{"id":1758,"date":"2026-06-25T12:35:30","date_gmt":"2026-06-25T04:35:30","guid":{"rendered":"https:\/\/www.izhuhn.cn\/?p=1758"},"modified":"2026-06-26T14:49:06","modified_gmt":"2026-06-26T06:49:06","slug":"gr00t-n1-7%e5%9c%a8%e6%98%87%e8%85%be910b%e4%b8%8a%e7%9a%84%e8%ae%ad%e7%bb%83%e9%83%a8%e7%bd%b2%e7%ae%80%e8%ae%b0","status":"publish","type":"post","link":"https:\/\/www.izhuhn.cn\/index.php\/2026\/06\/25\/gr00t-n1-7%e5%9c%a8%e6%98%87%e8%85%be910b%e4%b8%8a%e7%9a%84%e8%ae%ad%e7%bb%83%e9%83%a8%e7%bd%b2%e7%ae%80%e8%ae%b0\/","title":{"rendered":"GR00T N1.7\u5728\u6607\u817e910B\u4e0a\u7684\u8bad\u7ec3\u90e8\u7f72\u7b80\u8bb0"},"content":{"rendered":"\n<p class=\"wp-block-paragraph\">\u672c\u6587\u8bb0\u5f55\u5c06 NVIDIA GR00T N1.7\uff083B VLA \u6a21\u578b\uff09\u8fc1\u79fb\u5230\u534e\u4e3a Ascend 910B \u8fdb\u884c\u5206\u5e03\u5f0f\u8bad\u7ec3\u7684\u5168\u8fc7\u7a0b\u3002\u4ece\u57fa\u7840\u955c\u50cf\u9009\u578b\u3001\u73af\u5883\u9002\u914d\u3001mx_driving \u8865\u4e01\u3001\u4ee3\u7801 patch\u3001\u8c03\u4f18\u6392\u9519\u5230\u6700\u7ec8\u8dd1\u901a DeepSpeed ZeRO-2 8 \u5361\u8bad\u7ec3\uff0c\u9010\u4e00\u5c55\u5f00\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e00\u3001\u73af\u5883\u6982\u89c8<\/h2>\n\n\n\n<figure class=\"wp-block-table\"><table><thead><tr><th>\u7ec4\u4ef6<\/th><th>\u7248\u672c<\/th><\/tr><\/thead><tbody><tr><td>\u786c\u4ef6<\/td><td>8 \u00d7 Ascend 910B (aarch64)<\/td><\/tr><tr><td>\u57fa\u7840\u955c\u50cf<\/td><td>drivingsdk:26.0.0-cann9.0.0-910b-ubuntu22.04<\/td><\/tr><tr><td>Python<\/td><td>3.10 (conda \u5185\u7f6e\u73af\u5883 torch2.7.1)<\/td><\/tr><tr><td>PyTorch<\/td><td>2.7.1 + torch_npu 2.7.1.post2<\/td><\/tr><tr><td>CANN<\/td><td>9.0.0<\/td><\/tr><tr><td>transformers<\/td><td>4.57.3<\/td><\/tr><tr><td>diffusers<\/td><td>0.35.1<\/td><\/tr><tr><td>DeepSpeed<\/td><td>0.18.4 (ZeRO-2)<\/td><\/tr><tr><td>\u6570\u636e\u96c6<\/td><td>LIBERO-10 (379 episodes, 95K steps)<\/td><\/tr><tr><td>\u6b65\u65f6<\/td><td>~8s\/step (global_batch_size=640)<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e8c\u3001\u57fa\u7840\u955c\u50cf\u9009\u578b<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">2.1 \u955c\u50cf\u6765\u6e90<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u534e\u4e3a\u4e91 SWR \u63d0\u4f9b\u4e86\u5b98\u65b9\u9884\u7f6e\u955c\u50cf\uff0c\u96c6\u6210\u4e86\u5b8c\u6574\u7684 DrivingSDK \u5957\u4ef6\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>swr.cn-south-1.myhuaweicloud.com\/ascendhub\/drivingsdk:26.0.0-cann9.0.0-910b-ubuntu22.04<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e2a\u955c\u50cf\u7684\u5b9a\u4f4d\u662f\"\u5f00\u7bb1\u5373\u7528\"\u2014\u2014\u5185\u7f6e CANN 9.0.0 \u9a71\u52a8\u6808\u3001conda \u73af\u5883\u7ba1\u7406\u5668\u3001\u4ee5\u53ca\u9884\u88c5\u7684 torch2.7.1 conda \u73af\u5883\uff08\u542b torch_npu 2.7.1.post2\uff09\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2.2 \u4e3a\u4ec0\u4e48\u9009\u8fd9\u4e2a\u955c\u50cf<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6807\u51c6 PyTorch op \u7684 NPU \u540e\u7aef\u652f\u6301\uff08\u5168\u91cf\uff09<\/li>\n\n\n\n<li>torch_npu.optim.NpuFusedAdamW \u7b49\u878d\u5408\u4f18\u5316\u5668<\/li>\n\n\n\n<li>DeepSpeed ZeRO \u7cfb\u5217\u7684 hccl \u901a\u4fe1\u540e\u7aef<\/li>\n\n\n\n<li>Diffusers \/ Transformers \u7684 NPU patcher \u63a5\u53e3<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">2.3 Dockerfile \u53c2\u8003<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u867d\u7136\u57fa\u7840\u955c\u50cf\u81ea\u5e26\u5927\u90e8\u5206\u73af\u5883\uff0c\u4f46 GR00T N1.7 \u9700\u8981\u989d\u5916\u7684 ffmpeg \u89c6\u9891\u89e3\u7801\u5e93\u548c\u82e5\u5e72 Python \u5305\u3002Dockerfile \u6269\u5c55\u5c42\u5982\u4e0b\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>FROM swr.cn-south-1.myhuaweicloud.com\/ascendhub\/drivingsdk:26.0.0-cann9.0.0-910b-ubuntu22.04<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">CMD [\"bash\"]<br>```<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">&gt; \u5b9e\u9645\u90e8\u7f72\u4e2d\u672a\u4f7f\u7528\u5b8c\u6574 Dockerfile\uff0c\u800c\u662f\u5728\u5bb9\u5668\u5185\u9010\u5c42\u9a8c\u8bc1\u548c\u5b89\u88c5\uff0c\u66f4\u7075\u6d3b\u5730\u6392\u9519\u3002Dockerfile \u9002\u5408\u56fa\u5316\u73af\u5883\u540e\u4e00\u952e\u91cd\u5efa\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e09\u3001\u5bb9\u5668\u73af\u5883\u9a8c\u8bc1\u4e0e\u589e\u5f3a<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u955c\u50cf pull \u4e0b\u6765\u540e\uff0c\u7b2c\u4e00\u6b65\u4e0d\u662f\u76f4\u63a5\u8dd1\u8bad\u7ec3\uff0c\u800c\u662f<strong>\u9a8c\u8bc1\u73af\u5883\u548c\u8865\u9f50\u7f3a\u5931\u7ec4\u4ef6<\/strong>\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.1 \u786c\u4ef6\u53ef\u7528\u6027\u68c0\u67e5<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code>npu-smi info<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u68c0\u67e5 8 \u5361\u662f\u5426\u5168\u90e8\u53ef\u89c1\u3001\u9a71\u52a8\u7248\u672c\u662f\u5426\u5339\u914d CANN 9.0.0\u3002\u5173\u952e\u8f93\u51fa\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>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\n\n<h3 class=\"wp-block-heading\">3.2 conda \u73af\u5883\u786e\u8ba4<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u955c\u50cf\u81ea\u5e26 conda \u73af\u5883\u548c\u9884\u6fc0\u6d3b\u7684 <code>torch2.7.1<\/code>\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>conda activate torch2.7.1\npython -c \"import torch; print(torch.__version__)\"\npython -c \"import torch_npu; print(torch_npu.__version__)\"\n# \u8f93\u51fa: 2.7.1, 2.7.1.post2<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u786e\u8ba4 NPU \u8bbe\u5907\u6570\u548c\u901a\u4fe1\u80fd\u529b\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>python -c \"import torch; import torch_npu; print('NPU count:', torch.npu.device_count())\"\npython -c \"\nimport torch\nimport torch.distributed as dist\ndist.init_process_group(backend='hccl', init_method='env:\/\/', world_size=8, rank=0)\nprint('hccl OK')\ndist.destroy_process_group()\n\"<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">3.3 \u7cfb\u7edf\u4f9d\u8d56\u5b89\u88c5<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">GR00T \u7684\u6570\u636e pipeline \u4f7f\u7528 torchvision \u89e3\u7801\u89c6\u9891\uff0c\u4f9d\u8d56\u7cfb\u7edf\u7ea7 ffmpeg \u5e93\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>apt update &amp;&amp; apt install -y \\\n    libavcodec-dev libavformat-dev libavutil-dev \\\n    libswscale-dev libswresample-dev libavfilter-dev \\\n    libavdevice-dev pkg-config ffmpeg<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">3.4 Python \u4f9d\u8d56\u5b89\u88c5<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">GR00T N1.7 \u7684 <code>pyproject.toml<\/code> \u58f0\u660e\u4e86\u5f88\u591a NVIDIA \u4e13\u6709\u4f9d\u8d56\uff08flash-attn\u3001triton\u3001tensorrt\uff09\uff0c\u9700\u8981\u5728 NPU \u4e0a\u88c1\u526a\u6389\u3002\u5b9e\u9645\u5b89\u88c5\u7684\u547d\u4ee4\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>pip install --upgrade pip<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u91cc\u6709\u4e2a\u91cd\u8981\u7684\u6ce8\u610f\u4e8b\u9879\uff1a<strong>\u4e0d\u8981\u76f4\u63a5 <code>pip install -e .<\/code><\/strong>\uff0c\u56e0\u4e3a Isaac-GR00T \u7684 <code>pyproject.toml<\/code> \u4f1a\u5c1d\u8bd5\u5b89\u88c5 flash-attn\uff08CUDA \u4e13\u5c5e\uff09\uff0c\u5728\u6607\u817e\u4e0a\u4f1a\u76f4\u63a5\u62a5\u9519\u3002\u9700\u8981\u5148 patch pyproject.toml \u518d\u5b89\u88c5\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.5 DrivingSDK \u83b7\u53d6<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code># \u514b\u9686 DrivingSDK \u4ed3\u5e93\uff08\u542b mx_driving \u6e90\u7801\uff09\ngit clone https:\/\/gitcode.com\/Ascend\/DrivingSDK.git \/data\/DrivingSDK<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">DrivingSDK \u4e2d\u5305\u542b\u4e86 GR00T-N1.7 \u793a\u4f8b\u76ee\u5f55\u548c mx_driving \u7684 patcher \u6e90\u7801\uff0c\u540e\u9762 mx_driving \u8865\u4e01\u9700\u8981\u4ece\u8fd9\u91cc\u7684\u6e90\u7801\u62f7\u8d1d\u6587\u4ef6\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u56db\u3001mx_driving \u8865\u4e01\uff08\u6700\u5173\u952e\u7684\u4e00\u6b65\uff09<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">4.1 \u95ee\u9898\u5b9a\u4f4d<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">pip \u5b89\u88c5\u7684 mx_driving v1.0.20260421 \u662f\u65e7\u7248 wheel\uff0c\u5b58\u5728\u4e24\u4e2a\u5173\u952e\u7f3a\u9677\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\"><li><strong>API \u7f3a\u5931<\/strong>\uff1a\u7f3a\u5c11 GR00T N1.7 \u9700\u8981\u7684 <code>TransformersNPU<\/code> \u548c <code>DiffusersNPU<\/code> patcher \u7c7b<\/li><li><strong>Patcher API \u8fc7\u65e7<\/strong>\uff1a\u4f7f\u7528\u65e7\u7248\u94fe\u5f0f\u8c03\u7528\uff0c\u4e0d\u652f\u6301\u65b0\u7248 <code>Patcher().add().apply()<\/code> \u7684 fluent \u63a5\u53e3<\/li><\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u76f4\u63a5 import \u65f6\u62a5\u9519\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>from mx_driving.patcher import Patcher, TransformersNPU, DiffusersNPU\n# ImportError: cannot import name 'TransformersNPU'<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">4.2 \u4fee\u590d\u65b9\u6848<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ece DrivingSDK \u6e90\u7801\u4e2d\u62f7\u8d1d\u65b0\u7248 patcher \u6587\u4ef6\uff0c\u5408\u5e76\u65b0\u65e7 API \u540e\u91cd\u5efa wheel\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u4fee\u590d\u6587\u4ef6\u6e05\u5355<\/strong>\uff08\u4ece <code>\/data\/DrivingSDK\/mx_driving\/patcher\/<\/code> \u6e90\u7801\u62f7\u8d1d\uff09\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table><thead><tr><th>\u6587\u4ef6<\/th><th>\u6765\u6e90<\/th><th>\u8bf4\u660e<\/th><\/tr><\/thead><tbody><tr><td>`transformers_patch.py`<\/td><td>DrivingSDK \u6e90\u7801<\/td><td>\u65b0\u589e \u2014 TransformersNPU patcher<\/td><\/tr><tr><td>`diffusers_patch.py`<\/td><td>DrivingSDK \u6e90\u7801<\/td><td>\u65b0\u589e \u2014 DiffusersNPU patcher<\/td><\/tr><tr><td>`patch.py`<\/td><td>DrivingSDK \u6e90\u7801<\/td><td>\u65b0\u7248\u8865\u4e01\u7ba1\u7406\u6846\u67b6<\/td><\/tr><tr><td>`version.py`<\/td><td>DrivingSDK \u6e90\u7801<\/td><td>\u7248\u672c\u8ffd\u8e2a<\/td><\/tr><tr><td>`reporting.py`<\/td><td>DrivingSDK \u6e90\u7801<\/td><td>\u8865\u4e01\u5e94\u7528\u62a5\u544a<\/td><\/tr><tr><td>`patcher_logger.py`<\/td><td>DrivingSDK \u6e90\u7801<\/td><td>\u65e5\u5fd7\u7ec4\u4ef6<\/td><\/tr><tr><td>`patcher.py`<\/td><td>**\u91cd\u5199**<\/td><td>\u65b0\u7248 fluent API\uff1a`Patcher().add().apply()`<\/td><\/tr><tr><td>`__init__.py`<\/td><td>**\u91cd\u5199**<\/td><td>\u6865\u63a5\u65b0\u65e7\u5bfc\u51fa\uff0c\u52a0 `default_patcher_builder = None`<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">4.3 \u5173\u952e\u7ec6\u8282\uff1a_C.so \u6269\u5c55<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u91cd\u5efa wheel \u76f4\u63a5\u8986\u76d6\u5b89\u88c5\u4f1a<strong>\u4e22\u6389<\/strong>\u539f\u6709\u7684 C \u6269\u5c55 <code>_C.cpython-310-aarch64-linux-gnu.so<\/code>\u3002\u8fd9\u662f mx_driving \u7684\u5e95\u5c42\u7b97\u5b50\u7ed1\u5b9a\uff0c\u6ca1\u6709\u5b83\u4f1a\u62a5\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>ModuleNotFoundError: No module named 'mx_driving._C'<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u89e3\u51b3<\/strong>\uff1a\u4ece\u540c\u7248\u672c CANN \u73af\u5883\u7684\u5df2\u6709\u5b89\u88c5\uff08\u5982 bevfusion \u7684 conda \u73af\u5883\uff09\u4e2d\u62f7\u8d1d <code>_C.so<\/code> \u5230 wheel \u5305\u5185\uff0c\u6216\u5b89\u88c5 wheel \u540e\u624b\u52a8\u6062\u590d\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code># \u4ece\u540c\u7ea7 conda \u73af\u5883\u501f _C\ncp \/path\/to\/other\/env\/lib\/python3.10\/site-packages\/mx_driving\/_C*.so \\\n   \/opt\/conda\/envs\/torch2.7.1\/lib\/python3.10\/site-packages\/mx_driving\/<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">4.4 \u6700\u7ec8 wheel \u5305<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code>mx_driving-1.0.20260421.1-py3-none-any.whl<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6253\u5305\u542b 8 \u4e2a patcher \u6587\u4ef6 + \u65e7\u7248 <code>_C.so<\/code>\uff0c\u5b89\u88c5\u540e <code>Patcher(TransformersNPU, DiffusersNPU).add(...).apply()<\/code> \u548c\u65e7\u7248 API \u540c\u65f6\u53ef\u7528\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e94\u3001GR00T \u4ee3\u7801\u9002\u914d\uff085 \u4e2a\u6587\u4ef6 patch\uff09<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u6838\u5fc3\u9002\u914d\u5de5\u4f5c\u96c6\u4e2d\u5728 5 \u4e2a\u6587\u4ef6\uff0c\u8986\u76d6\u5206\u5e03\u5f0f\u540e\u7aef\u3001tensor \u5904\u7406\u3001\u6a21\u578b\u7ed3\u6784\u548c\u4f9d\u8d56\u58f0\u660e\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.1 \u5206\u5e03\u5f0f\u540e\u7aef\u2014\u2014nccl \u2192 nccl + hccl\uff083 \u5904\uff09<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u539f\u4ee3\u7801\u4ec5\u68c0\u67e5 <code>backend == \"nccl\"<\/code>\uff0c\u6607\u817e NPU \u7684\u5206\u5e03\u5f0f\u901a\u4fe1\u540e\u7aef\u4e3a <code>hccl<\/code>\u3002\u9700\u8981\u5728 3 \u4e2a\u6587\u4ef6\u4e2d\u6dfb\u52a0 hccl \u5206\u652f\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table><thead><tr><th>\u6587\u4ef6<\/th><th>\u51fd\u6570<\/th><th>\u4fee\u6539\u5185\u5bb9<\/th><\/tr><\/thead><tbody><tr><td>`dist_utils.py`<\/td><td>`_collective_device()`<\/td><td>\u52a0 `else backend == \"hccl\"` \u2192 return npu device<\/td><\/tr><tr><td>`utils.py`<\/td><td>device \u9009\u62e9<\/td><td>\u52a0 `backend in (\"nccl\", \"hccl\")`<\/td><\/tr><tr><td>`sharded_mixture_dataset.py`<\/td><td>`_get_default_pg_tensor_device()`<\/td><td>\u52a0 hccl \u2192 npu device<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">5.2 pin_memory\u2014\u2014trainer.py<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><code>pin_memory<\/code> \u662f CUDA \u4e13\u6709\u7279\u6027\uff0cNPU \u4e0a\u4f1a\u62a5\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>RuntimeError: cannot pin 'npuFloatType'<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u4fee\u590d<\/strong>\uff1a\u786c\u7f16\u7801 <code>pin_memory=False<\/code>\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>pin_memory_device = \"npu\"\npin_memory = False  # NPU \u4e0d\u652f\u6301 pin_memory\uff0c\u786c\u7f16\u7801\u5173\u6389<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">5.3 get_device_capability \u5bb9\u9519\u2014\u2014dit.py<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><code>_is_spark_sm121()<\/code> \u5728\u6bcf\u6b21\u524d\u5411\u4f20\u64ad\u65f6\u90fd\u8c03\u7528 <code>torch.cuda.get_device_capability()<\/code>\uff0cNPU \u4e0a\u8fd4\u56de <code>None<\/code> \u5bfc\u81f4\u89e3\u5305\u5f02\u5e38\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>try:\n    with warnings.catch_warnings():\n        warnings.filterwarnings(\"ignore\", \".*get_device_capability.*\")\n        major, minor = torch.cuda.get_device_capability()\n        return (major, minor) == (12, 1)\nexcept (TypeError, RuntimeError):\n    return False<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">5.4 Qwen3-VL \u9aa8\u5e72\u2014\u2014qwen3_backbone.py<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">GR00T N1.7 \u4f7f\u7528 Qwen3-VL \u4f5c\u4e3a\u89c6\u89c9-\u8bed\u8a00\u9aa8\u5e72\u3002\u539f\u4ee3\u7801\u5728\u521d\u59cb\u5316\u65f6\u5c1d\u8bd5 import flash_attn\uff0cNPU \u4e0a\u6ca1\u6709\u8fd9\u4e2a\u5e93\u3002\u52a0 <code>torch_npu<\/code> \u4f18\u5148\u68c0\u6d4b\uff0c\u786e\u4fdd\u5728\u65e0 flash_attn \u65f6\u4ecd\u8bbe\u7f6e\u6b63\u786e\u7684 attention \u5b9e\u73b0\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code># \u4f18\u5148\u68c0\u6d4b torch_npu\ntry:\n    import torch_npu\n    _has_npu = True\nexcept ImportError:\n    _has_npu = False<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">if _has_npu:<br>    # NPU: \u4f7f\u7528 transformers \u81ea\u5e26\u7684 flash_attention_2 \u5b9e\u73b0<br>    attn_implementation = \"flash_attention_2\"<br>else:<br>    # CUDA: \u539f\u6709\u903b\u8f91<br>    ...<br>```<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.5 pyproject.toml \u4f9d\u8d56\u88c1\u526a<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u539f\u58f0\u660e\u7684 NVIDIA \u4e13\u5c5e\u4f9d\u8d56\u5728\u6607\u817e\u4e0a\u4e0d\u53ef\u7528\uff0c\u9700\u5220\u9664\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code># \u5220\u9664\u4ee5\u4e0b\u884c\n# flash-attn&gt;=2.6.0,\n# triton&gt;=3.0.0,\n# tensorrt&gt;=10.0.0,\n# uv&gt;=0.5.0,<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">DeepSpeed \u548c NCCL \u76f8\u5173\u4f9d\u8d56\u4fdd\u7559\u2014\u2014DeepSpeed \u901a\u8fc7 hccl \u540e\u7aef\u517c\u5bb9 NPU\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.6 \u6700\u7ec8 patch \u6587\u4ef6<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5408\u5e76\u4e0a\u8ff0\u6539\u52a8\u4e3a\u4e00\u4e2a patch \u6587\u4ef6\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>gr00t_n1d7.patch  \u2014 10 \u4e2a\u6587\u4ef6\uff0c+296\/-101 \u884c<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 Isaac-GR00T \u4ed3\u5e93\u4e0a\u5e94\u7528\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>git clone https:\/\/github.com\/NVIDIA\/Isaac-GR00T.git\ncd Isaac-GR00T\ngit checkout e5749287857afd97b78f1147166137de29746392\ncp \/path\/to\/gr00t_n1d7.patch .\/\ngit apply gr00t_n1d7.patch\npip install -e .<\/code><\/pre>\n\n\n\n<h2 class=\"wp-block-heading\">\u516d\u3001\u8bad\u7ec3\u811a\u672c\u4e0e\u53c2\u6570\u8c03\u4f18<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">6.1 \u8bad\u7ec3\u811a\u672c<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e24\u4e2a\u8bad\u7ec3\u811a\u672c\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong><code>train_8p.sh<\/code><\/strong>\u2014\u2014\u5b8c\u6574\u8bad\u7ec3\uff0820000 \u6b65\uff09\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>#!\/bin\/bash\nexport CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7\nexport HCCL_CONNECT_TIMEOUT=600\nexport LD_LIBRARY_PATH=\/opt\/conda\/envs\/torch2.7.1\/lib\/python3.10\/site-packages\/torch\/lib:\\\n\/opt\/conda\/envs\/torch2.7.1\/lib\/python3.10\/site-packages\/torch_npu\/lib:$LD_LIBRARY_PATH<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">accelerate launch \\<br>    --num_processes 8 \\<br>    --num_machines 1 \\<br>    --machine_rank 0 \\<br>    --main_process_port 29500 \\<br>    --mixed_precision bf16 \\<br>    gr00t\/experiment\/launch_finetune.py \\<br>    --dataset libero \\<br>    --config_path configs\/train_config.yaml \\<br>    --output_dir results\/Checkpoints\/gr00t_n1.7_8p \\<br>    --num_train_steps 20000<br>```<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u5173\u952e\u53c2\u6570\u8bf4\u660e\uff1a<\/strong><\/p>\n\n\n\n<figure class=\"wp-block-table\"><table><thead><tr><th>\u53c2\u6570<\/th><th>\u503c<\/th><th>\u539f\u56e0<\/th><\/tr><\/thead><tbody><tr><td>------<\/td><td>-----<\/td><td>------<\/td><\/tr><tr><td>`dataloader_num_workers`<\/td><td>0<\/td><td>NPU \u4e0d\u652f\u6301 fork \u5b50\u8fdb\u7a0b\u521b\u5efa DataLoader<\/td><\/tr><tr><td>`embodiment_tag`<\/td><td>LIBERO_PANDA<\/td><td>N1.7 tag \u4f53\u7cfb\u53d8\u4e86\uff0c\u9884\u6ce8\u518c\u6807\u7b7e\u81ea\u5e26 modality config<\/td><\/tr><tr><td>`mixed_precision`<\/td><td>bf16<\/td><td>NPU \u539f\u751f\u652f\u6301 bf16\uff0c\u65e0\u9700 fp16 \u8f6c\u6362<\/td><\/tr><tr><td>`optimizer`<\/td><td>adamw_torch<\/td><td>NPU \u517c\u5bb9\u6807\u51c6 PyTorch AdamW\uff0c\u65e0\u9700\u81ea\u5b9a\u4e49<\/td><\/tr><tr><td>`tf32`<\/td><td>False<\/td><td>NPU \u7684 tf32 \u5b9e\u73b0\u7cbe\u5ea6\u4e0e CUDA \u4e0d\u540c\uff0c\u663e\u5f0f\u5173\u95ed<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">6.2 DeepSpeed ZeRO-2 \u914d\u7f6e<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code>{\n    \"zero_optimization\": {\n        \"stage\": 2,\n        \"allgather_partitions\": true,\n        \"reduce_scatter\": true,\n        \"contiguous_gradients\": true,\n        \"overlap_comm\": true\n    },\n    \"bf16\": {\n        \"enabled\": true\n    },\n    \"optimizer\": {\n        \"type\": \"AdamW\",\n        \"params\": {\n            \"lr\": 1e-4,\n            \"weight_decay\": 0.01\n        }\n    },\n    \"communication_data_type\": \"bf16\",\n    \"gradient_accumulation_steps\": 8\n}<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>hccl \u901a\u4fe1\u540e\u7aef\u9002\u914d<\/strong>\uff1aDeepSpeed \u5728 NPU \u4e0a\u4f1a\u81ea\u52a8\u68c0\u6d4b hccl \u540e\u7aef\uff0c\u5173\u952e\u73af\u5883\u53d8\u91cf\u786e\u4fdd\u901a\u4fe1\u6b63\u5e38\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>export HCCL_CONNECT_TIMEOUT=600       # 8 \u5361\u5efa\u8fde\u8d85\u65f6\nexport HCCL_ALGO=\"ring\"               # \u901a\u4fe1\u7b97\u6cd5\nexport HCCL_BUFFER_SIZE=200           # \u901a\u4fe1\u7f13\u51b2\u533a<\/code><\/pre>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e03\u3001\u6392\u9519\u8bb0\u5f55<\/h2>\n\n\n\n<figure class=\"wp-block-table\"><table><thead><tr><th>#<\/th><th>\u9519\u8bef<\/th><th>\u6839\u56e0<\/th><th>\u4fee\u590d<\/th><\/tr><\/thead><tbody><tr><td>1<\/td><td>`ImportError: TransformersNPU`<\/td><td>mx_driving pip \u5305\u7248\u672c\u8fc7\u65e7<\/td><td>\u4ece\u6e90\u7801\u8865 8 \u4e2a patcher \u6587\u4ef6\uff0c\u6253 wheel<\/td><\/tr><tr><td>2<\/td><td>`Two accelerators: npu and npu`<\/td><td>mx_driving._C \u548c torch_npu \u53cc\u91cd\u6ce8\u518c<\/td><td>`transfer_to_npu()` \u5728 import torch_npu \u524d\u751f\u6548<\/td><\/tr><tr><td>3<\/td><td>`Unknown tag: fourier_gr1`<\/td><td>N1.7 tag \u4f53\u7cfb\u53d8\u4e86<\/td><td>\u6539\u7528 `LIBERO_PANDA`<\/td><\/tr><tr><td>4<\/td><td>`Cosmos-Reason2-2B 403`<\/td><td>Gated HF \u6a21\u578b\u9700\u6388\u6743<\/td><td>\u7533\u8bf7\u6388\u6743\u540e\u4e0b\u8f7d\u5230\u672c\u5730<\/td><\/tr><tr><td>5<\/td><td>`No backend type for cpu tensor`<\/td><td>hccl \u4e0d\u8ba4 CPU tensor<\/td><td>3 \u5904 `nccl` \u2192 `nccl + hccl` \u9002\u914d<\/td><\/tr><tr><td>6<\/td><td>`Cannot re-init NPU in forked subprocess`<\/td><td>DataLoader fork \u9650\u5236<\/td><td>`num_workers=0`<\/td><\/tr><tr><td>7<\/td><td>`cannot pin 'npuFloatType'`<\/td><td>pin_memory \u662f CUDA \u7279\u6709<\/td><td>trainer.py \u786c\u7f16\u7801 False<\/td><\/tr><tr><td>8<\/td><td>`get_device_capability returns None`<\/td><td>NPU API \u672a\u5b9e\u73b0\u8be5\u63a5\u53e3<\/td><td>`catch_warnings` + `try-except` \u5bb9\u9519<\/td><\/tr><tr><td>9<\/td><td>`ModuleNotFoundError: _C`<\/td><td>wheel \u8986\u76d6\u5b89\u88c5\u4e22 C \u6269\u5c55<\/td><td>\u4ece\u540c\u7ea7\u73af\u5883\u62f7\u8d1d `_C.so` \u6062\u590d<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">\u6392\u9519\u8981\u70b9\u8be6\u89e3<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u9519\u8bef 2\uff08\u53cc\u91cd\u6ce8\u518c\uff09<\/strong>\u662f\u6700\u9690\u6666\u7684\u5751\u3002mx_driving \u7684 <code>__init__.py<\/code> \u5728 import \u65f6\u4f1a\u81ea\u52a8\u8c03\u7528 <code>transfer_to_npu()<\/code>\uff0c\u4f46\u5982\u679c\u6b64\u65f6 <code>torch_npu<\/code> \u5c1a\u672a import\uff0c\u6ce8\u518c\u987a\u5e8f\u4f1a\u9519\u4e71\u3002<strong>\u89e3\u51b3\u987a\u5e8f<\/strong>\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code># \u6b63\u786e\u7684 import \u987a\u5e8f\nimport torch_npu  # \u5fc5\u987b\u5148\u521d\u59cb\u5316 NPU backend\nfrom mx_driving.patcher import Patcher, TransformersNPU, DiffusersNPU<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u9519\u8bef 4\uff08HF Gated Model\uff09<\/strong>\uff1aGR00T N1.7 \u4f9d\u8d56 Cosmos-Reason2-2B\uff08NVIDIA \u53d1\u5e03\u7684 gated \u6a21\u578b\uff09\uff0c\u6607\u817e\u673a\u5668\u4e0a\u6ca1\u6709 HF token \u6388\u6743\u3002\u9700\u8981\u5728 HF \u5b98\u7f51\u7533\u8bf7\u8bbf\u95ee\u6743\u9650\uff0c\u7136\u540e\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>huggingface-cli login --token YOUR_HF_TOKEN\n# \u6216\u5c06\u6a21\u578b\u4e0b\u8f7d\u5230\u672c\u5730\u540e\u6302\u8f7d<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u9519\u8bef 5\uff08CPU tensor\uff09<\/strong>\uff1ahccl \u901a\u4fe1\u8981\u6c42\u6240\u6709 tensor \u5728 NPU \u4e0a\u3002\u90e8\u5206\u7b97\u5b50\uff08\u5982 dataset collate \u4e2d\u7684 pad\uff09\u4f1a\u8fd4\u56de CPU tensor\uff0c\u9700\u8981\u663e\u5f0f <code>.to(\"npu\")<\/code>\u3002\u8fd9 3 \u5904 \"nccl+hccl\" \u7684\u4fee\u6539\u672c\u8d28\u4e0a\u90fd\u662f\u786e\u4fdd\u901a\u4fe1 tensor \u6b63\u786e\u843d\u5728 NPU device \u4e0a\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u516b\u3001\u8bad\u7ec3\u7ed3\u679c<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">8.1 Loss \u66f2\u7ebf<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">8 \u5361 910B\uff0cDeepSpeed ZeRO-2\uff0cLIBERO-10 \u6570\u636e\u96c6\uff0894 shards\uff0c95K steps\uff09\uff0c\u6bcf\u5361\u7ea6 9GB HBM \u5360\u7528\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Loss \u4e0b\u964d\u66f2\u7ebf\u5065\u5eb7\uff0c\u65e0\u9707\u8361\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>Step     Loss\n0        1.28\n500      1.18\n1000     0.90\n2000     0.53\n5000     0.28\n10000    0.16\n20000    0.12<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">Grad Norm \u7a33\u5b9a\u5728 0.3\uff5e2.0 \u533a\u95f4\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">8.2 \u6027\u80fd\u6570\u636e<\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table><thead><tr><th>\u6307\u6807<\/th><th>\u503c<\/th><\/tr><\/thead><tbody><tr><td>\u6b65\u65f6<\/td><td>~8s\/step<\/td><\/tr><tr><td>Global batch size<\/td><td>640 (8 \u5361 \u00d7 16 per_device \u00d7 8 grad_accum)<\/td><\/tr><tr><td>\u6bcf\u5361\u663e\u5b58<\/td><td>~9GB HBM<\/td><\/tr><tr><td>\u5206\u5e03\u5f0f<\/td><td>DeepSpeed ZeRO-2 (hccl)<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">8.3 \u5173\u952e\u68c0\u67e5\u70b9\u9a8c\u8bc1<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li>MX-DRIVING PATCHER \u6b63\u786e\u5e94\u7528\u4e86\u4ee5\u4e0b NPU \u9002\u914d\uff1a<\/li>\n\n\n\n<li>Qwen3RMSNorm\uff08NPU \u7248 LayerNorm \u66ff\u6362\uff09<\/li>\n\n\n\n<li>RoPE\uff08NPU \u7248\u65cb\u8f6c\u4f4d\u7f6e\u7f16\u7801\uff09<\/li>\n\n\n\n<li>Attention scale\uff08NPU bf16 scale\uff09<\/li>\n\n\n\n<li>\u603b\u5171 9 \u4e2a patcher \u751f\u6548<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e5d\u3001\u8f93\u51fa\u7269<\/h2>\n\n\n\n<figure class=\"wp-block-table\"><table><thead><tr><th>\u6587\u4ef6<\/th><th>\u8bf4\u660e<\/th><\/tr><\/thead><tbody><tr><td>`gr00t_n1d7.patch`<\/td><td>\u9002\u914d\u8865\u4e01\uff0810 \u6587\u4ef6\uff0c+296\/-101 \u884c\uff09<\/td><\/tr><tr><td>`mx_driving-1.0.20260421.1-py3-none-any.whl`<\/td><td>mx_driving \u8865\u4e01 wheel<\/td><\/tr><tr><td>`train_8p.sh`<\/td><td>\u5b8c\u6574\u8bad\u7ec3\u811a\u672c\uff0820000 \u6b65\uff09<\/td><\/tr><tr><td>`train_performance_8p.sh`<\/td><td>\u6027\u80fd\u6d4b\u8bd5\u811a\u672c\uff081000 \u6b65 + FPS \u8ba1\u7b97\uff09<\/td><\/tr><tr><td>`deploy.sh`<\/td><td>\u4e00\u952e\u90e8\u7f72\u811a\u672c<\/td><\/tr><tr><td>`pyproject_ascend.toml`<\/td><td>\u9002\u914d\u540e\u4f9d\u8d56\u58f0\u660e\uff08\u53bb NVIDIA \u4e13\u6709\u5305\uff09<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ee5\u4e0a\u6587\u4ef6\u5747\u5728 <code>ascend_patch\/<\/code> \u76ee\u5f55\u4e0b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">---<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6574\u4e2a\u9002\u914d\u8fc7\u7a0b\u53bb\u6389\u955c\u50cf pull \u548c\u73af\u5883\u51c6\u5907\uff0c\u6838\u5fc3 patch \u5de5\u4f5c\u7ea6 2-3 \u5929\u3002\u5176\u4e2d mx_driving patcher \u7684 wheel \u6253\u5305\u548c _C.so \u6062\u590d\u5360\u4e86\u6700\u5927\u5934\uff0c\u5c5e\u4e8e\u5178\u578b\u7684\"\u975e\u5b98\u65b9\u73af\u5883\u9002\u914d\"\u7684\u6210\u672c\u3002\u4ee3\u7801\u5c42\u9762\u7684\u6539\u52a8\u4e0d\u5927\uff08+296\/-101 \u884c\uff09\uff0c\u4f46\u6bcf\u4e2a\u6539\u52a8\u90fd\u662f\u8e29\u5751\u540e\u624d\u786e\u5b9a\u7684\u3002\u6607\u817e\u751f\u6001\u6574\u4f53\u53ef\u7528\u6027\u5c1a\u53ef\uff0c\u4f46\u5728\u975e\u4e3b\u6d41\u6a21\u578b\uff08\u975e LLaMA\/GLM\uff09\u7684\u9002\u914d\u4e2d\uff0cmx_driving patcher \u7684\u7248\u672c\u517c\u5bb9\u6027\u662f\u9700\u8981\u91cd\u70b9\u5173\u6ce8\u7684\u98ce\u9669\u70b9\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u672c\u6587\u8bb0\u5f55\u5c06 NVIDIA GR00T N1.7\uff083B VLA \u6a21\u578b\uff09\u8fc1\u79fb\u5230\u534e\u4e3a Ascend 910B \u8fdb\u884c\u5206\u5e03\u5f0f\u8bad\u7ec3\u7684\u5168\u8fc7\u7a0b\u3002\u4ece &#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":[25,24,18,22,23,15,26],"class_list":["post-1758","post","type-post","status-publish","format-standard","hentry","category-uncategorized","tag-910b","tag-ascend","tag-deepspeed","tag-gr00t","tag-n1-7","tag-npu","tag-26"],"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\/1758","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=1758"}],"version-history":[{"count":6,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/posts\/1758\/revisions"}],"predecessor-version":[{"id":1767,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/posts\/1758\/revisions\/1767"}],"wp:attachment":[{"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/media?parent=1758"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/categories?post=1758"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/tags?post=1758"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}