u73af\u5883\u7684 Docker \u955c\u50cf\uff0c\u955c\u50cf\u5185\u5305\u542b PyTorch 2.6\u3001torch-npu\u3001Ray \u7b49\u6838\u5fc3\u4f9d\u8d56\uff1a<\/p>\n\n

docker pull rlinf\/rlinf:agentic-rlinf0.2-libero-cann9.0<\/code><\/pre>\n\n
> \u6ce8\u610f\uff1a\u8be5\u955c\u50cf\u5e26 CANN 9.0 \u9a71\u52a8\u6808\uff0c\u9700\u786e\u4fdd\u5bbf\u4e3b\u673a\u5df2\u5b89\u88c5\u5bf9\u5e94\u7248\u672c\u7684 NPU \u9a71\u52a8\u548c\u56fa\u4ef6\u3002<\/p>\n\n
2. \u542f\u52a8\u5bb9\u5668<\/h3>\n\n
docker run -itd \\\n  --name rlinf-recap \\\n  --network host \\\n  --ipc=host \\\n  --privileged \\\n  --device=\/dev\/davinci0 \\\n  --device=\/dev\/davinci1 \\\n  --device=\/dev\/davinci2 \\\n  --device=\/dev\/davinci3 \\\n  --device=\/dev\/davinci4 \\\n  --device=\/dev\/davinci5 \\\n  --device=\/dev\/davinci6 \\\n  --device=\/dev\/davinci7 \\\n  --device=\/dev\/davinci_manager \\\n  --device=\/dev\/hisi_hdc \\\n  -v \/data:\/data \\\n  rlinf\/rlinf:agentic-rlinf0.2-libero-cann9.0 \\\n  \/bin\/bash<\/code><\/pre>\n\n
\u5173\u952e\u53c2\u6570\u8bf4\u660e\uff1a<\/p>\n\n
  • --network host<\/code> + --ipc=host<\/code>\uff1aRay \u5206\u5e03\u5f0f\u901a\u4fe1\u5fc5\u9700<\/li>
  • --privileged<\/code> + \u8bbe\u5907\u6620\u5c04\uff1aNPU \u8bbe\u5907\u76f4\u901a<\/li>
  • -v \/data:\/data<\/code>\uff1a\u6302\u8f7d\u6570\u636e\u76ee\u5f55\uff0c\u5b58\u653e\u6570\u636e\u96c6\u548c\u6a21\u578b\u6743\u91cd<\/li><\/ul>\n\n
    3. \u786c\u4ef6\u53ef\u7528\u6027\u68c0\u67e5<\/h3>\n\n
    \u8fdb\u5165\u5bb9\u5668\u540e\u9a8c\u8bc1 NPU \u8bbe\u5907\uff1a<\/p>\n\n
    docker exec -it rlinf-recap bash\nnpu-smi info<\/code><\/pre>\n\n
    \u9884\u671f\u8f93\u51fa 8 \u5361\u5747\u4e3a OK \u72b6\u6001\uff1a<\/p>\n\n
    NPU-ID  Name   Health  Power(Tot)  Temp(C)\n0       910B   OK      65.0W       42\n1       910B   OK      65.0W       41\n...\uff088 \u5361\u5747 OK\uff09<\/code><\/pre>\n\n
    4. \u521b\u5efa\u865a\u62df\u73af\u5883<\/h3>\n\n
    RLinf \u4f7f\u7528\u81ea\u5efa .venv<\/code> \u800c\u975e\u955c\u50cf\u5185\u7f6e conda \u73af\u5883\uff0c\u901a\u8fc7 install.sh<\/code> \u5b89\u88c5\u4f9d\u8d56\uff1a<\/p>\n\n
    cd \/RLinf\nbash install.sh embodied --model openpi --env libero --platform ascend<\/code><\/pre>\n\n
    \u8be5\u547d\u4ee4\u4f1a\uff1a<\/p>\n\n
    • \u521b\u5efa .venv<\/code>\uff08Python 3.11\uff09<\/li>
    • \u5b89\u88c5 torch 2.6.0+cpu + torch-npu 2.6.0<\/li>
    • \u5b89\u88c5 transformers\u3001diffusers\u3001accelerate\u3001lerobot \u7b49\u4f9d\u8d56<\/li>
    • \u914d\u7f6e hccl \u5206\u5e03\u5f0f\u901a\u4fe1\u540e\u7aef<\/li><\/ul>\n\n
      \u9a8c\u8bc1\u73af\u5883\uff1a<\/p>\n\n
      source .venv\/bin\/activate\npython -c \"import torch; import torch_npu; print('NPU count:', torch.npu.device_count())\"\npython -c \"import ray; print('Ray version:', ray.__version__)\"<\/code><\/pre>\n\n
      5. \u51c6\u5907\u6570\u636e\u96c6<\/h3>\n\n
      \u4ece HuggingFace \u4e0b\u8f7d RECAP-Libero10 \u6570\u636e\u96c6\uff1a<\/p>\n\n
      pip install huggingface-hub\n\n# \u4e0b\u8f7d\u6570\u636e\u96c6\u5230 \/data \u76ee\u5f55\nhuggingface-cli download \\\n  --repo-type dataset RLinf\/RECAP-Libero10-Task0-48succ-Data \\\n  --local-dir \/data\/RECAP-Libero10-Task0-48succ-Data<\/code><\/pre>\n\n
      \u6570\u636e\u96c6\u76ee\u5f55\u7ed3\u6784\uff1a<\/p>\n\n
      \/data\/RECAP-Libero10-Task0-48succ-Data\/\n\u251c\u2500\u2500 libero10_task0_sft\/       # 48 \u6761\u6210\u529f\u8f68\u8ff9\uff088005 \u5e27\uff09\uff0cStep 2 \u8bad\u7ec3\u96c6\n\u2502   \u251c\u2500\u2500 data\/                  # \u56fe\u50cf\u5e27 + \u52a8\u4f5c\u6570\u636e\n\u2502   \u251c\u2500\u2500 meta\/                  # episodes \u5143\u4fe1\u606f\n\u2502   \u2514\u2500\u2500 info.json              # \u6570\u636e\u96c6\u7248\u672c\u4fe1\u606f\n\u251c\u2500\u2500 libero10_task0_train\/      # 4096 \u6761\u6210\u8d25\u6df7\u5408 rollout\n\u251c\u2500\u2500 libero10_task0_eval\/       # \u9a8c\u8bc1\u96c6<\/code><\/pre>\n\n
      6. \u7cfb\u7edf\u4f9d\u8d56\u8865\u5145<\/h3>\n\n
      RLinf \u63a8\u7406\/\u8bc4\u4f30\u9636\u6bb5\u53ef\u80fd\u9700\u8981\u989d\u5916\u4f9d\u8d56\uff1a<\/p>\n\n
      apt update && apt install -y \\\n  libavcodec-dev libavformat-dev libavutil-dev \\\n  libswscale-dev libswresample-dev ffmpeg<\/code><\/pre>\n\n
      7. \u5173\u952e\u73af\u5883\u53d8\u91cf<\/h3>\n\n
      \u8bad\u7ec3\u524d\u8bbe\u7f6e\u4ee5\u4e0b\u73af\u5883\u53d8\u91cf\uff1a<\/p>\n\n
      # NPU \u53ef\u89c1\u8bbe\u5907\uff088 \u5361\u5168\u7528\uff09\nexport CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7\n\n# hccl \u901a\u4fe1\u8d85\u65f6\uff088 \u5361\u5efa\u8fde\u9700\u8981\u66f4\u957f\u65f6\u95f4\uff09\nexport HCCL_CONNECT_TIMEOUT=600\n\n# \u5173\u95ed\u4fe1\u53f7\u5904\u7406\uff08\u907f\u514d Ray fork \u65f6\u4fe1\u53f7\u51b2\u7a81\uff09\nexport TORCHELASTIC_ENABLE_FILE_TIMER=0\n\n# torch-npu \u5e93\u8def\u5f84\uff08\u786e\u4fdd hccl so \u53ef\u627e\u5230\uff09\nexport LD_LIBRARY_PATH=\/RLinf\/.venv\/lib\/python3.11\/site-packages\/torch\/lib:\\\n\/RLinf\/.venv\/lib\/python3.11\/site-packages\/torch_npu\/lib:$LD_LIBRARY_PATH\n\n# tf32 \u5173\u95ed\uff08NPU \u7684 tf32 \u5b9e\u73b0\u7cbe\u5ea6\u4e0e CUDA \u4e0d\u540c\uff09\nexport ALLOW_TF32=0<\/code><\/pre>\n\n
      \n\n
      \u4e00\u3001\u80cc\u666f<\/h2>\n\n
      RLinf \u662f\u4e00\u4e2a\u57fa\u4e8e Ray \u7684\u673a\u5668\u4eba\u5f3a\u5316\u5b66\u4e60\u6846\u67b6\u3002\u5176\u6838\u5fc3\u7b97\u6cd5 RECAP \u901a\u8fc7\u56db\u9636\u6bb5\u6d41\u6c34\u7ebf\u5b9e\u73b0 classifier-free guidance (CFG) \u7b56\u7565\u4f18\u5316\uff1a\u5148\u8ba1\u7b97\u8f68\u8ff9\u56de\u62a5\uff08Step 1\uff09\uff0c\u518d\u8bad\u7ec3\u4ef7\u503c\u6a21\u578b\uff08Step 2\uff09\uff0c\u7136\u540e\u8ba1\u7b97\u4f18\u52bf\u6807\u7b7e\uff08Step 3\uff09\uff0c\u6700\u540e CFG \u8bad\u7ec3\u7b56\u7565\u6a21\u578b\uff08Step 4\uff09\u3002\u6574\u4e2a\u8fc7\u7a0b\u6d89\u53ca VLM backbone\uff08SigLIP + Gemma3\uff09\u30018 \u5361 FSDP \u5206\u5e03\u5f0f\u8bad\u7ec3\uff0c\u76ee\u6807\u662f\u628a\u8fd9\u5957 pipeline \u5b8c\u6574\u8dd1\u5728\u534e\u4e3a\u6607\u817e 910B \u4e0a\u3002<\/p>\n\n
      \u6570\u636e\u96c6\u8def\u5f84\uff1a<\/p>\n\n
      • libero10_task0_sft<\/code>\uff1a48 \u6761\u6210\u529f\u8f68\u8ff9\uff088005 \u5e27\uff09\uff0cStep 2 \u8bad\u7ec3\u96c6<\/li>
      • libero10_task0_train<\/code>\uff1a4096 \u6761\u6210\u8d25\u6df7\u5408\u7684 rollout\uff0cStep 2 eval + Step 1 \u8f93\u5165<\/li>
      • libero10_task0_eval<\/code>\uff1a\u9a8c\u8bc1\u96c6<\/li><\/ul>\n\n
        \u4e8c\u3001Step 1 \u2014 Compute Returns\uff08\u901a\u8fc7\uff09<\/h2>\n\n
        \u7eaf CPU \u8ba1\u7b97\uff0c\u65e0 NPU \u4f9d\u8d56\u3002\u5bf9 8005 \u5e27 SFT \u6570\u636e\u548c 4096 \u6761 rollout \u8f68\u8ff9\u5206\u522b\u8ba1\u7b97\u6298\u6263\u56de\u62a5\uff08G_t = r_t + \u03b3\u00b7G_{t+1}<\/code>\uff0c\u03b3=1.0\uff09\uff0c\u751f\u6210 meta\/returns.parquet<\/code>\u3002\u5168\u7a0b\u987a\u7545\u3002<\/p>\n\n
        python examples\/recap\/process\/compute_returns.py --config-name=compute_returns<\/code><\/pre>\n\n
        \u4e09\u3001Step 2 \u2014 Value Model SFT\uff088 \u5361 FSDP \u8bad\u7ec3\uff09<\/h2>\n\n
        \u8fd9\u662f\u6574\u4e2a\u9002\u914d\u8fc7\u7a0b\u4e2d\u8e29\u5751\u6700\u5bc6\u96c6\u7684\u9636\u6bb5\u3002<\/p>\n\n
        3.1 transfer_to_npu \u7684 Segfault<\/h3>\n\n
        torch_npu.contrib.transfer_to_npu<\/code> \u662f torch-npu \u7684 monkey-patch \u6a21\u5757\uff0c\u4f5c\u7528\u662f\u628a\u6240\u6709 torch.cuda.*<\/code> \u8c03\u7528\u81ea\u52a8\u66ff\u6362\u4e3a torch.npu.*<\/code>\u3002\u4f46\u5b83\u5728\u904d\u5386 torch \u6a21\u5757\u6811\u65f6\u89e6\u53d1\u4e86 transformers \u7684 lazy import\u2192image_transforms\u2192tensorflow\u2192pywrap_tensorflow.py\u2192self_check.py:63<\/code>\uff0cTF \u7684 C \u6269\u5c55\u5728 NumPy 2.x \u4e0b\u76f4\u63a5 segfault\u3002<\/p>\n\n
        \u4fee\u590d<\/strong>\uff1a\u6ce8\u91ca\u6389 train_value.py<\/code> \u91cc\u7684 from torch_npu.contrib import transfer_to_npu<\/code>\uff0c\u7136\u540e pip uninstall tensorflow<\/code>\u3002<\/p>\n\n
        \u4f46\u5b9e\u9645\u4e0a Ray worker \u6587\u4ef6 rlinf\/workers\/sft\/fsdp_value_sft_worker.py<\/code> \u4e5f import \u4e86\u5b83\uff0c\u6240\u4ee5\u6bcf\u4e2a NPU worker fork \u65f6\u8fd8\u662f\u4f1a\u89e6\u53d1\u3002\u5173\u952e\u662f\u8981\u786e\u4fdd\u73af\u5883\u4e2d\u6ca1\u6709 tensorflow\u3002<\/p>\n\n
        3.2 LeRobot Dataset \u79bb\u7ebf\u52a0\u8f7d<\/h3>\n\n
        Step 2 \u7684 config \u91cc\u6307\u5b9a\u4e86 eval_data_paths<\/code> \u6307\u5411\u672c\u5730 libero10_task0_train<\/code>\u3002LeRobot \u7684 LeRobotDataset.__init__<\/code> \u4f1a\u8c03\u7528 get_safe_version()<\/code>\uff0c\u8be5\u65b9\u6cd5\u5185\u90e8 get_repo_versions()<\/code> \u53bb hf-mirror \u67e5\u6570\u636e\u96c6\u7248\u672c\u4fe1\u606f\uff0c\u4f46\u8fd9\u4e2a\u672c\u5730\u6570\u636e\u96c6\u4e0d\u5728 HuggingFace \u4e0a\u2014\u2014404\u3002<\/p>\n\n
        \u4fee\u590d<\/strong>\uff1a\u76f4\u63a5 patch LeRobot \u6e90\u7801 lerobot\/common\/datasets\/utils.py<\/code>\uff1a<\/p>\n\n
        # get_repo_versions \u2014 \u6355\u6349\u5f02\u5e38\u8fd4\u56de\u7a7a\u5217\u8868\ndef get_repo_versions(repo_id: str):\n    try:\n        return _get_repo_versions_impl(repo_id)\n    except Exception:\n        return []\n\n# get_safe_version \u2014 hub_versions \u4e3a\u7a7a\u65f6\u76f4\u63a5\u7528\u672c\u5730\u7248\u672c\ndef get_safe_version(repo_id, version):\n    hub_versions = get_repo_versions(repo_id)\n    if not hub_versions:\n        return f\"v{target_version}\"  # \u7528\u672c\u5730 info.json \u7684 _version_\n    ...<\/code><\/pre>\n\n
        .pth<\/code> \u6587\u4ef6\u548c\u5165\u53e3\u811a\u672c monkey-patch \u90fd\u4e0d\u751f\u6548\uff08Ray workers \u72ec\u7acb fork\uff0c\u4e0d\u7ee7\u627f\u7236\u8fdb\u7a0b monkey-patch\uff09\uff0c\u5fc5\u987b\u6539\u6e90\u6587\u4ef6\u3002<\/p>\n\n
        3.3 Ray Timer Crash<\/h3>\n\n
        \u8bad\u7ec3\u8dd1 4 \u6b65\u6ca1\u95ee\u9898\uff0c\u7b2c 5 \u6b65\u89e6\u53d1 save+eval \u65f6\u5d29\uff1a<\/p>\n\n
        ValueError: Timer 'run_eval' has not been recorded.<\/code><\/pre>\n\n
        \u6839\u56e0\uff1atransfer_to_npu<\/code> monkey-patch \u4e86 torch.distributed.init_process_group<\/code>\uff08\u66ff\u6362 backend \u4e3a hccl\uff09\uff0c\u5e72\u6270\u4e86 Ray \u5185\u90e8 @Worker.timer<\/code> \u88c5\u9970\u5668\u7684\u8ba1\u65f6\u673a\u5236\u3002<\/p>\n\n
        \u4fee\u590d<\/strong>\uff1a\u628a pop_execution_time<\/code> \u7684 crash \u6539\u6210 no-op\uff1a<\/p>\n\n
        # rlinf\/scheduler\/worker\/worker.py:1295\nif tag not in self._timer_metrics:\n    return 0.0  # patched: Ray timer broken by transfer_to_npu<\/code><\/pre>\n\n
        3.4 \u8bad\u7ec3\u6307\u6807<\/h3>\n\n
        8 \u5361 FSDP\uff0cmicro_batch=16\u00d78=128 global_batch\uff0c8005 \u6837\u672c\uff0c\u7ea6 62 \u6b65\/epoch\u3002<\/p>\n\n
        \u524d 4 \u6b65\u8bad\u7ec3\u6307\u6807\uff08\u6a21\u578b\u5c1a\u672a\u6536\u655b\uff0c\u6b63\u5e38\uff09\uff1a<\/p>\n\n
        • loss ~6.0\uff0cMAE ~0.43<\/li>
        • predicted_value_std ~0.01\uff08\u6a21\u578b\u8f93\u51fa\u8fd1\u4e4e\u5e38\u6570\uff0c\u9884\u671f\u4e2d\uff09<\/li>
        • value_spearman ~-0.3 \u9010\u6b65\u6539\u5584\uff08step 4 \u5230 -0.02\uff09<\/li><\/ul>\n\n
          \u56db\u3001\u5df2\u5b8c\u6210\u4fee\u6539\u6e05\u5355<\/h2>\n\n
          \u6587\u4ef6<\/th>\u4fee\u6539<\/th>\u539f\u56e0<\/th><\/tr><\/thead>
          `examples\/recap\/value\/train_value.py:26`<\/td>import transfer_to_npu\uff08\u4fdd\u7559\uff09<\/td>NPU cuda\u2192npu \u6620\u5c04<\/td><\/tr>
          `rlinf\/workers\/sft\/fsdp_value_sft_worker.py:30`<\/td>import transfer_to_npu\uff08\u4fdd\u7559\uff09<\/td>Ray worker \u9700\u8981<\/td><\/tr>
          `rlinf\/scheduler\/worker\/worker.py:1295`<\/td>ValueError \u2192 return 0.0<\/td>Ray timer crash<\/td><\/tr>
          `lerobot\/utils.py:get_repo_versions`<\/td>try\/except \u5305\u88f9<\/td>\u79bb\u7ebf\u6570\u636e\u96c6 404<\/td><\/tr>
          `lerobot\/utils.py:get_safe_version`<\/td>hub_versions \u7a7a\u65f6\u7528\u672c\u5730\u7248<\/td>\u540c\u4e0a<\/td><\/tr>
          `examples\/recap\/value\/config\/libero_sft_value.yaml`<\/td>val_check_interval=50, save_interval=50<\/td>\u63a7\u5236\u9a8c\u8bc1\u9891\u7387<\/td><\/tr><\/tbody><\/table><\/figure>\n\n
          \u5907\u4efd\u76ee\u5f55\uff1a\/RLinf\/backups_adapt\/20260630_101234\/<\/code><\/p>\n\n
          \u4e94\u3001Step 3 \u2014 Compute Advantages\uff088 \u5361\u63a8\u7406\uff09<\/h2>\n\n
          batch_size \u8c03\u4f18<\/h3>\n\n
          Value Model \u63a8\u7406\u663e\u5b58\u53ea\u7528\u5230 7G\uff08910B \u5171 61G\uff09\uff0c\u521d\u59cb batch_size=16<\/code> \u65f6 GPU busytime \u4ec5 30%\uff0c\u6570\u636e\u52a0\u8f7d\uff08avg_fetch=1088ms\uff09\u8fdc\u8d85\u63a8\u7406\uff08avg_infer=456ms\uff09\u3002\u8c03\u5927\u5230 batch_size=128<\/code> \u540e batch \u6570\u4ece 64 \u964d\u5230 8\uff0cavg_infer=2831ms<\/code>\uff0cGPU busy \u63d0\u5347\u81f3 19% \u2014 \u6570\u636e\u52a0\u8f7d\u4ecd\u662f\u74f6\u9888\u4f46\u541e\u5410\u6709\u6539\u5584\u3002<\/p>\n\n
          # compute_advantages_npu.yaml\nadvantage:\n  batch_size: 128  # 16 \u2192 128<\/code><\/pre>\n\n
          \u8fd0\u884c\u7ed3\u679c<\/h3>\n\n
          \u6307\u6807<\/th>batch_size=16<\/th>batch_size=128<\/th><\/tr><\/thead>
          \u6279\u6b21\u6570<\/td>64<\/td>8<\/td><\/tr>
          \u603b\u8017\u65f6<\/td>~104s<\/td>~122s<\/td><\/tr>
          avg_infer\/\u6279<\/td>456ms<\/td>2831ms<\/td><\/tr>
          GPU busy<\/td>30%<\/td>19%<\/td><\/tr>
          samples\/s<\/td>37<\/td>23<\/td><\/tr><\/tbody><\/table><\/figure>\n\n
          > \u6ce8\u610f<\/strong>\uff1abatch_size \u7ffb 8 \u500d\u540e\u603b\u8017\u65f6\u53cd\u800c\u7565\u589e\uff08104s\u2192122s\uff09\uff0c\u74f6\u9888\u5728 avg_fetch=12s<\/code> \u7684\u89c6\u9891\u89e3\u7801 + \u9884\u5904\u7406\uff0c\u4e0d\u5728 NPU \u63a8\u7406\u3002\u8fdb\u4e00\u6b65\u4f18\u5316\u9700\u63d0\u9ad8 dataloader workers \u6216\u9884\u7f13\u5b58\u9884\u5904\u7406\u7ed3\u679c\u3002<\/p>\n\n
          \u7ed3\u679c\u7edf\u8ba1\uff08\u4e0e batch_size \u65e0\u5173\uff09<\/h3>\n\n
          \u6307\u6807<\/th>\u503c<\/th><\/tr><\/thead>
          \u63a8\u7406\u6837\u672c<\/td>8005 samples, 30 episodes<\/td><\/tr>
          V(o_t) mean<\/td>-0.4851<\/td><\/tr>
          Advantage mean<\/td>-0.0033<\/td><\/tr>
          Advantage range<\/td>[-0.0435, 0.5022]<\/td><\/tr>
          Positive rate<\/td>30.0% (2402\/8005)<\/td><\/tr>
          SFT \u6a21\u5f0f<\/td>\u5168\u90e8\u6807\u8bb0\u4e3a True<\/td><\/tr>
          \u8f93\u51fa<\/td>`advantages.parquet`, `mixture_config.yaml`<\/td><\/tr><\/tbody><\/table><\/figure>\n\n
          compute_advantages.py \u4fee\u590d<\/h3>\n\n
          • \u6dfb\u52a0 hccl<\/code> \u540e\u7aef override\uff08\u539f\u4ee3\u7801\u53ea\u68c0\u6d4b nccl\uff09<\/li>
          • \u6dfb\u52a0 torch.npu.set_device()<\/code> \u517c\u5bb9\uff08torch.cuda.set_device<\/code> \u5728 NPU \u4e0a\u4f1a\u62a5\u9519\uff09<\/li><\/ul>\n\n
            \u516d\u3001Step 4 \u2014 CFG Training\uff088 \u5361 FSDP \u8bad\u7ec3\uff09<\/h2>\n\n
            \u4fee\u590d\u6e05\u5355<\/h3>\n\n
            \u6587\u4ef6<\/th>\u4fee\u590d\u5185\u5bb9<\/th><\/tr><\/thead>
            `rlinf\/workers\/sft\/fsdp_cfg_worker.py`<\/td>NPU device \u521d\u59cb\u5316\uff1a`torch.cuda.set_device` \u2192 `torch.npu.set_device`<\/td><\/tr>
            `examples\/recap\/cfg\/config\/libero_cfg_openpi_npu.yaml`<\/td>\u65b0\u5efa NPU \u914d\u7f6e\uff0c\u6307\u5411 `\/data\/pi05_base` + \u5b9e\u9645\u6570\u636e\u96c6\u8def\u5f84<\/td><\/tr>
            `\/data\/pi05_base\/physical-intelligence\/libero\/norm_stats.json`<\/td>\u4ece\u6570\u636e\u96c6\u751f\u6210 LIBERO action\/state \u5f52\u4e00\u5316\u7edf\u8ba1\uff08mean\/std\/q01\/q99\uff09<\/td><\/tr>
            NPU OOM \u4fee\u590d<\/td>`micro_batch_size` 32\u21924, `global_batch_size` 512\u2192256<\/td><\/tr><\/tbody><\/table><\/figure>\n\n
            norm_stats.json \u751f\u6210<\/h3>\n\n
            pi05 \u901a\u7528\u6a21\u578b\u7f3a\u5c11 LIBERO \u73af\u5883\u4e13\u5c5e\u7684 norm_stats.json<\/code>\u3002\u4ece\u6570\u636e\u96c6\u7684 8005 \u6837\u672c\u4e2d\u63d0\u53d6 action(7\u7ef4) \u548c state(8\u7ef4) \u7684\u7edf\u8ba1\u4fe1\u606f\uff0c\u5199\u5165 physical-intelligence\/libero\/norm_stats.json<\/code>\uff1a<\/p>\n\n
            # \u4ece parquet \u6587\u4ef6\u8bfb\u53d6\nfor f in parquet_files:\n    df = pd.read_parquet(f)\n    actions = np.stack(df[\"actions\"].values)  # (N, 7)\n    states = np.stack(df[\"state\"].values)      # (N, 8)\n# \u8ba1\u7b97 mean\/std\/q01\/q99 \u2192 norm_stats.json<\/code><\/pre>\n\n
            \u663e\u5b58\u4e0e batch size<\/h3>\n\n
            pi05 14B \u6a21\u578b\u5728 sharding_strategy: no_shard<\/code> \u4e0b\u6bcf\u5361\u5b8c\u6574\u52a0\u8f7d\u6a21\u578b\u3002\u521d\u59cb micro_batch_size=32<\/code> \u65f6\u524d\u5411\u5df2\u5360 56G\/61G\uff0cbackward \u65f6 OOM\uff1a<\/p>\n\n
            RuntimeError: NPU out of memory. Tried to allocate 970.00 MiB\n  (NPU 0; 60.96 GiB total; 56.37 GiB already allocated)<\/code><\/pre>\n\n
            \u51cf\u5c0f micro_batch_size=4<\/code> \u540e\u987a\u5229\u8fd0\u884c\u3002gradient accumulation \u81ea\u52a8\u8865\u507f\uff08256\/(4\u00d78)=8 \u6b65\u7d2f\u79ef\uff09\u3002<\/p>\n\n
            \u5feb\u901f\u9a8c\u8bc1\u8fd0\u884c\uff0815 \u6b65\uff09<\/h3>\n\n
            # libero_cfg_openpi_npu_validate.yaml\nrunner:\n  max_steps: 15          # \u4ec5 15 \u6b65\u9a8c\u8bc1\n  save_interval: 5       # \u6bcf 5 \u6b65\u4fdd\u5b58\n  val_check_interval: -1 # \u5173\u95ed\u6821\u9a8c\uff08CFG worker \u4e0d\u652f\u6301 eval\uff09\n\nactor:\n  micro_batch_size: 4\n  global_batch_size: 256\n  optim:\n    lr_warmup_steps: 3\n    total_training_steps: 15<\/code><\/pre>\n\n
            \u6307\u6807<\/th>\u503c<\/th><\/tr><\/thead>
            \u901f\u5ea6<\/td>~13-23s\/it<\/td><\/tr>
            Loss<\/td>0.0703 \u2192 0.0433\uff0815 \u6b65\u7a33\u5b9a\u4e0b\u964d\uff09<\/td><\/tr>
            Grad Norm<\/td>0.815 \u2192 0.336\uff08\u6536\u655b\uff09<\/td><\/tr>
            Checkpoints<\/td>global_step_5, _10, _15 \u2705<\/td><\/tr>
            \u663e\u5b58<\/td>~56G\/61G\uff08no_shard \u6a21\u5f0f\uff09<\/td><\/tr><\/tbody><\/table><\/figure>\n\n
            Eval \u9a8c\u8bc1<\/h3>\n\n
            fsdp_cfg_worker.py<\/code> \u786c\u7f16\u7801 self.eval_data_loader = None<\/code>\uff0c\u672a\u5b9e\u73b0 eval \u903b\u8f91\uff08\u4e0d\u50cf fsdp_sft_worker.py<\/code> \u4f1a\u4ece val_data_paths<\/code> \u6784\u5efa eval dataloader\uff09\u3002\u5982\u9700 validation\uff0c\u9700\u5728 CFG worker \u4e2d\u6dfb\u52a0 eval dataloader \u652f\u6301\u3002\u5feb\u901f\u9a8c\u8bc1\u9636\u6bb5\u6682\u65f6\u5173\u95ed\u6821\u9a8c\uff08val_check_interval: -1<\/code>\uff09\u3002<\/p>\n\n
            \u6570\u636e\u96c6\u8bf4\u660e<\/h3>\n\n
            \u6570\u636e\u96c6<\/th>Episodes<\/th>Frames<\/th>\u7528\u9014<\/th><\/tr><\/thead>
            `libero10_task0_sft`<\/td>30<\/td>8,005<\/td>\ud83d\udfe2 \u8bad\u7ec3\u96c6\uff08Step 2\/3\/4\uff09<\/td><\/tr>
            `libero10_task0_train`<\/td>4,096<\/td>1,561,874<\/td>\ud83d\udfe1 \u5907\u7528\u8bad\u7ec3\/\u9a8c\u8bc1\u96c6<\/td><\/tr>
            `libero10_task0_eval`<\/td>64<\/td>25,493<\/td>\ud83d\udd35 \u7eaf\u9a8c\u8bc1\u96c6<\/td><\/tr><\/tbody><\/table><\/figure>\n\n
            \u5df2\u77e5\u95ee\u9898<\/h3>\n\n
            • aten::_sample_dirichlet<\/code> NPU \u4e0d\u652f\u6301\uff0cfallback CPU<\/li>
            • NO_SHARD<\/code> \u7b56\u7565 PyTorch \u5df2\u6807\u8bb0 deprecated\uff0c\u540e\u7eed\u53ef\u6539 shard_grad_op<\/code><\/li>
            • \u901f\u5ea6 ~13s\/it \u504f\u6162\uff08pi05 14B + no_shard + \u975e\u878d\u5408\u7b97\u5b50\uff09<\/li>
            • CFG worker \u4e0d\u652f\u6301 eval dataloader<\/li><\/ul>\n\n
              \u4e03\u3001\u4e0a\u6e38\u56de\u9988\u8ba1\u5212<\/h2>\n\n
              \u9002\u914d\u8fc7\u7a0b\u4e2d\u79ef\u7d2f\u7684\u6539\u52a8\u8ba1\u5212\u56de\u9988 RLinf \u4e3b\u4ed3\u5e93\uff1a<\/p>\n\n
              \u4f18\u5148\u7ea7<\/th>\u5185\u5bb9<\/th>\u8bf4\u660e<\/th><\/tr><\/thead>
              P0<\/td>`torch.npu.set_device()` \u517c\u5bb9<\/td>\u4e09\u4e2a worker + compute_advantages\uff0c\u7528 try\/except \u5305\u88f9\uff0c\u4e0d\u5f71\u54cd CUDA<\/td><\/tr>
              P0<\/td>FSDP device \u4fee\u590d<\/td>`cls_embedding.weight.device` \u2192  buffer device\uff0c\u901a\u7528\u4fee\u590d<\/td><\/tr>
              P1<\/td>`drop_last=True` \u4fee\u590d<\/td>eval \u6b7b\u9501\uff0c\u5bf9 CUDA \u4e5f\u6709\u76ca<\/td><\/tr>
              P1<\/td>HCCL backend \u81ea\u52a8\u68c0\u6d4b<\/td>`compute_advantages.py` \u7684\u5206\u5e03\u5f0f\u540e\u7aef<\/td><\/tr>
              P2<\/td>NPU \u914d\u7f6e\u6a21\u677f<\/td>`compute_advantages_npu.yaml`, `libero_cfg_openpi_npu.yaml`<\/td><\/tr>
              P2<\/td>CFG eval dataloader<\/td>\u8865\u5145 `val_data_paths` \u2192 `eval_data_loader` \u6784\u5efa\u903b\u8f91<\/td><\/tr><\/tbody><\/table><\/figure>\n\n
              \u5b8c\u6574 patch: RLinf-ascend-910b.patch<\/code>\uff0c\u4f7f\u7528\u6307\u5357: RLinf-ascend-910b-guide.md<\/code>\u3002<\/p>\n\n
              \u7b80\u8bb0\u3002<\/p>","protected":false},"excerpt":{"rendered":"
              \u73af\u5883\u6982\u89c8 \u7ec4\u4ef6 \u7248\u672c\/\u8bf4\u660e \u786c\u4ef6 8 \u00d7 Ascend 910B (aarch64) \u57fa\u7840\u955c\u50cf `rlinf\/rlinf:agen …<\/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":[49],"tags":[25,24,46,52,48,45,50,44,43,47,16],"class_list":["post-1777","post","type-post","status-publish","format-standard","hentry","category-49","tag-910b","tag-ascend","tag-cann","tag-docker","tag-fsdp","tag-pi0-5","tag-ray","tag-recap","tag-rlinf","tag-torch-npu","tag-16"],"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\/1777","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=1777"}],"version-history":[{"count":6,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/posts\/1777\/revisions"}],"predecessor-version":[{"id":1783,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/posts\/1777\/revisions\/1783"}],"wp:attachment":[{"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/media?parent=1777"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/categories?post=1777"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/tags?post=1777"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}