{"id":42458,"date":"2025-08-21T18:11:23","date_gmt":"2025-08-21T10:11:23","guid":{"rendered":"https:\/\/www.kdjingpai.com\/?p=42458"},"modified":"2025-08-21T18:11:37","modified_gmt":"2025-08-21T10:11:37","slug":"hrmyongyufuzabai","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/en\/hrmyongyufuzabai\/","title":{"rendered":"HRM\uff1a\u7528\u4e8e\u590d\u6742\u63a8\u7406\u7684\u5206\u5c42\u63a8\u7406\u6a21\u578b"},"content":{"rendered":"<p>HRM (Hierarchical Reasoning Model) \u662f\u4e00\u4e2a\u4ec5\u67092700\u4e07\u53c2\u6570\u7684\u5c42\u7ea7\u5f0f\u63a8\u7406\u6a21\u578b\uff0c\u65e8\u5728\u89e3\u51b3\u4eba\u5de5\u667a\u80fd\u9886\u57df\u4e2d\u590d\u6742\u7684\u63a8\u7406\u4efb\u52a1\u3002\u8be5\u6a21\u578b\u7684\u8bbe\u8ba1\u7075\u611f\u6765\u6e90\u4e8e\u4eba\u8111\u7684\u5c42\u7ea7\u5f0f\u3001\u591a\u65f6\u95f4\u5c3a\u5ea6\u7684\u4fe1\u606f\u5904\u7406\u65b9\u5f0f\u3002 \u5b83\u901a\u8fc7\u4e00\u4e2a\u9ad8\u5c42\u6a21\u5757\uff08\u8d1f\u8d23\u7f13\u6162\u3001\u62bd\u8c61\u7684\u89c4\u5212\uff09\u548c\u4e00\u4e2a\u4f4e\u5c42\u6a21\u5757\uff08\u5904\u7406\u5feb\u901f\u3001\u5177\u4f53\u7684\u8ba1\u7b97\uff09\u76f8\u4e92\u4f9d\u8d56\u7684\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u7ed3\u6784\uff0c\u5728\u6ca1\u6709\u660e\u786e\u4e2d\u95f4\u8fc7\u7a0b\u76d1\u7763\u7684\u60c5\u51b5\u4e0b\uff0c\u901a\u8fc7\u5355\u6b21\u524d\u5411\u4f20\u64ad\u6267\u884c\u5e8f\u5217\u63a8\u7406\u4efb\u52a1\u3002 HRM\u65e0\u9700\u9884\u8bad\u7ec3\u6216\u601d\u7ef4\u94fe\uff08CoT\uff09\u6570\u636e\uff0c\u5728\u4ec5\u4f7f\u75281000\u4e2a\u8bad\u7ec3\u6837\u672c\u7684\u60c5\u51b5\u4e0b\uff0c\u5373\u5728\u6570\u72ec\u548c\u5927\u578b\u8ff7\u5bab\u5bfb\u8def\u7b49\u590d\u6742\u4efb\u52a1\u4e0a\u53d6\u5f97\u4e86\u63a5\u8fd1\u5b8c\u7f8e\u7684\u6027\u80fd\uff0c\u5e76\u5728\u901a\u7528\u4eba\u5de5\u667a\u80fd\u5173\u952e\u57fa\u51c6\u201c\u62bd\u8c61\u4e0e\u63a8\u7406\u8bed\u6599\u5e93\u201d\uff08ARC\uff09\u4e0a\u8d85\u8d8a\u4e86\u8bb8\u591a\u66f4\u5927\u89c4\u6a21\u7684\u6a21\u578b\u3002<\/p>\n<h2>\u529f\u80fd\u5217\u8868<\/h2>\n<ul>\n<li><strong>\u9ad8\u6548\u63a8\u7406<\/strong>\uff1a\u91c7\u7528\u65b0\u9896\u7684\u5faa\u73af\u67b6\u6784\uff0c\u5728\u4fdd\u6301\u8bad\u7ec3\u7a33\u5b9a\u6027\u548c\u6548\u7387\u7684\u540c\u65f6\uff0c\u83b7\u5f97\u4e86\u5de8\u5927\u7684\u8ba1\u7b97\u6df1\u5ea6\u3002<\/li>\n<li><strong>\u65e0\u9700\u9884\u8bad\u7ec3<\/strong>\uff1a\u6a21\u578b\u53ef\u4ee5\u76f4\u63a5\u4ece\u5c11\u91cf\u6837\u672c\u4e2d\u5b66\u4e60\uff0c\u65e0\u9700\u5927\u89c4\u6a21\u7684\u9884\u8bad\u7ec3\u8fc7\u7a0b\u3002<\/li>\n<li><strong>\u4f4e\u6570\u636e\u9700\u6c42<\/strong>\uff1a\u4ec5\u97001000\u4e2a\u8bad\u7ec3\u6837\u672c\u5373\u53ef\u5728\u590d\u6742\u63a8\u7406\u4efb\u52a1\u4e0a\u8fbe\u5230\u9ad8\u6027\u80fd\u3002<\/li>\n<li><strong>\u53cc\u6a21\u5757\u7ed3\u6784<\/strong>\uff1a\u5305\u542b\u4e00\u4e2a\u9ad8\u5c42\u6a21\u5757\u8d1f\u8d23\u62bd\u8c61\u89c4\u5212\uff0c\u4e00\u4e2a\u4f4e\u5c42\u6a21\u5757\u5904\u7406\u5feb\u901f\u8ba1\u7b97\u3002<\/li>\n<li><strong>\u5e7f\u6cdb\u9002\u7528\u6027<\/strong>\uff1a\u5728\u591a\u79cd\u590d\u6742\u7684\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u8868\u73b0\u51fa\u8272\uff0c\u4f8b\u5982\uff1a\n<ul>\n<li>\u9ad8\u96be\u5ea6\u4e5d\u5bab\u683c\u6570\u72ec\uff08Sudoku 9&#215;9 Extreme\uff09<\/li>\n<li>30&#215;30\u8ff7\u5bab\u8def\u5f84\u5bfb\u627e\uff08Maze 30&#215;30 Hard\uff09<\/li>\n<li>\u62bd\u8c61\u4e0e\u63a8\u7406\u8bed\u6599\u5e93\uff08ARC-AGI-2\uff09<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u5f00\u6e90<\/strong>\uff1a\u4ee3\u7801\u5728GitHub\u4e0a\u5f00\u6e90\uff0c\u5e76\u63d0\u4f9b\u9884\u8bad\u7ec3\u597d\u7684\u6a21\u578b\u68c0\u67e5\u70b9\u3002<\/li>\n<\/ul>\n<h2>\u4f7f\u7528\u5e2e\u52a9<\/h2>\n<p>HRM\u7684\u5b89\u88c5\u548c\u4f7f\u7528\u9700\u8981\u7279\u5b9a\u7684\u8f6f\u786c\u4ef6\u73af\u5883\uff0c\u4e3b\u8981\u662f\u652f\u6301CUDA\u7684NVIDIA GPU\u3002\u4ee5\u4e0b\u662f\u8be6\u7ec6\u7684\u5b89\u88c5\u548c\u4f7f\u7528\u6d41\u7a0b\u3002<\/p>\n<h3><strong>\u73af\u5883\u51c6\u5907<\/strong><\/h3>\n<ol>\n<li><strong>\u5b89\u88c5CUDA<\/strong>:<br \/>\nHRM\u4f9d\u8d56CUDA\u6269\u5c55\uff0c\u9996\u5148\u9700\u8981\u786e\u4fdd\u7cfb\u7edf\u4e2d\u5df2\u5b89\u88c5NVIDIA\u9a71\u52a8\u548cCUDA\u5de5\u5177\u5305\u3002\u63a8\u8350\u5b89\u88c5CUDA 12.6\u7248\u672c\u3002<\/p>\n<pre><code># \u4e0b\u8f7dCUDA 12.6\u5b89\u88c5\u7a0b\u5e8f\r\nwget -q --show-progress --progress=bar:force:noscroll -O cuda_installer.run https:\/\/developer.download.nvidia.com\/compute\/cuda\/12.6.3\/local_installers\/cuda_12.6.3_560.35.05_linux.run\r\n# \u4ee5\u9759\u9ed8\u6a21\u5f0f\u5b89\u88c5\r\nsudo sh cuda_installer.run --silent --toolkit --override\r\n# \u8bbe\u7f6eCUDA\u73af\u5883\u53d8\u91cf\r\nexport CUDA_HOME=\/usr\/local\/cuda-12.6\r\n<\/code><\/pre>\n<\/li>\n<li><strong>\u5b89\u88c5PyTorch<\/strong>:<br \/>\n\u6839\u636e\u5df2\u5b89\u88c5\u7684CUDA\u7248\u672c\uff0c\u5b89\u88c5\u5bf9\u5e94\u7684PyTorch\u3002<\/p>\n<pre><code>pip3 install torch torchvision torchaudio --index-url https:\/\/download.pytorch.org\/whl\/cu126\r\n<\/code><\/pre>\n<\/li>\n<li><strong>\u5b89\u88c5\u6784\u5efa\u5de5\u5177<\/strong>:<br \/>\n\u4e3a\u4e86\u7f16\u8bd1CUDA\u6269\u5c55\uff0c\u9700\u8981\u5b89\u88c5\u4e00\u4e9b\u989d\u5916\u7684\u5305\u3002<\/p>\n<pre><code>pip3 install packaging ninja wheel setuptools setuptools-scm\r\n<\/code><\/pre>\n<\/li>\n<li><strong>\u5b89\u88c5FlashAttention<\/strong>:<br \/>\n\u6839\u636e\u4f60\u7684GPU\u578b\u53f7\u5b89\u88c5\u76f8\u5e94\u7248\u672c\u7684FlashAttention\u3002<\/p>\n<ul>\n<li><strong>Hopper\u67b6\u6784GPU (\u4f8b\u5982 H100)<\/strong>:\n<pre><code>git clone git@github.com:Dao-AILab\/flash-attention.git\r\ncd flash-attention\/hopper\r\npython setup.py install\r\n<\/code><\/pre>\n<\/li>\n<li><strong>Ampere\u67b6\u6784\u6216\u66f4\u65e9\u7684GPU (\u4f8b\u5982 RTX 30\u7cfb\u5217, 40\u7cfb\u5217)<\/strong>:\n<pre><code>pip3 install flash-attn\r\n<\/code><\/pre>\n<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u5b89\u88c5\u9879\u76ee\u4f9d\u8d56<\/strong>:<br \/>\n\u514b\u9686HRM\u4ed3\u5e93\u5e76\u5b89\u88c5<code>requirements.txt<\/code>\u4e2d\u7684\u4f9d\u8d56\u3002<\/p>\n<pre><code>git clone https:\/\/github.com\/sapientinc\/HRM.git\r\ncd HRM\r\npip install -r requirements.txt\r\n<\/code><\/pre>\n<\/li>\n<li><strong>W&amp;B\u96c6\u6210\uff08\u53ef\u9009\uff09<\/strong>:<br \/>\n\u9879\u76ee\u4f7f\u7528Weights &amp; Biases\u8fdb\u884c\u5b9e\u9a8c\u8ddf\u8e2a\u3002\u5982\u679c\u4f60\u9700\u8981\u53ef\u89c6\u5316\u8bad\u7ec3\u8fc7\u7a0b\uff0c\u8bf7\u767b\u5f55W&amp;B\u3002<\/p>\n<pre><code>wandb login\r\n<\/code><\/pre>\n<\/li>\n<\/ol>\n<h3><strong>\u5feb\u901f\u4e0a\u624b\uff1a\u8bad\u7ec3\u4e00\u4e2a\u6570\u72ec\u6c42\u89e3AI<\/strong><\/h3>\n<p>\u8fd9\u662f\u4e00\u4e2a\u5728\u5355\u5f20\u6d88\u8d39\u7ea7GPU\uff08\u4f8b\u5982RTX 4070\uff09\u4e0a\u8bad\u7ec3\u9ad8\u96be\u5ea6\u6570\u72ec\u6c42\u89e3\u6a21\u578b\u7684\u793a\u4f8b\u3002<\/p>\n<ol>\n<li><strong>\u6784\u5efa\u6570\u636e\u96c6<\/strong>:<br \/>\n\u9996\u5148\uff0c\u9700\u8981\u4e0b\u8f7d\u5e76\u6784\u5efa\u6570\u72ec\u6570\u636e\u96c6\u3002\u8fd9\u4e2a\u547d\u4ee4\u4f1a\u751f\u62101000\u4e2a\u9ad8\u96be\u5ea6\u6570\u72ec\u6837\u672c\uff0c\u5e76\u8fdb\u884c\u6570\u636e\u589e\u5f3a\u3002<\/p>\n<pre><code>python dataset\/build_sudoku_dataset.py --output-dir data\/sudoku-extreme-1k-aug-1000 --subsample-size 1000 --num-aug 1000\r\n<\/code><\/pre>\n<\/li>\n<li><strong>\u5f00\u59cb\u8bad\u7ec3<\/strong>:<br \/>\n\u4f7f\u7528\u4ee5\u4e0b\u547d\u4ee4\u5728\u5355\u4e2aGPU\u4e0a\u5f00\u59cb\u8bad\u7ec3\u3002<\/p>\n<pre><code>OMP_NUM_THREADS=8 python pretrain.py data_path=data\/sudoku-extreme-1k-aug-1000 epochs=20000 eval_interval=2000 global_batch_size=384 lr=7e-5 puzzle_emb_lr=7e-5 weight_decay=1.0 puzzle_emb_weight_decay=1.0\r\n<\/code><\/pre>\n<p>\u5728RTX 4070\u4e0a\uff0c\u8fd9\u4e2a\u8fc7\u7a0b\u5927\u7ea6\u9700\u898110\u5c0f\u65f6\u3002<\/li>\n<\/ol>\n<h3><strong>\u5927\u89c4\u6a21\u5b9e\u9a8c<\/strong><\/h3>\n<p>\u5bf9\u4e8e\u66f4\u5927\u89c4\u6a21\u7684\u5b9e\u9a8c\uff0c\u4f8b\u5982ARC\u6216\u5b8c\u6574\u7684\u6570\u72ec\u6570\u636e\u96c6\uff0c\u63a8\u8350\u4f7f\u7528\u591aGPU\u73af\u5883\uff08\u4f8b\u59828\u5361\uff09\u3002<\/p>\n<ol>\n<li><strong>\u521d\u59cb\u5316\u5b50\u6a21\u5757<\/strong>:\n<pre><code>git submodule update --init --recursive\r\n<\/code><\/pre>\n<\/li>\n<li><strong>\u51c6\u5907\u6570\u636e\u96c6<\/strong>:<br \/>\n\u6839\u636e\u9700\u8981\u89e3\u51b3\u7684\u95ee\u9898\uff0c\u6784\u5efa\u76f8\u5e94\u7684\u6570\u636e\u96c6\u3002<\/p>\n<ul>\n<li><strong>ARC-2<\/strong>:\n<pre><code>python dataset\/build_arc_dataset.py --dataset-dirs dataset\/raw-data\/ARC-AGI-2\/data --output-dir data\/arc-2-aug-1000\r\n<\/code><\/pre>\n<\/li>\n<li><strong>\u8ff7\u5bab (Maze)<\/strong>:\n<pre><code>python dataset\/build_maze_dataset.py\r\n<\/code><\/pre>\n<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u542f\u52a8\u591aGPU\u8bad\u7ec3<\/strong>:<br \/>\n\u4f7f\u7528<code>torchrun<\/code>\u542f\u52a8\u5206\u5e03\u5f0f\u8bad\u7ec3\u3002\u4ee5\u8bad\u7ec3\u9ad8\u96be\u5ea6\u6570\u72ec\uff081000\u6837\u672c\uff09\u4e3a\u4f8b\uff1a<\/p>\n<pre><code>OMP_NUM_THREADS=8 torchrun --nproc-per-node 8 pretrain.py data_path=data\/sudoku-extreme-1k-aug-1000 epochs=20000 eval_interval=2000 lr=1e-4 puzzle_emb_lr=1e-4 weight_decay=1.0 puzzle_emb_weight_decay=1.0\r\n<\/code><\/pre>\n<p>\u57288\u4e2aGPU\u7684\u73af\u5883\u4e0b\uff0c\u8fd9\u4e2a\u8bad\u7ec3\u8fc7\u7a0b\u5927\u7ea6\u53ea\u9700\u898110\u5206\u949f\u3002<\/li>\n<\/ol>\n<h3><strong>\u8bc4\u4f30\u6a21\u578b<\/strong><\/h3>\n<p>\u4f60\u53ef\u4ee5\u4f7f\u7528\u9884\u8bad\u7ec3\u597d\u7684\u6a21\u578b\u68c0\u67e5\u70b9\uff0c\u6216\u8005\u5bf9\u81ea\u5df1\u8bad\u7ec3\u7684\u6a21\u578b\u8fdb\u884c\u8bc4\u4f30\u3002<\/p>\n<ol>\n<li><strong>\u4e0b\u8f7d\u68c0\u67e5\u70b9<\/strong>:<br \/>\n\u5b98\u65b9\u4ed3\u5e93\u63d0\u4f9b\u4e86\u9488\u5bf9ARC\u3001\u6570\u72ec\u548c\u8ff7\u5bab\u4efb\u52a1\u7684\u9884\u8bad\u7ec3\u6a21\u578b\u3002<\/li>\n<li><strong>\u8fd0\u884c\u8bc4\u4f30\u811a\u672c<\/strong>:<br \/>\n\u901a\u8fc7W&amp;B\u754c\u9762\u67e5\u770b<code>eval\/exact_accuracy<\/code>\u6307\u6807\u3002\u5bf9\u4e8eARC\u4efb\u52a1\uff0c\u9700\u8981\u989d\u5916\u8fd0\u884c\u8bc4\u4f30\u811a\u672c\uff0c\u5e76\u4f7f\u7528Jupyter Notebook\u8fdb\u884c\u7ed3\u679c\u5206\u6790\u3002<\/p>\n<pre><code>OMP_NUM_THREADS=8 torchrun --nproc-per-node 8 evaluate.py checkpoint=&lt;CHECKPOINT_PATH&gt;\r\n<\/code><\/pre>\n<p>\u4e4b\u540e\uff0c\u6253\u5f00<code>arc_eval.ipynb<\/code>\u6765\u6700\u7ec8\u786e\u5b9a\u548c\u68c0\u67e5\u7ed3\u679c\u3002<\/li>\n<\/ol>\n<h3><strong>\u6ce8\u610f\u4e8b\u9879<\/strong><\/h3>\n<ul>\n<li>\u5c0f\u6837\u672c\u5b66\u4e60\u7684\u51c6\u786e\u7387\u901a\u5e38\u4f1a\u6709\u00b12\u4e2a\u70b9\u7684\u6d6e\u52a8\u3002<\/li>\n<li>\u5bf9\u4e8e\u9ad8\u96be\u5ea6\u6570\u72ec\u7684\u5343\u6837\u672c\u6570\u636e\u96c6\uff0c\u8bad\u7ec3\u540e\u671f\u53ef\u80fd\u4f1a\u51fa\u73b0\u8fc7\u62df\u5408\u5bfc\u81f4\u6570\u503c\u4e0d\u7a33\u5b9a\u3002\u5efa\u8bae\u5728\u8bad\u7ec3\u51c6\u786e\u7387\u63a5\u8fd1100%\u65f6\u91c7\u7528\u65e9\u671f\u505c\u6b62\u7b56\u7565\u3002<\/li>\n<\/ul>\n<h2>\u5e94\u7528\u573a\u666f<\/h2>\n<ol>\n<li><strong>\u8ba4\u77e5\u79d1\u5b66\u7814\u7a76<\/strong><br \/>\nHRM\u7684\u8bbe\u8ba1\u6a21\u62df\u4e86\u4eba\u8111\u7684\u5c42\u7ea7\u5316\u4fe1\u606f\u5904\u7406\u673a\u5236\uff0c\u4e3a\u7814\u7a76\u4eba\u7c7b\u7684\u89c4\u5212\u3001\u63a8\u7406\u548c\u89e3\u51b3\u95ee\u9898\u80fd\u529b\u63d0\u4f9b\u4e86\u4e00\u4e2a\u53ef\u8ba1\u7b97\u7684\u6a21\u578b\uff0c\u6709\u52a9\u4e8e\u63a2\u7d22\u901a\u7528\u4eba\u5de5\u667a\u80fd\u7684\u5b9e\u73b0\u8def\u5f84\u3002<\/li>\n<li><strong>\u590d\u6742\u89c4\u5212\u4e0e\u8c03\u5ea6<\/strong><br \/>\n\u5728\u7269\u6d41\u3001\u673a\u5668\u4eba\u8def\u5f84\u89c4\u5212\u3001\u81ea\u52a8\u5316\u751f\u4ea7\u8c03\u5ea6\u7b49\u9886\u57df\uff0cHRM\u80fd\u591f\u4e0d\u4f9d\u8d56\u5927\u91cf\u6570\u636e\uff0c\u5feb\u901f\u627e\u5230\u6700\u4f18\u6216\u63a5\u8fd1\u6700\u4f18\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u4f8b\u5982\u89e3\u51b3\u5927\u578b\u8ff7\u5bab\u5bfb\u8def\u95ee\u9898\u3002<\/li>\n<li><strong>\u6e38\u620fAI<\/strong><br \/>\n\u53ef\u7528\u4e8e\u5f00\u53d1\u80fd\u591f\u89e3\u51b3\u590d\u6742\u7b56\u7565\u6e38\u620f\uff08\u5982\u6570\u72ec\uff09\u7684AI\uff0c\u5176\u9ad8\u6548\u7684\u63a8\u7406\u80fd\u529b\u4f7f\u5176\u80fd\u591f\u5728\u6ca1\u6709\u9884\u8bad\u7ec3\u7684\u60c5\u51b5\u4e0b\u5b66\u4e60\u5e76\u638c\u63e1\u6e38\u620f\u89c4\u5219\u3002<\/li>\n<li><strong>\u672c\u5730\u5316AI\u5e94\u7528<\/strong><br \/>\n\u7531\u4e8e\u6a21\u578b\u53c2\u6570\u91cf\u5c0f\uff0cHRM\u6709\u6f5c\u529b\u88ab\u90e8\u7f72\u5728\u8d44\u6e90\u6709\u9650\u7684\u672c\u5730\u8bbe\u5907\u4e0a\uff0c\u7528\u4e8e\u5bb6\u5ead\u81ea\u52a8\u5316\u7b49\u573a\u666f\uff0c\u6267\u884c\u7279\u5b9a\u7684\u903b\u8f91\u63a8\u7406\u4efb\u52a1\u3002<\/li>\n<\/ol>\n<h2>QA<\/h2>\n<ol>\n<li><strong>HRM\u4e0e\u4f20\u7edf\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u6709\u4f55\u4e0d\u540c\uff1f<\/strong><br \/>\nHRM\u4e0d\u4f9d\u8d56\u81ea\u7136\u8bed\u8a00\u8fdb\u884c\u63a8\u7406\uff0c\u800c\u662f\u901a\u8fc7\u4e24\u4e2a\u4e0d\u540c\u65f6\u95f4\u5c3a\u5ea6\u7684\u5faa\u73af\u6a21\u5757\u8fdb\u884c\u7b26\u53f7\u64cd\u4f5c\u548c\u63a8\u7406\u3002\u5b83\u4e0d\u9700\u8981\u50cfLLM\u90a3\u6837\u8fdb\u884c\u5927\u89c4\u6a21\u7684\u9884\u8bad\u7ec3\uff0c\u4e5f\u907f\u514d\u4e86\u601d\u7ef4\u94fe\uff08CoT\uff09\u65b9\u6cd5\u56fa\u6709\u7684\u8106\u5f31\u6027\u3001\u9ad8\u6570\u636e\u9700\u6c42\u548c\u9ad8\u5ef6\u8fdf\u7b49\u95ee\u9898\u3002<\/li>\n<li><strong>\u8bad\u7ec3HRM\u9700\u8981\u4ec0\u4e48\u6837\u7684\u786c\u4ef6\uff1f<\/strong><br \/>\n\u8bad\u7ec3HRM\u9700\u8981\u652f\u6301CUDA\u7684NVIDIA GPU\u3002\u867d\u7136\u53ef\u4ee5\u5728\u5355\u5f20\u6d88\u8d39\u7ea7GPU\uff08\u5982RTX 3060\u62164070\uff09\u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0c\u4f46\u8fd9\u4f1a\u82b1\u8d39\u8f83\u957f\u65f6\u95f4\u3002\u5b98\u65b9\u63a8\u8350\u4f7f\u7528\u591aGPU\u73af\u5883\uff08\u59828\u5361\u670d\u52a1\u5668\uff09\u4ee5\u5b9e\u73b0\u9ad8\u6548\u8bad\u7ec3\uff0c\u4f8b\u5982\u57288\u4e2aGPU\u4e0a\u8bad\u7ec3\u6570\u72ec\u6a21\u578b\u4ec5\u970010\u5206\u949f\u3002<\/li>\n<li><strong>HRM\u7684\u8bad\u7ec3\u6570\u636e\u91cf\u8981\u6c42\u9ad8\u5417\uff1f<\/strong><br \/>\n\u4e0d\u9ad8\u3002HRM\u7684\u4e00\u5927\u4f18\u52bf\u5c31\u662f\u5bf9\u8bad\u7ec3\u6570\u636e\u7684\u9700\u6c42\u91cf\u6781\u5c0f\u3002\u5728\u6570\u72ec\u3001\u8ff7\u5bab\u548cARC\u7b49\u590d\u6742\u4efb\u52a1\u4e0a\uff0c\u5b83\u4ec5\u4f7f\u75281000\u4e2a\u8bad\u7ec3\u6837\u672c\u5c31\u8fbe\u5230\u4e86\u975e\u5e38\u9ad8\u7684\u6027\u80fd\u6c34\u5e73\u3002<\/li>\n<li><strong>\u5982\u4f55\u590d\u73b0\u8bba\u6587\u4e2d\u62a5\u544a\u7684\u6027\u80fd\uff1f<\/strong><br \/>\n\u7531\u4e8e\u5c0f\u6837\u672c\u5b66\u4e60\u7684\u968f\u673a\u6027\uff0c\u5355\u6b21\u8bad\u7ec3\u7684\u7ed3\u679c\u53ef\u80fd\u4f1a\u6709\u8f7b\u5fae\u6ce2\u52a8\u3002\u8bba\u6587\u4f5c\u8005\u6307\u51fa\uff0c\u5bf9\u4e8e\u6570\u72ec\u4efb\u52a1\uff0c\u51c6\u786e\u7387\u7684\u6807\u51c6\u5dee\u7ea6\u4e3a2%\u3002\u4e3a\u4e86\u8fbe\u5230\u6700\u4f73\u6027\u80fd\uff0c\u53ef\u80fd\u9700\u8981\u5fae\u8c03\u8bad\u7ec3\u65f6\u957f\uff0c\u5e76\u5728\u6a21\u578b\u5f00\u59cb\u8fc7\u62df\u5408\u4e4b\u524d\u8fdb\u884c\u65e9\u671f\u505c\u6b62\u3002<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"<p>HRM (Hierarchical Reasoning Model) \u662f\u4e00\u4e2a\u4ec5\u67092700\u4e07\u53c2\u6570\u7684\u5c42\u7ea7\u5f0f\u63a8\u7406\u6a21\u578b\uff0c\u65e8\u5728\u89e3\u51b3\u4eba\u5de5\u667a\u80fd\u9886\u57df\u4e2d\u590d\u6742\u7684\u63a8\u7406\u4efb\u52a1\u3002\u8be5\u6a21\u578b\u7684\u8bbe\u8ba1\u7075\u611f\u6765\u6e90\u4e8e\u4eba\u8111\u7684\u5c42\u7ea7\u5f0f\u3001\u591a\u65f6\u95f4\u5c3a\u5ea6\u7684\u4fe1\u606f\u5904\u7406\u65b9\u5f0f\u3002 \u5b83\u901a\u8fc7\u4e00\u4e2a\u9ad8\u5c42\u6a21\u5757\uff08\u8d1f\u8d23\u7f13&#8230;<\/p>\n","protected":false},"author":1,"featured_media":32782,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20,403,392],"tags":[230],"class_list":["post-42458","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tool","category-dedicated-model","category-models","tag-aikaiyuanxiangmu"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts\/42458","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/comments?post=42458"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts\/42458\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/media\/32782"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/media?parent=42458"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/categories?post=42458"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/tags?post=42458"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}