{"id":21251,"date":"2026-08-14T15:55:09","date_gmt":"2026-08-14T07:55:09","guid":{"rendered":"https:\/\/92it.top\/?p=21251"},"modified":"2026-08-14T16:21:22","modified_gmt":"2026-08-14T08:21:22","slug":"pytorch-%e5%92%8c-transformer-%e7%ae%80%e5%8d%95%e4%bb%8b%e7%bb%8d","status":"publish","type":"post","link":"https:\/\/92it.top\/?p=21251","title":{"rendered":"PyTorch \u548c Transformer \u7b80\u5355\u4ecb\u7ecd"},"content":{"rendered":"\n<p><strong>\u524d\u8a00 \ud83d\udd16<\/strong><\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p><strong>PyTorch \u662f\u4e00\u4e2a Python \u8bed\u8a00\u7684\u6df1\u5ea6\u5b66\u4e60\u5de5\u5177\u7bb1\uff08AI \u6846\u67b6\uff09\uff0c<\/strong>PyTorch \u6700\u521d\u7531 Meta Platforms \u7684\u4eba\u5de5\u667a\u80fd\u7814\u7a76\u56e2\u961f\u5f00\u53d1\uff0c\u73b0\u5728\u5c5e\u4e8eLinux \u57fa\u91d1\u4f1a\u7684\u4e00\u90e8\u5206<strong>\u3002<\/strong>PyTorch \u4ee5\u5176\u7075\u6d3b\u6027\u548c\u6613\u7528\u6027\u800c\u95fb\u540d\uff0c\u7279\u522b\u9002\u5408\u4e8e\u6df1\u5ea6\u5b66\u4e60\u7814\u7a76\u548c\u5f00\u53d1\u3002<\/p>\n\n\n\n<p>\u5b83\u672c\u8eab\u4e0d\u662f\u5927\u6a21\u578b\uff0c\u4e0d\u662f AI\uff0c\u4e0d\u80fd\u76f4\u63a5\u804a\u5929\u3001\u4e0d\u80fd\u751f\u6210\u56fe\u7247\u3002\u5b83\u5c31\u76f8\u5f53\u4e8e\u4e00\u5957\u9ad8\u7ea7 \u201c\u5de5\u7a0b\u673a\u68b0\u5de5\u5177\u7bb1\u201d\uff1a\u63d0\u4f9b\u5f20\u91cf\u3001\u77e9\u9635\u8fd0\u7b97\u3001GPU \u8c03\u5ea6\u3001\u81ea\u52a8\u6c42\u5bfc\u5168\u5957\u5de5\u5177\u3002<\/p>\n\n\n\n<p>AI \u5de5\u7a0b\u5e08\u62ff\u7740\u8fd9\u5957\u5de5\u5177\u7bb1\uff0c\u53bb\u642d\u5efa\u3001\u8bad\u7ec3\u3001\u8fd0\u884c\u795e\u7ecf\u7f51\u7edc\uff08\u5927\u6a21\u578b\u3001\u56fe\u50cf\u6a21\u578b\uff09\u3002<\/p>\n\n\n\n<pre class=\"wp-block-preformatted\">\u7c7b\u6bd4\u73b0\u5b9e\uff1a<br>PyTorch = \u4fee\u7406\u5382\u5168\u5957\u5de5\u5177\uff08\u6273\u624b\u3001\u710a\u673a\u3001\u53d8\u901f\u7bb1\u6d4b\u8bd5\u53f0\uff09<br>\u5927\u6a21\u578b (Qwen\u3001Flux) = \u7528\u8fd9\u5957\u5de5\u5177\u7ec4\u88c5\u51fa\u6765\u7684\u6c7d\u8f66\u3002<br>\u5de5\u5177\u672c\u8eab\u4e0d\u4f1a\u8dd1\uff0c\u5de5\u7a0b\u5e08\u4f7f\u7528\u5de5\u5177\uff0c\u9020\u51fa\u53ef\u4ee5\u8dd1\u7684\u8f66\u3002<\/pre>\n\n\n\n<p>\u8bb8\u591a\u6df1\u5ea6\u5b66\u4e60\u8f6f\u4ef6\u90fd\u662f\u57fa\u4e8e PyTorch \u6784\u5efa\u7684\uff0c\u5305\u62ec\u7279\u65af\u62c9\u81ea\u52a8\u9a7e\u9a76\u3001Uber \u7684 Pyro\u3001Hugging Face \u7684 Transformers\u3001 PyTorch Lightning \u548c Catalyst\u3002<\/p>\n\n\n\n<p>PyTorch \u4e3b\u8981\u6709\u4e24\u5927\u7279\u5f81\uff1a<\/p>\n\n\n\n<ul>\n<li>\u7c7b\u4f3c\u4e8e NumPy \u7684\u5f20\u91cf\u8ba1\u7b97\uff0c\u80fd\u5728 GPU \u6216 MPS \u7b49\u786c\u4ef6\u52a0\u901f\u5668\u4e0a\u52a0\u901f\u3002<\/li>\n\n\n\n<li>\u57fa\u4e8e\u5e26\u81ea\u52a8\u5fae\u5206\u7cfb\u7edf\u7684\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u3000\u3000<\/p>\n\n\n\n<p><strong>PyTorch \u6838\u5fc3\u7684\u51e0\u5927\u57fa\u7840\u7279\u5f81 \ud83d\udd16<\/strong><\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p><strong>\ud83d\udd391. \u5f20\u91cf Tensor \u2014\u2014 AI \u4e16\u754c\u7684\u901a\u7528\u6570\u636e\u5bb9\u5668<\/strong><\/p>\n\n\n\n<p>\u666e\u901a Python \u53ea\u80fd\u5904\u7406\u6570\u5b57\u3001\u5217\u8868\u3002<\/p>\n\n\n\n<p>PyTorch \u4e2d\u7684\u6838\u5fc3\u6570\u636e\u7ed3\u6784\u662f <strong>\u5f20\u91cf\uff08Tensor\uff09<\/strong>\uff0c\u5b83\u662f\u4e00\u4e2a\u591a\u7ef4\u77e9\u9635\uff0c\u53ef\u4ee5\u5728 CPU \u6216 GPU \u4e0a\u9ad8\u6548\u5730\u8fdb\u884c\u8ba1\u7b97\u3002\u5f20\u91cf\u7684\u64cd\u4f5c\u652f\u6301\u81ea\u52a8\u6c42\u5bfc\uff08Autograd\uff09\u673a\u5236\uff0c\u4f7f\u5f97\u5728\u53cd\u5411\u4f20\u64ad\u8fc7\u7a0b\u4e2d\u81ea\u52a8\u8ba1\u7b97\u68af\u5ea6\uff0c\u8fd9\u5bf9\u4e8e\u6df1\u5ea6\u5b66\u4e60\u4e2d\u7684\u68af\u5ea6\u4e0b\u964d\u4f18\u5316\u7b97\u6cd5\u81f3\u5173\u91cd\u8981\u3002<\/p>\n\n\n\n<p><strong>\u5f20\u91cf\u5c31\u662f\u591a\u7ef4\u6570\u5b57\u6570\u7ec4\uff0c\u662f PyTorch \u7684\u6838\u5fc3\u3002<\/strong><\/p>\n\n\n\n<ul>\n<li>\u6587\u5b57\u3001\u56fe\u7247\uff0c\u4e0d\u80fd\u76f4\u63a5\u5582\u7ed9\u795e\u7ecf\u7f51\u7edc\u3002 <\/li>\n\n\n\n<li>\u6587\u5b57\u7ecf\u8fc7\u5206\u8bcd\u5668\u53d8\u6210\u6570\u5b57 ID\uff0c\u56fe\u7247\u8f6c\u6210\u50cf\u7d20\u6570\u5b57\uff0c\u5168\u90e8\u5c01\u88c5\u6210<strong>\u5f20\u91cf<\/strong>\u3002<\/li>\n\n\n\n<li> \u5f20\u91cf\u53ef\u4ee5\u642c\u5bb6\uff1a\u53ef\u4ee5\u653e\u5728 CPU \u5185\u5b58\uff0c\u4e5f\u53ef\u4ee5\u642c\u5230 GPU\uff08N \u5361 CUDA \/ Mac MPS\uff09\uff0c\u5229\u7528\u663e\u5361\u9ad8\u901f\u5e76\u884c\u8ba1\u7b97\u3002<\/li>\n\n\n\n<li>\u63d0\u4f9b\u4e86\u7c7b\u4f3c NumPy \u7684\u63a5\u53e3\uff0c\u652f\u6301\u5143\u7d20\u7ea7\u8fd0\u7b97\u3002<\/li>\n\n\n\n<li>\u652f\u6301\u81ea\u52a8\u6c42\u5bfc\uff0c\u53ef\u4ee5\u65b9\u4fbf\u5730\u8fdb\u884c\u68af\u5ea6\u8ba1\u7b97\u3002<\/li>\n<\/ul>\n\n\n\n<pre class=\"wp-block-preformatted\">\u5927\u6a21\u578b\u91cc\u9762\u4ebf\u7ea7\u522b\u7684\u6743\u91cd\u53c2\u6570\uff0c\u5168\u90e8\u5b58\u50a8\u4e3a\u5f20\u91cf\u3002.safetensors\u6587\u4ef6\uff0c\u672c\u8d28\u5c31\u662f\u628a\u5927\u91cf\u5f20\u91cf\u5e8f\u5217\u5316\u4fdd\u5b58\u5230\u78c1\u76d8\u3002<\/pre>\n\n\n\n<p>\u3000\u3000<\/p>\n\n\n\n<p><strong>\ud83d\udd392. \u81ea\u52a8\u6c42\u5bfc Autograd\uff08\u8bad\u7ec3 AI \u7684\u7075\u9b42\uff09<\/strong><\/p>\n\n\n\n<p>\u8fd9\u662f AI \u6846\u67b6\u6700\u786c\u6838\u7684\u80fd\u529b\u3002\u8bad\u7ec3\u6a21\u578b\u7684\u65f6\u5019\uff1a<\/p>\n\n\n\n<ul>\n<li>\u8f93\u5165\u5f20\u91cf\u8d70\u7f51\u7edc\u8ba1\u7b97\uff0c\u5f97\u5230\u8f93\u51fa\uff1b<\/li>\n\n\n\n<li>\u5bf9\u6bd4\u6807\u51c6\u7b54\u6848\uff0c\u7b97\u51fa\u8bef\u5dee loss\uff1b<\/li>\n\n\n\n<li>loss.backward()\u81ea\u52a8\u53cd\u5411\u6c42\u5bfc\uff0c\u81ea\u52a8\u7b97\u51fa\u6bcf\u4e00\u4e2a\u6743\u91cd\u8be5\u5f80\u54ea\u4e2a\u65b9\u5411\u4fee\u6539\uff1b<\/li>\n\n\n\n<li>\u4f18\u5316\u5668\u81ea\u52a8\u66f4\u65b0 safetensors \u91cc\u9762\u4ebf\u4e07\u6743\u91cd\u6570\u503c\u3002<\/li>\n<\/ul>\n\n\n\n<pre class=\"wp-block-preformatted\">\u5982\u679c\u6ca1\u6709\u81ea\u52a8\u6c42\u5bfc\uff0c\u4eba\u7c7b\u9700\u8981\u624b\u5199\u6210\u5343\u4e0a\u4e07\u884c\u5fae\u79ef\u5206\uff0c\u51e0\u4e4e\u4e0d\u53ef\u80fd\u8bad\u7ec3\u51fa\u5927\u6a21\u578b\u3002<\/pre>\n\n\n\n<p>\u26a0\ufe0f\u533a\u5206\u4e24\u79cd\u6a21\u5f0f<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"generic\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\u63a8\u7406\u6a21\u5f0f\uff1a\u53ea\u524d\u5411\u8ba1\u7b97\uff08ComfyUI \u751f\u56fe\u3001\u5927\u6a21\u578b\u5bf9\u8bdd\uff09\uff0c\u4e0d\u9700\u8981\u53cd\u5411\u6c42\u5bfc\u3002\n\u8bad\u7ec3 \/ \u5fae\u8c03\u6a21\u5f0f\uff1a\u524d\u5411 + \u53cd\u5411\u6c42\u5bfc\uff0c\u66f4\u65b0\u6743\u91cd\uff0c\u4ea7\u51fa\u65b0\u7684 safetensors\u3002<\/pre>\n\n\n\n<p>\u3000\u3000<\/p>\n\n\n\n<p><strong>\ud83d\udd393. GPU \u52a0\u901f<\/strong><\/p>\n\n\n\n<p>PyTorch \u5b8c\u5168\u652f\u6301\u5728 GPU \u4e0a\u8fd0\u884c\uff0c\u4ee5\u52a0\u901f\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u7684\u8bad\u7ec3\u3002\u901a\u8fc7\u7b80\u5355\u7684 <code>.to(device)<\/code> \u65b9\u6cd5\uff0c\u7528\u6237\u53ef\u4ee5\u5c06\u6a21\u578b\u548c\u5f20\u91cf\u8f6c\u79fb\u5230 GPU \u4e0a\u8fdb\u884c\u8ba1\u7b97\u3002PyTorch \u652f\u6301\u591a GPU \u8bad\u7ec3\uff0c\u80fd\u591f\u5229\u7528 NVIDIA CUDA \u6280\u672f\u663e\u8457\u63d0\u9ad8\u8ba1\u7b97\u6548\u7387\u3002<\/p>\n\n\n\n<p>GPU \u652f\u6301\uff1a<\/p>\n\n\n\n<ul>\n<li>\u81ea\u52a8\u9009\u62e9 GPU \u6216 CPU\u3002<\/li>\n\n\n\n<li>\u652f\u6301\u901a\u8fc7 CUDA \u52a0\u901f\u8fd0\u7b97\u3002<\/li>\n\n\n\n<li>\u652f\u6301\u591a GPU \u5e76\u884c\u8ba1\u7b97\uff08<code>DataParallel<\/code> \u6216 <code>torch.distributed<\/code>\uff09\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u3000\u3000<\/p>\n\n\n\n<p><strong>\ud83d\udd394.\u6a21\u578b\u5b9a\u4e49\u4e0e\u8bad\u7ec3<\/strong><\/p>\n\n\n\n<p>PyTorch \u63d0\u4f9b\u4e86 <code>torch.nn<\/code> \u6a21\u5757\uff0c\u5141\u8bb8\u7528\u6237\u901a\u8fc7\u7ee7\u627f <code>nn.Module<\/code> \u7c7b\u6765\u5b9a\u4e49\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\u3002\u4f7f\u7528 <code>forward<\/code> \u51fd\u6570\u6307\u5b9a\u524d\u5411\u4f20\u64ad\uff0c\u81ea\u52a8\u53cd\u5411\u4f20\u64ad\uff08\u901a\u8fc7 <code>autograd<\/code>\uff09\u548c\u68af\u5ea6\u8ba1\u7b97\u4e5f\u7531 PyTorch \u5185\u90e8\u5904\u7406\u3002<\/p>\n\n\n\n<p>\u795e\u7ecf\u7f51\u7edc\u6a21\u5757\uff08torch.nn\uff09\uff1a<\/p>\n\n\n\n<ul>\n<li>\u63d0\u4f9b\u4e86\u5e38\u7528\u7684\u5c42\uff08\u5982\u7ebf\u6027\u5c42\u3001\u5377\u79ef\u5c42\u3001\u6c60\u5316\u5c42\u7b49\uff09\u3002<\/li>\n\n\n\n<li>\u652f\u6301\u5b9a\u4e49\u590d\u6742\u7684\u795e\u7ecf\u7f51\u7edc\u67b6\u6784\uff08\u5305\u62ec\u591a\u8f93\u5165\u3001\u591a\u8f93\u51fa\u7684\u7f51\u7edc\uff09\u3002<\/li>\n\n\n\n<li>\u517c\u5bb9\u4e0e\u4f18\u5316\u5668\uff08\u5982 <code>torch.optim<\/code>\uff09\u4e00\u8d77\u4f7f\u7528\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u3000\u3000<\/p>\n\n\n\n<p><strong>\ud83d\udd395.\u751f\u6001\u7cfb\u7edf\u4e0e\u793e\u533a\u652f\u6301<\/strong><\/p>\n\n\n\n<p>PyTorch \u4f5c\u4e3a\u4e00\u4e2a\u5f00\u6e90\u9879\u76ee\uff0c\u62e5\u6709\u4e00\u4e2a\u5e9e\u5927\u7684\u793e\u533a\u548c\u751f\u6001\u7cfb\u7edf\u3002\u5b83\u4e0d\u4ec5\u5728\u5b66\u672f\u754c\u5f97\u5230\u4e86\u5e7f\u6cdb\u7684\u5e94\u7528\uff0c\u4e5f\u5728\u5de5\u4e1a\u754c\uff0c\u7279\u522b\u662f\u5728\u8ba1\u7b97\u673a\u89c6\u89c9\u3001\u81ea\u7136\u8bed\u8a00\u5904\u7406\u7b49\u9886\u57df\u4e2d\u5f97\u5230\u4e86\u5e7f\u6cdb\u90e8\u7f72\u3002PyTorch \u8fd8\u63d0\u4f9b\u4e86\u8bb8\u591a\u4e0e\u6df1\u5ea6\u5b66\u4e60\u76f8\u5173\u7684\u5de5\u5177\u548c\u5e93\uff0c\u5982\uff1a<\/p>\n\n\n\n<ul>\n<li><strong>torchvision<\/strong>\uff1a\u7528\u4e8e\u8ba1\u7b97\u673a\u89c6\u89c9\u4efb\u52a1\u7684\u6570\u636e\u96c6\u548c\u6a21\u578b\u3002<\/li>\n\n\n\n<li><strong>torchtext<\/strong>\uff1a\u7528\u4e8e\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4efb\u52a1\u7684\u6570\u636e\u96c6\u548c\u9884\u5904\u7406\u5de5\u5177\u3002<\/li>\n\n\n\n<li><strong>torchaudio<\/strong>\uff1a\u7528\u4e8e\u97f3\u9891\u5904\u7406\u7684\u5de5\u5177\u5305\u3002<\/li>\n\n\n\n<li><strong>PyTorch Lightning<\/strong>\uff1a\u4e00\u79cd\u7b80\u5316 PyTorch \u4ee3\u7801\u7684\u9ad8\u5c42\u5e93\uff0c\u4e13\u6ce8\u4e8e\u7814\u7a76\u548c\u5b9e\u9a8c\u7684\u5feb\u901f\u8fed\u4ee3\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u3000\u3000<\/p>\n\n\n\n<p><strong>\ud83d\udd396.PyTorch \u7279\u6027<\/strong><\/p>\n\n\n\n<ul>\n<li><strong>\u52a8\u6001\u8ba1\u7b97\u56fe<\/strong>\uff08Dynamic Computation Graphs\uff09\uff1a PyTorch \u7684\u8ba1\u7b97\u56fe\u662f\u52a8\u6001\u7684\uff0c\u8fd9\u610f\u5473\u7740\u5b83\u4eec\u5728\u8fd0\u884c\u65f6\u6784\u5efa\uff0c\u5e76\u4e14\u53ef\u4ee5\u968f\u65f6\u6539\u53d8\u3002\u8fd9\u4e3a\u5b9e\u9a8c\u548c\u8c03\u8bd5\u63d0\u4f9b\u4e86\u6781\u5927\u7684\u7075\u6d3b\u6027\uff0c\u56e0\u4e3a\u5f00\u53d1\u8005\u53ef\u4ee5\u9010\u884c\u6267\u884c\u4ee3\u7801\uff0c\u67e5\u770b\u4e2d\u95f4\u7ed3\u679c\u3002<\/li>\n\n\n\n<li><strong>\u81ea\u52a8\u5fae\u5206<\/strong>\uff08Automatic Differentiation\uff09\uff1a PyTorch \u7684\u81ea\u52a8\u5fae\u5206\u7cfb\u7edf\u5141\u8bb8\u5f00\u53d1\u8005\u8f7b\u677e\u5730\u8ba1\u7b97\u68af\u5ea6\uff0c\u8fd9\u5bf9\u4e8e\u8bad\u7ec3\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u81f3\u5173\u91cd\u8981\u3002\u5b83\u901a\u8fc7\u53cd\u5411\u4f20\u64ad\u7b97\u6cd5\u81ea\u52a8\u8ba1\u7b97\u51fa\u635f\u5931\u51fd\u6570\u5bf9\u6a21\u578b\u53c2\u6570\u7684\u68af\u5ea6\u3002<\/li>\n\n\n\n<li><strong>\u5f20\u91cf\u8ba1\u7b97<\/strong>\uff08Tensor Computation\uff09\uff1a PyTorch \u63d0\u4f9b\u4e86\u7c7b\u4f3c\u4e8e NumPy \u7684\u5f20\u91cf\u64cd\u4f5c\uff0c\u8fd9\u4e9b\u64cd\u4f5c\u53ef\u4ee5\u5728 CPU \u548c GPU \u4e0a\u6267\u884c\uff0c\u4ece\u800c\u52a0\u901f\u8ba1\u7b97\u8fc7\u7a0b\u3002\u5f20\u91cf\u662f PyTorch \u4e2d\u7684\u57fa\u672c\u6570\u636e\u7ed3\u6784\uff0c\u7528\u4e8e\u5b58\u50a8\u548c\u64cd\u4f5c\u6570\u636e\u3002<\/li>\n\n\n\n<li><strong>\u4e30\u5bcc\u7684 API<\/strong>\uff1a PyTorch \u63d0\u4f9b\u4e86\u5927\u91cf\u7684\u9884\u5b9a\u4e49\u5c42\u3001\u635f\u5931\u51fd\u6570\u548c\u4f18\u5316\u7b97\u6cd5\uff0c\u8fd9\u4e9b\u90fd\u662f\u6784\u5efa\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u7684\u5e38\u7528\u7ec4\u4ef6\u3002<\/li>\n\n\n\n<li><strong>\u591a\u8bed\u8a00\u652f\u6301<\/strong>\uff1a PyTorch \u867d\u7136\u4ee5 Python \u4e3a\u4e3b\u8981\u63a5\u53e3\uff0c\u4f46\u4e5f\u63d0\u4f9b\u4e86 C++ \u63a5\u53e3\uff0c\u5141\u8bb8\u66f4\u5e95\u5c42\u7684\u96c6\u6210\u548c\u63a7\u5236\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u3000\u3000<\/p>\n\n\n\n<p><strong>\u4e0e\u5176\u4ed6\u6846\u67b6\u7684\u5bf9\u6bd4 \ud83d\udd16<\/strong><\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p>PyTorch \u7531\u4e8e\u5176\u7075\u6d3b\u6027\u3001\u6613\u7528\u6027\u548c\u793e\u533a\u652f\u6301\uff0c\u5df2\u7ecf\u6210\u4e3a\u5f88\u591a\u6df1\u5ea6\u5b66\u4e60\u7814\u7a76\u8005\u548c\u5f00\u53d1\u8005\u7684\u9996\u9009\u6846\u67b6\u3002<\/p>\n\n\n\n<p><strong>\ud83d\udd39TensorFlow vs PyTorch<\/strong><\/p>\n\n\n\n<ul>\n<li>PyTorch \u7684\u52a8\u6001\u8ba1\u7b97\u56fe\u4f7f\u5f97\u5b83\u66f4\u52a0\u7075\u6d3b\uff0c\u9002\u5408\u5feb\u901f\u5b9e\u9a8c\u548c\u7814\u7a76\uff1b\u800c TensorFlow \u7684\u9759\u6001\u8ba1\u7b97\u56fe\u5728\u751f\u4ea7\u73af\u5883\u4e2d\u66f4\u5177\u4f18\u5316\u7a7a\u95f4\u3002<\/li>\n\n\n\n<li>PyTorch \u5728\u8c03\u8bd5\u65f6\u66f4\u52a0\u65b9\u4fbf\uff0cTensorFlow \u5219\u5728\u90e8\u7f72\u4e0a\u66f4\u52a0\u6210\u719f\uff0c\u652f\u6301\u66f4\u5e7f\u6cdb\u7684\u786c\u4ef6\u548c\u5e73\u53f0\u3002<\/li>\n\n\n\n<li>\u8fd1\u5e74\u6765\uff0cTensorFlow \u4e5f\u5f15\u5165\u4e86\u52a8\u6001\u56fe\uff08\u5982 TensorFlow 2.x\uff09\uff0c\u4f7f\u5f97\u4e24\u8005\u5728\u529f\u80fd\u4e0a\u8d8b\u4e8e\u63a5\u8fd1\u3002<\/li>\n\n\n\n<li>\u5176\u4ed6\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\uff0c\u5982 Keras\u3001Caffe \u7b49\u4e5f\u6709\u4e00\u5b9a\u5e94\u7528\uff0c\u4f46 PyTorch \u7531\u4e8e\u5176\u7075\u6d3b\u6027\u3001\u6613\u7528\u6027\u548c\u793e\u533a\u652f\u6301\uff0c\u5df2\u7ecf\u6210\u4e3a\u5f88\u591a\u6df1\u5ea6\u5b66\u4e60\u7814\u7a76\u8005\u548c\u5f00\u53d1\u8005\u7684\u9996\u9009\u6846\u67b6\u3002<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-table is-style-stripes\"><table><thead><tr><th>\u7279\u6027<\/th><th>TensorFlow<\/th><th>PyTorch<\/th><\/tr><\/thead><tbody><tr><td>\u5f00\u53d1\u516c\u53f8<\/td><td>Google<\/td><td>Facebook (FAIR)<\/td><\/tr><tr><td>\u8ba1\u7b97\u56fe\u7c7b\u578b<\/td><td>\u9759\u6001\u8ba1\u7b97\u56fe\uff08\u5b9a\u4e49\u540e\u518d\u6267\u884c\uff09<\/td><td>\u52a8\u6001\u8ba1\u7b97\u56fe\uff08\u5b9a\u4e49\u5373\u6267\u884c\uff09<\/td><\/tr><tr><td>\u7075\u6d3b\u6027<\/td><td>\u4f4e\uff08\u8ba1\u7b97\u56fe\u5728\u7f16\u8bd1\u65f6\u6784\u5efa\uff0c\u4e0d\u6613\u4fee\u6539\uff09<\/td><td>\u9ad8\uff08\u8ba1\u7b97\u56fe\u5728\u6267\u884c\u65f6\u52a8\u6001\u521b\u5efa\uff0c\u6613\u4e8e\u4fee\u6539\u548c\u8c03\u8bd5\uff09<\/td><\/tr><tr><td>\u8c03\u8bd5<\/td><td>\u8f83\u96be\uff08\u9700\u8981\u4f7f\u7528 tf.debugging \u6216\u5916\u90e8\u5de5\u5177\u8c03\u8bd5\uff09<\/td><td>\u5bb9\u6613\uff08\u53ef\u4ee5\u76f4\u63a5\u5728 Python \u4e2d\u8fdb\u884c\u8c03\u8bd5\uff09<\/td><\/tr><tr><td>\u6613\u7528\u6027<\/td><td>\u4f4e\uff08\u8f83\u590d\u6742\uff0cAPI \u8f83\u591a\uff0c\u5b66\u4e60\u66f2\u7ebf\u8f83\u9661\u5ced\uff09<\/td><td>\u9ad8\uff08API \u7b80\u6d01\uff0c\u8bed\u6cd5\u66f4\u52a0\u63a5\u8fd1 Python\uff0c\u5bb9\u6613\u4e0a\u624b\uff09<\/td><\/tr><tr><td>\u90e8\u7f72<\/td><td>\u5f3a\uff08\u652f\u6301\u5e7f\u6cdb\u7684\u786c\u4ef6\uff0c\u5982 TensorFlow Lite\u3001TensorFlow.js\uff09<\/td><td>\u8f83\u5f31\uff08\u90e8\u7f72\u5de5\u5177\u548c\u5e73\u53f0\u76f8\u5bf9\u8f83\u5c11\uff0c\u867d\u7136\u6709 TensorFlow \u652f\u6301\uff09<\/td><\/tr><tr><td>\u793e\u533a\u652f\u6301<\/td><td>\u5f88\u5f3a\uff08\u6210\u719f\u4e14\u5e9e\u5927\u7684\u793e\u533a\uff0c\u5e7f\u6cdb\u7684\u6559\u7a0b\u548c\u6587\u6863\uff09<\/td><td>\u5f88\u5f3a\uff08\u793e\u533a\u6d3b\u8dc3\uff0c\u7279\u522b\u662f\u5728\u5b66\u672f\u754c\uff0c\u5feb\u901f\u53d1\u5c55\u7684\u751f\u6001\uff09<\/td><\/tr><tr><td>\u6a21\u578b\u8bad\u7ec3<\/td><td>\u652f\u6301\u5206\u5e03\u5f0f\u8bad\u7ec3\uff0c\u652f\u6301\u591a\u79cd\u8bbe\u5907\uff08\u5982 CPU\u3001GPU\u3001TPU\uff09<\/td><td>\u652f\u6301\u5206\u5e03\u5f0f\u8bad\u7ec3\uff0c\u652f\u6301\u591a GPU\u3001CPU \u548c TPU<\/td><\/tr><tr><td>API \u5c42\u7ea7<\/td><td>\u9ad8\u7ea7 API\uff1aKeras\uff1b\u4f4e\u7ea7 API\uff1aTensorFlow Core<\/td><td>\u9ad8\u7ea7 API\uff1aTorchVision\u3001TorchText \u7b49\uff1b\u4f4e\u7ea7 API\uff1aTorch<\/td><\/tr><tr><td>\u6027\u80fd<\/td><td>\u9ad8\uff08\u4f18\u5316\u65b9\u9762\u6210\u719f\uff0c\u9002\u5408\u751f\u4ea7\u73af\u5883\uff09<\/td><td>\u9ad8\uff08\u9002\u5408\u7814\u7a76\u548c\u539f\u578b\u5f00\u53d1\uff0c\u751f\u4ea7\u6027\u80fd\u4e5f\u5728\u63d0\u5347\uff09<\/td><\/tr><tr><td>\u81ea\u52a8\u6c42\u5bfc<\/td><td>\u901a\u8fc7 tf.GradientTape \u5b9e\u73b0\u52a8\u6001\u6c42\u5bfc\uff08\u8f83\u590d\u6742\uff09<\/td><td>\u901a\u8fc7 autograd \u52a8\u6001\u6c42\u5bfc\uff08\u66f4\u7b80\u6d01\u548c\u76f4\u89c2\uff09<\/td><\/tr><tr><td>\u8c03\u4f18\u4e0e\u53ef\u6269\u5c55\u6027<\/td><td>\u5f3a\uff08\u652f\u6301\u5728\u591a\u5e73\u53f0\u4e0a\u8fd0\u884c\uff0c\u5982 TensorFlow Serving \u7b49\uff09<\/td><td>\u8f83\u5f31\uff08\u867d\u7136\u5728\u5b66\u672f\u548c\u5b9e\u9a8c\u73af\u5883\u4e2d\u8868\u73b0\u4f18\u8d8a\uff0c\u4f46\u751f\u4ea7\u73af\u5883\u652f\u6301\u76f8\u5bf9\u8f83\u5c11\uff09<\/td><\/tr><tr><td>\u6846\u67b6\u7075\u6d3b\u6027<\/td><td>\u8f83\u4f4e\uff08TensorFlow 2.x \u5f15\u5165\u4e86\u52a8\u6001\u56fe\u7279\u6027\uff0c\u4f46\u4ecd\u4e0d\u5b8c\u5168\u7075\u6d3b\uff09<\/td><td>\u9ad8\uff08\u52a8\u6001\u56fe\u5e26\u6765\u66f4\u9ad8\u7684\u7075\u6d3b\u6027\uff09<\/td><\/tr><tr><td>\u652f\u6301\u591a\u79cd\u8bed\u8a00<\/td><td>\u652f\u6301\u591a\u79cd\u8bed\u8a00\uff08Python, C++, Java, JavaScript, etc.\uff09<\/td><td>\u4e3b\u8981\u652f\u6301 Python\uff08\u4f46\u4e5f\u6709 C++ API\uff09<\/td><\/tr><tr><td>\u517c\u5bb9\u6027\u4e0e\u8fc1\u79fb<\/td><td>TensorFlow 2.x \u4e0e\u65e7\u7248\u672c\u517c\u5bb9\u6027\u8f83\u597d<\/td><td>\u4e0e TensorFlow \u517c\u5bb9\u6027\u5dee\uff0c\u8fc1\u79fb\u8f83\u96be<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p>\u3000\u3000<\/p>\n\n\n\n<p><strong>\ud83d\udd39PyTorch vs NumPy<\/strong><\/p>\n\n\n\n<figure class=\"wp-block-table is-style-stripes\"><table><thead><tr><th>\u7279\u6027<\/th><th>PyTorch<\/th><th>NumPy<\/th><\/tr><\/thead><tbody><tr><td>\u76ee\u6807<\/td><td>\u6df1\u5ea6\u5b66\u4e60\u4e13\u7528<\/td><td>\u901a\u7528\u79d1\u5b66\u8ba1\u7b97<\/td><\/tr><tr><td>GPU \u652f\u6301<\/td><td>\u539f\u751f\u652f\u6301 CUDA<\/td><td>\u4e0d\u76f4\u63a5\u652f\u6301<\/td><\/tr><tr><td>\u81ea\u52a8\u5fae\u5206<\/td><td>\u5185\u7f6e\u81ea\u52a8\u6c42\u5bfc<\/td><td>\u9700\u8981\u624b\u52a8\u8ba1\u7b97\u68af\u5ea6<\/td><\/tr><tr><td>\u795e\u7ecf\u7f51\u7edc<\/td><td>\u4e30\u5bcc\u7684\u795e\u7ecf\u7f51\u7edc\u6a21\u5757<\/td><td>\u9700\u8981\u4ece\u96f6\u5b9e\u73b0<\/td><\/tr><tr><td>\u5b66\u4e60\u6210\u672c<\/td><td>\u76f8\u5bf9\u8f83\u9ad8<\/td><td>\u76f8\u5bf9\u8f83\u4f4e<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p>\u3000\u3000<\/p>\n\n\n\n<p><strong>\u5f88\u591a\u5927\u6a21\u578b\u90fd\u662f PyTorch \u8bad\u7ec3\u51fa\u6765\u7684\u5417\uff1f\ud83d\udd16<\/strong><\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p>\u662f\u7684\uff0c\u73b0\u5728\u51e0\u4e4e\u6240\u6709\u4e3b\u6d41\u5f00\u6e90\u5927\u6a21\u578b\uff1aQwen\u3001Llama\u3001DeepSeek\u3001Z\u2011Image\u2011Turbo\u3001Flux\uff0c\u7edd\u5927\u591a\u6570\u90fd\u662f\u7528 PyTorch \u5b8c\u6210\u8bad\u7ec3\u3002<\/p>\n\n\n\n<ul>\n<li>\u65e9\u671f\u90e8\u5206\u8001\u6a21\u578b\u6709\u7528 TensorFlow\uff1b <\/li>\n\n\n\n<li>\u5c11\u91cf\u5b9e\u9a8c\u5ba4\u7528 JAX\uff1b <\/li>\n\n\n\n<li>\u4f46\u5f00\u6e90\u5708\u5b50\u3001\u56fd\u5185\u5927\u5382\uff0c<strong>PyTorch \u662f\u7edd\u5bf9\u4e3b\u6d41<\/strong>\u3002<\/li>\n<\/ul>\n\n\n\n<pre class=\"wp-block-preformatted\">\u8bad\u7ec3\u7684\u65f6\u5019\uff1a<br>PyTorch \u6846\u67b6\u8d1f\u8d23\uff1a\u5f20\u91cf\u7ba1\u7406\u3001\u77e9\u9635\u8fd0\u7b97\u3001GPU \u8c03\u5ea6 (CUDA)\u3001\u81ea\u52a8\u6c42\u5bfc\u53cd\u5411\u4f20\u64ad\u3001\u66f4\u65b0\u6743\u91cd\u3002<br>\u8bad\u7ec3\u8dd1\u5728\u5927\u91cf NVIDIA \u663e\u5361\u96c6\u7fa4\u4e0a\u3002<\/pre>\n\n\n\n<p><strong>\u5927\u6a21\u578b\u8bad\u7ec3\u5b8c\u6210\u4e4b\u540e\uff0c\u4ea7\u51fa\u4ec0\u4e48\uff08\u91cd\u70b9\uff09<\/strong><\/p>\n\n\n\n<p>\u8bad\u7ec3\u8dd1\u5b8c\uff0c\u4e0d\u4f1a\u76f4\u63a5\u51fa\u6765\u4e00\u4e2a\u53ef\u4ee5\u53cc\u51fb\u8fd0\u884c\u7684\u7a0b\u5e8f\u3002\u8f93\u51fa\u4e00\u5806\u6587\u4ef6\uff0c\u8fd9\u5c31\u662f<strong>PyTorch \u539f\u751f\u4ea7\u7269<\/strong>\uff1a<\/p>\n\n\n\n<ol>\n<li><strong><code>config.json<\/code><\/strong> \n<ul>\n<li>\u6a21\u578b\u914d\u7f6e\u6587\u4ef6\uff1a\u5c42\u6570\u3001\u6ce8\u610f\u529b\u5934\u6570\u3001\u4e0a\u4e0b\u6587\u7a97\u53e3\u5927\u5c0f\u3001hidden \u7ef4\u5ea6\u7b49\uff0c\u8bb0\u5f55\u795e\u7ecf\u7f51\u7edc \u201c\u56fe\u7eb8\u201d\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong><code>.safetensors<\/code>\uff08\u5206\u7247\u6743\u91cd\u6587\u4ef6\uff09<\/strong> \n<ul>\n<li>\u771f\u6b63\u7684\u6a21\u578b\u53c2\u6570\uff0c\u4ebf\u4e07\u6d6e\u70b9\u6570\u6743\u91cd\u3002 <\/li>\n\n\n\n<li>\u6bd4\u5982\uff1a<code>model\u201100001\u2011of\u201100026.safetensors<\/code><\/li>\n\n\n\n<li>\u8fd9\u5c31\u662f<strong>PyTorch \u539f\u751f\u6743\u91cd<\/strong>\uff0c\u53ea\u6709\u4e00\u5806\u6570\u5b57\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u5206\u8bcd\u5668\u76f8\u5173\uff1a<code>tokenizer.json<\/code>\u3001<code>vocab.json<\/code>\n<ul>\n<li>\u8d1f\u8d23\u6587\u5b57\u8f6c\u6570\u5b57 token\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u6709\u4e9b\u8fd8\u4f1a\u9644\u5e26<code>model.py<\/code>\u6a21\u578b\u4ee3\u7801\uff08\u7f51\u7edc\u7ed3\u6784\u5b9e\u73b0\uff09<\/li>\n<\/ol>\n\n\n\n<p>\u6574\u5957\u5408\u5728\u4e00\u8d77\uff0c\u624d\u662f\u5b8c\u6574\u539f\u7248\u6a21\u578b\uff0c\u53ea\u6709 safetensors\uff0c\u6ca1\u6709 config\uff0c\u662f\u52a0\u8f7d\u4e0d\u4e86\u6a21\u578b\u7684\u3002<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"generic\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\u5927\u91cf\u8bad\u7ec3\u6570\u636e \u2192 PyTorch\u6846\u67b6 + \u591a\u5757N\u5361\u96c6\u7fa4\u8bad\u7ec3\n        \u2193\n\u3010\u4ea7\u51fa\u3011config.json + \u4e00\u5806safetensors\u6743\u91cd + \u5206\u8bcd\u5668\u6587\u4ef6\uff08\u539f\u751fPyTorch\u4ea7\u7269\uff09<\/pre>\n\n\n\n<p>\u3000\u3000<\/p>\n\n\n\n<p><strong>Transformer \u662f\u4ec0\u4e48\uff1f\u4e3a\u4ec0\u4e48\u6709 PyTorch \u8fd8\u9700\u8981 Transformer\uff1f\ud83d\udd16<\/strong><\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p>\u5148\u5206\u6e05\u4e24\u4e2a\u89d2\u8272\uff0c\u8fd9\u662f\u5f88\u591a\u4eba\u6df7\u6dc6\u7684\u5173\u952e\u70b9\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-preformatted\">PyTorch\uff1a\u901a\u7528\u8ba1\u7b97\u5de5\u5177\u7bb1\uff08\u5e95\u5c42\u79ef\u6728\uff0c\u4ec0\u4e48\u6df1\u5ea6\u5b66\u4e60\u90fd\u80fd\u505a\uff09<br>Transformer\uff1a\u4e00\u79cd\u7279\u5b9a\u795e\u7ecf\u7f51\u7edc\u7684\u300c\u7b97\u6cd5 \/ \u7f51\u7edc\u67b6\u6784\u300d\uff08\u4e13\u95e8\u5904\u7406\u5e8f\u5217\u3001\u6587\u5b57\u7684\u4e00\u5957\u62fc\u88c5\u65b9\u6848\uff09<br><\/pre>\n\n\n\n<p><strong>\u901a\u4fd7\u6bd4\u55bb\uff1a<\/strong><\/p>\n\n\n\n<ul>\n<li><strong>PyTorch = \u4e94\u91d1\u5de5\u5177\u7bb1<\/strong>\uff1a\u91cc\u9762\u6709\u87ba\u4e1d\u3001\u87ba\u6bcd\u3001\u6273\u624b\u3001\u77e9\u9635\u4e58\u6cd5\u5de5\u5177\u3001\u5f20\u91cf\u3001\u81ea\u52a8\u6c42\u5bfc\u3002\u5de5\u5177\u7bb1\u672c\u8eab<strong>\u4e0d\u4ee3\u8868\u67d0\u4e2a\u673a\u5668<\/strong>\uff0c\u4f60\u53ef\u4ee5\u7528\u5b83\u642d CNN \u56fe\u50cf\u5377\u79ef\u7f51\u7edc\u3001RNN\u3001Transformer\uff0c\u968f\u4fbf\u4ec0\u4e48\u7f51\u7edc\u3002<\/li>\n\n\n\n<li><strong>Transformer = \u4e00\u5f20\u5b8c\u6574\u7684\u673a\u5668\u7ec4\u88c5\u56fe\u7eb8<\/strong>\uff1a\u89c4\u5b9a\u600e\u4e48\u628a\u5de5\u5177\u7bb1\u91cc\u7684\u96f6\u4ef6\u62fc\u51fa\u4e00\u53f0\u4e13\u95e8\u5904\u7406\u6587\u5b57\u7684\u673a\u5668\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u5de5\u5177\u7bb1 (PyTorch) \u4e0d\u80fd\u76f4\u63a5\u4ea7\u51fa\u5927\u6a21\u578b\uff1b\u4f60\u9700\u8981\u6309\u7167 Transformer \u8fd9\u5f20\u56fe\u7eb8\uff0c\u62ff\u5de5\u5177\u7bb1\u7684\u96f6\u4ef6\u628a\u673a\u5668\u62fc\u51fa\u6765\u3002<\/p>\n\n\n\n<p><strong>\ud83d\udd39Transformer \u662f\u4ec0\u4e48<\/strong>\uff1f<\/p>\n\n\n\n<p>Transformer \u662f 2017 \u5e74\u8bba\u6587\u63d0\u51fa\u7684\u795e\u7ecf\u7f51\u7edc\u7ed3\u6784\uff0c\u6838\u5fc3\u53d1\u660e\u662f \u81ea\u6ce8\u610f\u529b\u673a\u5236\uff08Self\u2011Attention\uff09\u3002\u4e13\u95e8\u89e3\u51b3\uff1a\u6587\u5b57\u3001\u5bf9\u8bdd\u8fd9\u7c7b\u5e8f\u5217\u6570\u636e\u3002<\/p>\n\n\n\n<ul>\n<li>\u4ee5\u524d\u8001\u7f51\u7edc RNN\uff1a\u6587\u5b57\u8981\u4e00\u4e2a\u5b57\u4e00\u4e2a\u987a\u5e8f\u4e32\u884c\u7b97\uff0c\u6162\uff0c\u957f\u6587\u672c\u8bb0\u4e0d\u4f4f\u3002<\/li>\n\n\n\n<li>Transformer \u7684\u81ea\u6ce8\u610f\u529b\uff1a\u53ef\u4ee5\u540c\u65f6\u770b\u6574\u6bb5\u6240\u6709 token\uff0c\u770b\u8bcd\u8bed\u4e4b\u95f4\u7684\u5173\u8054\u3002<\/li>\n<\/ul>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"generic\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\u4e3e\u4f8b\u5b50\u53e5\u5b50\uff1a\u5c0f\u660e\u628a\u676f\u5b50\u9012\u7ed9\u4e86\u4ed6\n\u81ea\u6ce8\u610f\u529b\u53ef\u4ee5\u8bc6\u522b\u51fa\uff1a\u201c\u4ed6\u201d \u6307\u7684\u662f\u5c0f\u660e\u3002\n\u8fd9\u5c31\u662f\u5927\u6a21\u578b\u7406\u89e3\u4e0a\u4e0b\u6587\u7684\u6839\u6e90\u3002<\/pre>\n\n\n\n<p>Transformer \u7531\u4e24\u5927\u5757\u7ec4\u6210\uff1a<\/p>\n\n\n\n<ol>\n<li><strong>\u7f16\u7801\u5668 Encoder<\/strong>\uff1a\u7406\u89e3\u8f93\u5165\uff08\u7ffb\u8bd1\u6a21\u578b\u7528\uff09<\/li>\n\n\n\n<li><strong>\u89e3\u7801\u5668 Decoder<\/strong>\uff1a\u751f\u6210\u6587\u5b57\uff0c\u73b0\u5728\u6240\u6709\u5927\u8bed\u8a00\u6a21\u578b\u53ea\u7528 Decoder \u90e8\u5206\u3002 Qwen\u3001Llama\u3001DeepSeek \u5168\u90e8\u90fd\u662f Decoder\u2011only Transformer\u3002<\/li>\n<\/ol>\n\n\n\n<p>\u5b83\u5185\u90e8\u7684\u7ec4\u4ef6\uff1a<\/p>\n\n\n\n<ul>\n<li>\u591a\u5934\u81ea\u6ce8\u610f\u529b Multi\u2011Head Attention<\/li>\n\n\n\n<li>\u524d\u5411\u7f51\u7edc FFN<\/li>\n\n\n\n<li>\u5f52\u4e00\u5316\u5c42 RMSNorm\/LayerNorm<\/li>\n\n\n\n<li>\u4f4d\u7f6e\u7f16\u7801 RoPE\uff0c\u544a\u8bc9\u6a21\u578b\u6587\u5b57\u7684\u5148\u540e\u987a\u5e8f<\/li>\n<\/ul>\n\n\n\n<p>\u8fd9\u4e9b\u7ec4\u4ef6\uff0c\u5168\u90e8\u90fd\u662f\u7528 PyTorch \u7684 API \u5199\u51fa\u6765\u7684\u3002<\/p>\n\n\n\n<p>Attention \u672c\u8d28\u5c31\u662f\u5927\u91cf\u77e9\u9635\u4e58\u6cd5\uff0c\u77e9\u9635\u4e58\u6cd5\u5c31\u662f PyTorch \u63d0\u4f9b\u7684\u7b97\u5b50\u3002<\/p>\n\n\n\n<p>\u3000\u3000<\/p>\n\n\n\n<p><strong>\ud83d\udd39\u4e3a\u4ec0\u4e48\u6709 PyTorch\uff0c\u8fd8\u9700\u8981 Transformer\uff1f<\/strong><\/p>\n\n\n\n<p>PyTorch \u53ea\u662f\u57fa\u7840\u5de5\u5177\uff0c\u5b83\u4e0d\u77e5\u9053 \u201c\u600e\u4e48\u7406\u89e3\u8bed\u8a00\u201d\uff0cPyTorch \u4f1a\u505a\u77e9\u9635\u4e58\u3001\u5f20\u91cf\u3001\u6c42\u5bfc\uff0c\u4f46\u5b83\u4e0d\u61c2\u4ec0\u4e48\u662f\u53e5\u5b50\u3001\u4ec0\u4e48\u662f\u4e0a\u4e0b\u6587\u5173\u7cfb\u3002PyTorch \u6ca1\u6709\u5185\u7f6e \u201c\u5927\u6a21\u578b\u201d\u3002\u4f60\u5fc5\u987b\u4eba\u5199\u4ee3\u7801\uff0c\u628a Attention\u3001FFN \u4e00\u5c42\u5c42\u7ec4\u88c5\u8d77\u6765\uff0c\u8fd9\u4e2a\u7ec4\u88c5\u65b9\u6848\u5c31\u662f Transformer\u3002<\/p>\n\n\n\n<ul>\n<li>PyTorch\uff1a\u901a\u7528\u6df1\u5ea6\u5b66\u4e60\u5de5\u5177\u5e93\uff0c\u63d0\u4f9b\u5f20\u91cf\u3001\u77e9\u9635\u8fd0\u7b97\u3001GPU \u8c03\u5ea6\u3001\u81ea\u52a8\u6c42\u5bfc\u3002\u672c\u8eab\u4e0d\u61c2\u8bed\u8a00\u3002<\/li>\n\n\n\n<li>Transformer\uff1a\u4e00\u5957\u795e\u7ecf\u7f51\u7edc\u67b6\u6784\uff08\u7b97\u6cd5\u56fe\u7eb8\uff09\uff0c\u4f9d\u9760\u81ea\u6ce8\u610f\u529b\uff0c\u64c5\u957f\u5904\u7406\u6587\u5b57\u5e8f\u5217\uff0c\u73b0\u5728\u5927\u6a21\u578b\u7684\u6807\u51c6\u65b9\u6848\u3002<\/li>\n\n\n\n<li>PyTorch \u7528\u6765\u5b9e\u73b0\u8fd0\u884c Transformer\uff1b\u4e8c\u8005\u4e0d\u662f\u4e8c\u9009\u4e00\uff0c\u662f\u56fe\u7eb8 vs \u5de5\u5177\u7684\u5173\u7cfb\u3002<\/li>\n\n\n\n<li>Transformer \u8fd9\u5957\u7b97\u6cd5\u903b\u8f91\uff0c\u4e0d\u53ea\u53ef\u4ee5\u8dd1\u5728 PyTorch\uff0c\u4e5f\u53ef\u4ee5\u7528 C++(llama\u2011cpp)\u3001MLX\u3001JAX \u5b9e\u73b0\u3002<\/li>\n\n\n\n<li>config.json\u91cc\u9762\u4fdd\u5b58\u7684\uff0c\u5c31\u662f\u8fd9\u4e2a Transformer \u7f51\u7edc\u7684\u56fe\u7eb8\u53c2\u6570\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u3000\u3000<\/p>\n\n\n\n<p><strong>\ud83d\udd39Transformer \u662f\u5927\u6a21\u578b\u7684\u6838\u5fc3\u8bbe\u8ba1\u67b6\u6784<\/strong><\/p>\n\n\n\n<p>Transformer \u662f\u5927\u6a21\u578b\u7684\u6838\u5fc3\u8bbe\u8ba1\u67b6\u6784\uff08\u795e\u7ecf\u7f51\u7edc\u56fe\u7eb8\uff09\uff0c\u4f46 Transformer \u2260 \u5927\u6a21\u578b\u672c\u8eab\u3002<\/p>\n\n\n\n<p>\u6253\u4e2a\u6bd4\u65b9\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-preformatted\">Transformer = \u6c7d\u8f66\u7684\u8bbe\u8ba1\u56fe\u7eb8\uff08\u89c4\u5b9a\u53d1\u52a8\u673a\u600e\u4e48\u5e03\u5c40\u3001\u53d8\u901f\u7bb1\u7ed3\u6784\uff09<br>\u5927\u6a21\u578b = \u6309\u7167\u8fd9\u4efd\u56fe\u7eb8\u9020\u51fa\u6765\u7684\u4e00\u53f0\u5b8c\u6574\u6c7d\u8f66\u3002<\/pre>\n\n\n\n<ol>\n<li>\u73b0\u5728\u51e0\u4e4e\u6240\u6709\u5f00\u6e90\u5927\u8bed\u8a00\u6a21\u578b Qwen\u3001Llama\u3001DeepSeek\uff0c\u7528\u7684\u90fd\u662f <strong>Transformer \u7684 Decoder\u2011only\uff08\u4ec5\u89e3\u7801\u5668\uff09\u53d8\u4f53<\/strong>\u3002<code>config.json<\/code> \u91cc\u9762\u5b58\u7684\uff1a\u5c42\u6570\u3001\u6ce8\u610f\u529b\u5934\u6570\u3001\u9690\u85cf\u7ef4\u5ea6\u3001RoPE \u53c2\u6570\uff0c\u5168\u90e8\u5c31\u662f\u8fd9\u4efd Transformer \u56fe\u7eb8\u4e0a\u7684\u53c2\u6570\u3002<\/li>\n\n\n\n<li>\u6709\u4e86\u67b6\u6784\u56fe\u7eb8\uff08Transformer\uff09\uff0c\u8fd8\u9700\u8981\u4e24\u6837\u4e1c\u897f\u624d\u53eb\u4e00\u4e2a\u53ef\u7528\u5927\u6a21\u578b\uff1a\n<ul>\n<li>\u6d77\u91cf\u8bad\u7ec3\u6570\u636e\uff0c\u5582\u7ed9 PyTorch \u53bb\u8bad\u7ec3\uff1b<\/li>\n\n\n\n<li>\u8bad\u7ec3\u51fa\u6765\u7684\u6570\u5341\u4ebf\u6743\u91cd\uff08safetensors\uff09\uff0c\u4e5f\u5c31\u662f\u6a21\u578b\u5b66\u5230\u7684\u77e5\u8bc6\u3002<\/li>\n\n\n\n<li>\u540c\u6837\u4e00\u5f20 Transformer \u56fe\u7eb8\uff0c\u7528\u4e0d\u540c\u6570\u636e\u8bad\u7ec3\uff0c\u53ef\u4ee5\u5f97\u5230\u5b8c\u5168\u4e0d\u4e00\u6837\u80fd\u529b\u7684\u6a21\u578b\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>Transformer \u4e0d\u662f\u53ea\u6709\u5927\u8bed\u8a00\u6a21\u578b\u5728\u7528\n<ul>\n<li>\u591a\u6a21\u6001 Qwen\u2011VL\uff1a\u4e3b\u5e72\u4f9d\u65e7\u662f Transformer Decoder\uff0c\u989d\u5916\u52a0\u4e00\u4e2a\u56fe\u50cf\u7f16\u7801\u5668\uff1b<\/li>\n\n\n\n<li>Flux \/ Z\u2011Image\u2011Turbo\uff08DiT\uff09\uff1a\u4e5f\u662f\u57fa\u4e8e Transformer \u67b6\u6784\u505a\u56fe\u50cf\u751f\u6210\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u91cd\u8981\uff1a<strong>\u67b6\u6784\u662f\u7b97\u6cd5\u601d\u60f3\uff0c\u53ef\u4ee5\u7528\u4e0d\u540c\u5de5\u5177\u5b9e\u73b0<\/strong>\n<ul>\n<li>PyTorch\uff1aPython \u5b9e\u73b0 Transformer\uff0c\u7528\u6765\u8bad\u7ec3\u3001\u5fae\u8c03\uff1b<\/li>\n\n\n\n<li>llama\u2011cpp\uff1aC++ \u624b\u5199\u91cd\u5199\u4e00\u904d\u4e00\u6a21\u4e00\u6837\u7684 Transformer Decoder \u903b\u8f91\uff0c\u8bfb GGUF \u505a\u63a8\u7406\uff0c<strong>\u4e0d\u518d\u4f9d\u8d56 PyTorch<\/strong>\u3002<\/li>\n\n\n\n<li>\u67b6\u6784\uff08Transformer\uff09\u6ca1\u53d8\uff0c\u53ea\u662f\u6362\u4e86\u4e00\u5957\u4ee3\u7801\u5de5\u5177\u53bb\u8fd0\u884c\u5b83\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n\n\n\n<p>Transformer \u662f\u5f53\u4ee3\u5927\u6a21\u578b\u7684\u6807\u51c6\u8bbe\u8ba1\u67b6\u6784\uff08\u7f51\u7edc\u7ed3\u6784\u56fe\u7eb8\uff09\uff1b<\/p>\n\n\n\n<pre class=\"wp-block-preformatted\">\u5927\u6a21\u578b = Transformer \u67b6\u6784 + \u8bad\u7ec3\u5f97\u5230\u7684\u6743\u91cd\u77e5\u8bc6\u3002<\/pre>\n\n\n\n<p>\u3000\u3000<\/p>\n\n\n\n<p><strong>\u5404\u5bb6\u5927\u6a21\u578b\u90fd\u662f\u5e95\u5c42\u90fd\u662f Transformer\u67b6\u6784\uff0c\u4e3a\u4ec0\u4e48\u6a21\u578b\u51fa\u6765\u5dee\u8ddd\u5de8\u5927<\/strong>\uff1f<strong>\ud83d\udd16<\/strong><\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p>\u8fd9\u4e2a\u95ee\u9898\u7279\u522b\u5173\u952e\uff1a\u90fd\u662f PyTorch\uff0c\u5e95\u5c42\u90fd\u662f Transformer\uff0c\u8f93\u5165\u90fd\u662f token \/ \u5f20\u91cf\uff0c\u4f46\u5404\u5bb6\u6a21\u578b\u51fa\u6765\u5dee\u8ddd\u5de8\u5927\uff0c\u5e76\u4e0d\u662f\u5957\u540c\u4e00\u4e2a\u6a21\u677f\u5c31\u5b8c\u4e8b\u3002<\/p>\n\n\n\n<p>\u6253\u4e2a\u6bd4\u65b9\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-preformatted\">\u5927\u5bb6\u90fd\u7528\u540c\u4e00\u5957\u4e50\u9ad8\u79ef\u6728\uff08PyTorch \u6846\u67b6\u3001Transformer \u57fa\u7840\u7ec4\u4ef6\uff09\u3002<br>\u4f46\u4e0d\u540c\u5de5\u7a0b\u5e08\uff0c\u9009\u7684\u96f6\u4ef6\u6570\u91cf\u3001\u62fc\u88c5\u65b9\u5f0f\u3001\u8bad\u7ec3\u6570\u636e\u3001\u8bad\u7ec3\u914d\u65b9\u3001\u76ee\u6807\u4e0d\u4e00\u6837\uff0c\u62fc\u51fa\u6765\u7684\u6210\u54c1\u5b8c\u5168\u4e0d\u4e00\u6837\u3002<br>\u79ef\u6728\u5e93\u662f\u4e00\u6837\u7684\uff0c\u4e0d\u7b49\u4e8e\u9020\u51fa\u6765\u7684\u8f66\u90fd\u4e00\u6837\u3002<\/pre>\n\n\n\n<p><strong>\u5927\u6a21\u578b\u5dee\u522b\u5230\u5e95\u5728\u54ea<\/strong><\/p>\n\n\n\n<p><strong><strong>\ud83d\udd39<\/strong>1\u3001\u8bad\u7ec3\u7528\u7684\u6570\u636e\u4e0d\u4e00\u6837\uff08\u6700\u5927\u7684\u533a\u522b\uff09<\/strong><\/p>\n\n\n\n<p>\u540c\u6837\u7684\u7f51\u7edc\u7ed3\u6784\uff0c\u5582\u4e0d\u540c\u6570\u636e\uff0c\u6a21\u578b\u6027\u683c\u3001\u77e5\u8bc6\u3001\u80fd\u529b\u5929\u5dee\u5730\u522b\u3002<\/p>\n\n\n\n<ul>\n<li>Qwen\uff1a\u5927\u91cf\u4e2d\u6587\u4e92\u8054\u7f51\u3001\u4e66\u7c4d\u3001\u591a\u8bed\u8a00\u6570\u636e\uff1b\u4e2d\u6587\u80fd\u529b\u5f3a\u3002<\/li>\n\n\n\n<li>Llama\uff1a\u82f1\u6587\u4e3a\u4e3b\uff0c\u4e2d\u6587\u539f\u59cb\u7248\u672c\u5f31\u3002<\/li>\n\n\n\n<li>DeepSeek\uff1a\u5927\u91cf\u4ee3\u7801\u3001\u63a8\u7406\u7c7b\u6587\u672c\uff0c\u64c5\u957f\u505a\u9898\u5199\u4ee3\u7801\u3002<\/li>\n<\/ul>\n\n\n\n<pre class=\"wp-block-preformatted\">\u8bad\u7ec3 = \u6a21\u578b\u5728\u6d77\u91cf\u6570\u636e\u91cc\u9762\u7edf\u8ba1\u5b66\u4e60\u8bed\u8a00\u89c4\u5f8b\u3002<br>\u6570\u636e\u7684\u8d28\u91cf\u3001\u8bed\u79cd\u3001\u9898\u6750\u3001\u6e05\u6d17\u8fc7\u6ee4\u65b9\u5f0f\uff0c\u76f4\u63a5\u51b3\u5b9a\u6a21\u578b\u61c2\u4ec0\u4e48\u3001\u4e0d\u61c2\u4ec0\u4e48\u3001\u4f1a\u4e0d\u4f1a\u80e1\u8bf4\u3002<br>\u54ea\u6015\u7f51\u7edc\u4e00\u6a21\u4e00\u6837\uff0c\u6362\u4e00\u5957\u6570\u636e\u96c6\uff0c\u6a21\u578b\u5c31\u53d8\u6210\u53e6\u4e00\u4e2a \u201c\u8111\u5b50\u201d\u3002<\/pre>\n\n\n\n<p>\u3000\u3000<\/p>\n\n\n\n<p><strong><strong>\ud83d\udd39<\/strong>2\u3001\u6a21\u578b\u7ed3\u6784\u7ec6\u8282\u4e0d\u4e00\u6837\uff08\u867d\u7136\u90fd\u662f Transformer\uff09<\/strong><\/p>\n\n\n\n<p>\u90fd\u53eb Transformer\uff0c\u4f46\u53ef\u4ee5\u6539\u5f88\u591a\u5730\u65b9\uff1a<\/p>\n\n\n\n<ol>\n<li>\u5c42\u6570\u3001\u5934\u6570\u3001\u9690\u85cf\u7ef4\u5ea6\uff1a27B\u300114B\u30017B\uff0c\u89c4\u6a21\u672c\u8eab\u5c31\u4e0d\u540c\u3002<\/li>\n\n\n\n<li>\u53d8\u4f53\uff1aMoE \u6df7\u5408\u4e13\u5bb6 \/ \u7a20\u5bc6\u6a21\u578b\uff1bRMSNorm \/ LayerNorm\uff1b\u4e0d\u540c\u7684\u6fc0\u6d3b\u51fd\u6570\uff08SwiGLU \u7b49\u7b49\uff09\u3002<\/li>\n\n\n\n<li>\u4f4d\u7f6e\u7f16\u7801\uff1aRoPE\uff0c\u4e0d\u540c\u7684\u65cb\u8f6c\u57fa\u6570\u3001\u6700\u5927\u4e0a\u4e0b\u6587\u7a97\u53e3\u8bbe\u7f6e\u3002<\/li>\n\n\n\n<li>\u591a\u6a21\u6001\uff1a\u56fe\u50cf\u7f16\u7801\u5668\u9009\u578b\u4e0d\u4e00\u6837\uff08Qwen\u2011VL \u548c LLaVA \u7684\u89c6\u89c9\u6a21\u5757\u5b8c\u5168\u4e0d\u540c\uff09\u3002<\/li>\n<\/ol>\n\n\n\n<p>\u8fd9\u4e9b\u7ec6\u8282\u5168\u90e8\u5199\u5728 <code>config.json<\/code>\uff0c\u4f60\u6253\u5f00\u4e24\u4e2a\u4e0d\u540c\u6a21\u578b\u7684 config.json \u5bf9\u6bd4\uff0c\u5f88\u591a\u53c2\u6570\u662f\u4e0d\u4e00\u6837\u7684\u3002<\/p>\n\n\n\n<p>\u3000\u3000<\/p>\n\n\n\n<p><strong><strong>\ud83d\udd39<\/strong>3\u3001\u8bad\u7ec3\u914d\u65b9\u4e0d\u4e00\u6837\uff08\u8d85\u53c2\uff09<\/strong><\/p>\n\n\n\n<p>\u5c31\u7b97\u6570\u636e\u3001\u7f51\u7edc\u5b8c\u5168\u76f8\u540c\uff0c\u8bad\u7ec3\u914d\u65b9\u4e0d\u540c\u7ed3\u679c\u4e5f\u4e0d\u540c\uff1a<\/p>\n\n\n\n<ul>\n<li>\u5b66\u4e60\u7387\u591a\u5927<\/li>\n\n\n\n<li>batch \u5927\u5c0f<\/li>\n\n\n\n<li>\u8bad\u7ec3\u591a\u5c11\u4e07\u4ebf token<\/li>\n\n\n\n<li>\u4f18\u5316\u5668\uff08AdamW \u5404\u7c7b\u53d8\u79cd\uff09<\/li>\n\n\n\n<li>dropout\u3001\u6743\u91cd\u8870\u51cf<\/li>\n\n\n\n<li>\u9884\u70ed\u3001\u8c03\u5ea6\u5668\u7b56\u7565<\/li>\n<\/ul>\n\n\n\n<p>\u5c31\u50cf\u540c\u6837\u7684\u9762\u7c89\u6c34\uff0c\u4e0d\u540c\u6e29\u5ea6\u3001\u53d1\u9175\u65f6\u95f4\uff0c\u70e4\u51fa\u6765\u9762\u5305\u53e3\u611f\u5b8c\u5168\u4e0d\u540c\u3002<\/p>\n\n\n\n<p>\u3000\u3000<\/p>\n\n\n\n<p><strong><strong>\ud83d\udd39<\/strong>4\u3001\u9884\u8bad\u7ec3\u4e4b\u540e\uff0c\u540e\u7eed\u5fae\u8c03\u4e0d\u4e00\u6837<\/strong><\/p>\n\n\n\n<p>\u9884\u8bad\u7ec3\u662f\u5b66\u901a\u7528\u77e5\u8bc6\uff1b<\/p>\n\n\n\n<p>\u4e4b\u540e\u8fd8\u8981\u505a\uff1aSFT \u76d1\u7763\u5fae\u8c03\u3001RLHF\/RLAIF \u4eba\u7c7b\u5bf9\u9f50\u3002<\/p>\n\n\n\n<ul>\n<li>\u6709\u7684\u4fa7\u91cd\u5bf9\u8bdd\u53cb\u597d\uff1b<\/li>\n\n\n\n<li>\u6709\u7684\u4fa7\u91cd\u63a8\u7406\uff1b<\/li>\n\n\n\n<li>\u6709\u7684\u504f\u5411\u7f16\u7801\uff1b<\/li>\n\n\n\n<li>\u6709\u7684\u505a\u53bb\u9650\u5236\uff08uncensored\uff09\u793e\u533a\u4e8c\u6b21\u5fae\u8c03\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u5f88\u591a\u793e\u533a\u9b54\u6539\u6a21\u578b\uff08\u6bd4\u5982Qwen3.6\u201127B\u2011Uncensored\uff09\uff1a\u4e3b\u5e72\u6743\u91cd\u6765\u81ea\u5b98\u65b9 Qwen\uff0c\u53ea\u662f\u505a\u4e8c\u6b21\u5fae\u8c03\uff0c\u5c31\u6027\u683c\u5927\u53d8\u3002<\/p>\n\n\n\n<p>\u3000\u3000<\/p>\n\n\n\n<p><strong><strong>\ud83d\udd39<\/strong>5\u3001\u5206\u8bcd\u5668 tokenizer \u4e0d\u4e00\u6837<\/strong><\/p>\n\n\n\n<p>\u540c\u6837\u4e00\u53e5\u4e2d\u6587\uff1a\u4e0d\u540c\u6a21\u578b\u5206\u8bcd\u5668\u5207\u51fa\u6765\u7684 token ID \u5e8f\u5217\u4e0d\u4e00\u6837\u3002<\/p>\n\n\n\n<ul>\n<li>token \u7c92\u5ea6\u4e0d\u540c\uff0c\u4f1a\u5f71\u54cd\u957f\u6587\u672c\u3001\u4e2d\u6587\u7406\u89e3\u3001\u7b26\u53f7\u5904\u7406\u3002<\/li>\n\n\n\n<li>tokenizer \u662f\u72ec\u7acb\u6587\u4ef6\uff0c\u4e0d\u662f\u7f51\u7edc\u6743\u91cd\uff0c\u4f46\u662f\u76f4\u63a5\u5f71\u54cd\u8f93\u5165\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u3000\u3000<\/p>\n\n\n\n<p><strong>\ud83d\udd396\u3001\u5bf9\u9f50\u76ee\u6807\u4e0d\u540c<\/strong><\/p>\n\n\n\n<ul>\n<li>\u6709\u7684\u6a21\u578b\u76ee\u6807\uff1a\u5b89\u5168\u3001\u4fdd\u5b88\uff1b<\/li>\n\n\n\n<li>\u6709\u7684\u6a21\u578b\u76ee\u6807\uff1a\u5c3d\u53ef\u80fd\u8f93\u51fa\u771f\u5b9e\u56de\u7b54\uff0c\u5c11\u62d2\u7edd\uff1b<\/li>\n\n\n\n<li>\u6709\u7684\u4f18\u5148\u6570\u5b66\u63a8\u7406\uff1b<\/li>\n\n\n\n<li>\u6709\u7684\u4f18\u5148\u4ee3\u7801\u3002<\/li>\n\n\n\n<li>\u5bf9\u9f50\u9636\u6bb5\u5c31\u5851\u9020\u4e86\u6a21\u578b\u7684 \u201c\u6027\u683c\u201d\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u3000\u3000<\/p>\n\n\n\n<p><strong>\ud83d\udd39<\/strong>7.<strong>\u68b3\u7406\u6574\u6761\u94fe\u8def<\/strong><\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"generic\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\u539f\u59cb\u8bed\u6599\u5e93A\uff08Qwen\uff09 \/ \u8bed\u6599\u5e93B\uff08Llama\uff09\n        \u2193\n\u6e05\u6d17\u8fc7\u6ee4\uff08\u6bcf\u5bb6\u7b56\u7565\u4e0d\u4e00\u6837\uff09\n        \u2193\ntokenizer\u5206\u8bcd\uff0c\u8f6c\u4e3a\u5f20\u91cf\u8f93\u5165PyTorch\n        \u2193\nTransformer\u7f51\u7edc\uff08\u7ed3\u6784\u7ec6\u8282\u3001\u89c4\u6a21\u5404\u5bb6\u6709\u5dee\u5f02\uff09\n        \u2193\n\u8bad\u7ec3\u8d85\u53c2\u914d\u65b9\u4e0d\u540c\uff0c\u53cd\u590d\u66f4\u65b0safetensors\u6743\u91cd\n        \u2193\n\u9884\u8bad\u7ec3\u57fa\u7840\u6a21\u578b\n        \u2193\nSFT \/ RLHF\u5fae\u8c03\uff08\u5404\u5bb6\u5fae\u8c03\u6570\u636e\u96c6\u3001\u76ee\u6807\u4e0d\u4e00\u6837\uff09\n        \u2193\n\u6700\u7ec8\u4ea7\u51fa config + safetensors\u6743\u91cd<\/pre>\n\n\n\n<p>\u5168\u90e8\u90fd\u8dd1\u5728 PyTorch \u4e0a\uff0c\u4f46<strong>\u6570\u636e\u3001\u7ed3\u6784\u7ec6\u8282\u3001\u8bad\u7ec3\u914d\u65b9\u3001\u5fae\u8c03\uff0c\u6bcf\u4e00\u5904\u90fd\u53ef\u4ee5\u62c9\u5f00\u5dee\u8ddd<\/strong>\u3002<\/p>\n\n\n\n<p>Transformer \u53ea\u662f\u4e00\u5957<strong>\u8ba1\u7b97\u9aa8\u67b6<\/strong>\uff0c\u771f\u6b63\u7684\u77e5\u8bc6\u5168\u90e8\u5b58\u5728<code>safetensors<\/code>\u91cc\u9762\u90a3\u51e0\u5341\u4ebf\u4e2a\u6d6e\u70b9\u6570\u6743\u91cd\u3002 \u8bad\u7ec3\u7684\u672c\u8d28\uff0c\u5c31\u662f\u4e0d\u65ad\u8c03\u6574\u8fd9\u5806\u6570\u5b57\u3002 \u4e0d\u540c\u6570\u636e\u3001\u4e0d\u540c\u8bad\u7ec3\u8fc7\u7a0b\uff0c\u6700\u7ec8\u6536\u655b\u5f97\u5230\u7684\u90a3\u5806\u6d6e\u70b9\u6570\u5b8c\u5168\u4e0d\u4e00\u6837\u3002<\/p>\n\n\n\n<p>\u3000\u3000<\/p>\n\n\n\n<p><strong>\ud83d\udd398.\u4e3e\u4e2a\u73b0\u5b9e\u4f8b\u5b50<\/strong><\/p>\n\n\n\n<p>Qwen\u201127B \u548c Llama3\u201127B\uff1a<\/p>\n\n\n\n<ul>\n<li>\u90fd\u7528 PyTorch \u8bad\u7ec3<\/li>\n\n\n\n<li>\u90fd\u662f Transformer\uff0c\u8f93\u5165\u90fd\u662f token \u5f20\u91cf \u4f46\u662f\uff1a<\/li>\n<\/ul>\n\n\n\n<ol>\n<li>\u8bad\u7ec3\u6570\u636e\u4e00\u4e2a\u4fa7\u91cd\u4e2d\u6587\uff0c\u4e00\u4e2a\u82f1\u6587\u4e3a\u4e3b<\/li>\n\n\n\n<li>config \u7f51\u7edc\u53c2\u6570\u7ec6\u8282\u6709\u5dee\u5f02<\/li>\n\n\n\n<li>\u5206\u8bcd\u5668\u5b8c\u5168\u4e0d\u540c<\/li>\n\n\n\n<li>\u5fae\u8c03\u5bf9\u9f50\u6570\u636e\u96c6\u4e0d\u4e00\u6837 \u2192 \u5b9e\u9645\u5bf9\u8bdd\u3001\u5199\u4e2d\u6587\u3001\u63a8\u7406\u80fd\u529b\u8868\u73b0\u660e\u663e\u4e0d\u4e00\u6837\u3002<\/li>\n<\/ol>\n\n\n\n<p>\u3000\u3000<\/p>\n\n\n\n<p><strong>Mac M \u7cfb\u5217 PyTorch MPS \u5b8c\u6574\u5c0f\u793a\u4f8b \ud83d\udd16<\/strong><\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p>PyTorch MPS\uff0c <strong>MPS = Metal Performance Shaders<\/strong>\uff0c\u662f\u82f9\u679c\u7ed9 M \u7cfb\u5217\u82af\u7247\uff08M1\/M2\/M3\/M4\/M5\uff09\u505a\u7684<strong>GPU \u52a0\u901f\u540e\u7aef<\/strong>\uff0c\u5bf9\u6807 N \u5361\u7684 CUDA\u3001AMD \u7684 ROCm\u3002<\/p>\n\n\n\n<pre class=\"wp-block-preformatted\">CUDA \u2192 Nvidia \u663e\u5361<br>ROCm \u2192 AMD \u663e\u5361<br>MPS \u2192 Apple Silicon\uff08M \u82af\u7247 Mac\uff09<\/pre>\n\n\n\n<p>Metal \u662f\u82f9\u679c\u5e95\u5c42 GPU \u7f16\u7a0b API\uff1bMPS \u662f\u6784\u5efa\u5728 Metal \u4e4b\u4e0a\uff0c\u4e13\u95e8\u505a\u673a\u5668\u5b66\u4e60\u77e9\u9635\u8fd0\u7b97\u7684\u9ad8\u6027\u80fd\u8ba1\u7b97\u5e93<\/p>\n\n\n\n<p>PyTorch \u4ece <strong>1.12 \u7248\u672c\u5f00\u59cb\u5b98\u65b9\u5185\u7f6e MPS \u540e\u7aef<\/strong>\uff0c\u53ef\u4ee5\u628a\u5f20\u91cf\u3001\u795e\u7ecf\u7f51\u7edc\u8fd0\u7b97\u4ea4\u7ed9 Mac \u5185\u7f6e GPU \u8dd1\uff0c\u4e0d\u518d\u53ea\u9760 CPU \u6162\u541e\u541e\u8ba1\u7b97<\/p>\n\n\n\n<p><strong>\u65b0\u5efa\u865a\u62df\u73af\u5883<\/strong><\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"generic\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">python -m venv python-venv<\/pre>\n\n\n\n<p><strong>\u6fc0\u6d3b\u865a\u62df\u73af\u5883<\/strong><\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"generic\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">source python-venv\/bin\/activate<\/pre>\n\n\n\n<p><strong>\u5b89\u88c5torch<\/strong><\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"generic\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">python -m pip install torch    <\/pre>\n\n\n\n<p><strong>\u8fd0\u884c\u4ee3\u7801<\/strong><\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"generic\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">python demo_torch_mps.py<\/pre>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"generic\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">import torch\n\n# 1. \u5224\u65ad\u8bbe\u5907\uff1a\u4f18\u5148MPS\uff0c\u6ca1\u6709\u5c31\u56de\u9000CPU\nif torch.backends.mps.is_available():\n    device = torch.device(\"mps\")\n    print(\"\u2705 \u4f7f\u7528 MPS (Metal GPU)\")\nelse:\n    device = torch.device(\"cpu\")\n    print(\"\u26a0\ufe0f MPS\u4e0d\u53ef\u7528\uff0c\u4f7f\u7528 CPU\")\n\n# 2. \u521b\u5efa\u5f20\u91cf\uff0c\u642c\u8fd0\u5230GPU(MPS)\n# 3\u884c\uff0c5\u5217\u7684\u968f\u673a\u6570\u5b57\u5f20\u91cf\nx = torch.randn(3, 5).to(device)\ny = torch.randn(5, 4).to(device)\n\nprint(\"\\nx \u5f20\u91cf shape:\", x.shape)\nprint(\"y \u5f20\u91cf shape:\", y.shape)\nprint(\"x \u7684\u8bbe\u5907:\", x.device)\n\n# 3. GPU\u4e0a\u505a\u77e9\u9635\u4e58\u6cd5\uff08AI\u6700\u6838\u5fc3\u8fd0\u7b97\uff09\nz = torch.matmul(x, y)\nprint(\"\\n\u77e9\u9635\u4e58\u6cd5\u7ed3\u679cz shape:\", z.shape)\nprint(z)\n\n# ----------------------\n# \u6f14\u793a\uff1a\u63a8\u7406(\u53ea\u524d\u5411) vs \u8bad\u7ec3(\u524d\u5411+\u53cd\u5411\u6c42\u68af\u5ea6)\n# ----------------------\n# \u7b80\u5355\u795e\u7ecf\u7f51\u7edc\nnet = torch.nn.Linear(4, 2).to(device)\nout = net(z)\n\nloss = out.sum()\nloss.backward()   # \u53cd\u5411\u4f20\u64ad\uff0c\u8bad\u7ec3\u624d\u4f1a\u7528\u5230\uff01MPS\u652f\u6301\u53cd\u5411\n\nprint(\"\\n\u2705 \u68af\u5ea6\u8ba1\u7b97\u5b8c\u6210\uff0c\u6743\u91cd\u5b58\u5728\u68af\u5ea6\uff1a\")\nprint(net.weight.grad is not None)  # True\u4ee3\u8868\u53cd\u5411\u6c42\u5bfc\u6b63\u5e38\u5de5\u4f5c\n\n# 4. \u628aMPS\u4e0a\u7684\u5f20\u91cf\u62ff\u56de\u5230CPU\uff0c\u624d\u80fd\u6253\u5370\/numpy\u5904\u7406\nz_cpu = z.to(\"cpu\")\nprint(\"\\n\u8f6c\u56deCPU\u540e\u7684\u5f20\u91cf\u8bbe\u5907\uff1a\", z_cpu.device)<\/pre>\n\n\n\n<p><strong>\u8fd0\u884c\u7ed3\u679c<\/strong><\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"generic\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\u2705 \u4f7f\u7528 MPS (Metal GPU)\n\nx \u5f20\u91cf shape: torch.Size([3, 5])\ny \u5f20\u91cf shape: torch.Size([5, 4])\nx \u7684\u8bbe\u5907: mps:0\n\n\u77e9\u9635\u4e58\u6cd5\u7ed3\u679cz shape: torch.Size([3, 4])\ntensor([[-0.1038,  1.3596,  2.0733,  3.7567],\n        [ 3.2019, -0.1455, -1.4333,  3.9464],\n        [ 2.4885, -2.1966, -0.9626,  1.0828]], device='mps:0')\n\n\u2705 \u68af\u5ea6\u8ba1\u7b97\u5b8c\u6210\uff0c\u6743\u91cd\u5b58\u5728\u68af\u5ea6\uff1a\nTrue\n\n\u8f6c\u56deCPU\u540e\u7684\u5f20\u91cf\u8bbe\u5907\uff1a cpu\n<\/pre>\n\n\n\n<p><strong>\u5173\u952e\u77e5\u8bc6\u70b9<\/strong><\/p>\n\n\n\n<ul>\n<li>.to(device)\uff1a\u5f20\u91cf\u642c\u5bb6\uff0c\u653e\u5230 MPS (GPU \u7edf\u4e00\u5185\u5b58)\uff0c\u540e\u7eed\u8fd0\u7b97\u8d70 Metal \u52a0\u901f\u3002<\/li>\n\n\n\n<li>loss.backward()\uff1a\u53cd\u5411\u4f20\u64ad\uff0c\u8bad\u7ec3\u624d\u7528\uff1bComfyUI \u751f\u56fe\u53ea\u505a\u524d\u5411\uff0c\u4e0d\u9700\u8981\u8fd9\u884c\u3002<\/li>\n\n\n\n<li>MPS \u4e0a\u7684\u5f20\u91cf\u4e0d\u80fd\u76f4\u63a5\u8f6c numpy\uff0c\u5fc5\u987b\u5148 .to(&#8220;cpu&#8221;)\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u5982\u679c\u9047\u5230\u90e8\u5206\u7b97\u5b50 MPS \u4e0d\u652f\u6301\u3001\u62a5\u9519\uff0c\u53ef\u4ee5\u5f00\u542f\u964d\u7ea7\u73af\u5883\u53d8\u91cf\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"generic\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">export PYTORCH_ENABLE_MPS_FALLBACK=1\npython demo_torch_mps.py<\/pre>\n","protected":false},"excerpt":{"rendered":"<p>\u524d\u8a00 \ud83d\udd16 PyTorch \u662f\u4e00\u4e2a Python \u8bed\u8a00\u7684\u6df1\u5ea6\u5b66\u4e60\u5de5\u5177\u7bb1\uff08AI \u6846\u67b6\uff09\uff0cPyTorch \u6700\u521d\u7531 M [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[48,37],"tags":[],"_links":{"self":[{"href":"https:\/\/92it.top\/index.php?rest_route=\/wp\/v2\/posts\/21251"}],"collection":[{"href":"https:\/\/92it.top\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/92it.top\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/92it.top\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/92it.top\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=21251"}],"version-history":[{"count":9,"href":"https:\/\/92it.top\/index.php?rest_route=\/wp\/v2\/posts\/21251\/revisions"}],"predecessor-version":[{"id":21258,"href":"https:\/\/92it.top\/index.php?rest_route=\/wp\/v2\/posts\/21251\/revisions\/21258"}],"wp:attachment":[{"href":"https:\/\/92it.top\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=21251"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/92it.top\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=21251"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/92it.top\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=21251"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}