{"id":35510,"date":"2024-12-26T09:00:02","date_gmt":"2024-12-26T01:00:02","guid":{"rendered":"https:\/\/17aitech.com\/?p=35510"},"modified":"2024-12-26T09:00:02","modified_gmt":"2024-12-26T01:00:02","slug":"%e5%a4%a9%e4%b8%8b%e8%8b%a6%e8%8b%b1%e4%bc%9f%e8%be%be%e4%b9%85%e7%9f%a3%ef%bc%81pytorch%e5%ae%98%e6%96%b9%e5%85%8dcuda%e5%8a%a0%e9%80%9f%e6%8e%a8%e7%90%86%ef%bc%8ctriton%e6%97%b6%e4%bb%a3%e8%a6%81","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=35510","title":{"rendered":"\u5929\u4e0b\u82e6\u82f1\u4f1f\u8fbe\u4e45\u77e3\uff01PyTorch\u5b98\u65b9\u514dCUDA\u52a0\u901f\u63a8\u7406\uff0cTriton\u65f6\u4ee3\u8981\u6765\uff1f"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-09-09-9\" target=\"_blank\">\u5929\u4e0b\u82e6\u82f1\u4f1f\u8fbe\u4e45\u77e3\uff01PyTorch\u5b98\u65b9\u514dCUDA\u52a0\u901f\u63a8\u7406\uff0cTriton\u65f6\u4ee3\u8981\u6765\uff1f<\/a><\/p>\n<blockquote data-author-name=\"\" data-content-utf8-length=\"76\" data-source-title=\"\" data-type=\"2\" data-url=\"\"><p>\u8fd1\u65e5\uff0cPyTorch \u5b98\u65b9\u5206\u4eab\u4e86\u5982\u4f55\u5b9e\u73b0\u65e0 CUDA \u8ba1\u7b97\uff0c\u5bf9\u5404\u4e2a\u5185\u6838\u8fdb\u884c\u4e86\u5fae<mark data-type=\"concepts\" data-id=\"308c3a45-0fee-4ec6-858e-85b15f440fc0\">\u57fa\u51c6<\/mark>\u6d4b\u8bd5\u6bd4\u8f83\uff0c\u5e76\u8ba8\u8bba\u4e86\u672a\u6765\u5982\u4f55\u8fdb\u4e00\u6b65\u6539\u8fdb Triton \u5185\u6838\u4ee5\u7f29\u5c0f\u4e0e CUDA \u7684\u5dee\u8ddd\u3002<\/p><\/blockquote>\n<p>\u5728\u505a\u5927<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\uff08LLM\uff09\u7684\u8bad\u7ec3\u3001\u5fae\u8c03\u548c\u63a8\u7406\u65f6\uff0c\u4f7f\u7528\u82f1\u4f1f\u8fbe\u7684 GPU \u548c CUDA \u662f\u5e38\u89c1\u7684\u505a\u6cd5\u3002\u5728\u66f4\u5927\u7684<mark data-type=\"tech_methods\" data-id=\"1a0e9c5e-6502-4cd7-8683-6b5ca6c48be2\">\u673a\u5668\u5b66\u4e60<\/mark>\u7f16\u7a0b\u4e0e\u8ba1\u7b97\u8303\u7574\uff0c\u540c\u6837\u4e25\u91cd\u4f9d\u8d56 CUDA\uff0c\u4f7f\u7528\u5b83\u52a0\u901f\u7684<mark data-type=\"tech_methods\" data-id=\"1a0e9c5e-6502-4cd7-8683-6b5ca6c48be2\">\u673a\u5668\u5b66\u4e60<\/mark>\u6a21\u578b\u53ef\u4ee5\u5b9e\u73b0\u66f4\u5927\u7684\u6027\u80fd\u63d0\u5347\u3002<\/p>\n<p>\u867d\u7136 CUDA \u5728\u52a0\u901f\u8ba1\u7b97\u9886\u57df\u5360\u636e\u4e3b\u5bfc\u5730\u4f4d\uff0c\u5e76\u6210\u4e3a\u82f1\u4f1f\u8fbe\u91cd\u8981\u7684\u62a4\u57ce\u6cb3\u4e4b\u4e00\u3002\u4f46\u5176\u4ed6\u4e00\u4e9b\u5de5\u4f5c\u7684\u51fa\u73b0\u6b63\u5728\u5411 CUDA \u53d1\u8d77\u6311\u6218\uff0c\u6bd4\u5982 OpenAI \u63a8\u51fa\u7684 Triton\uff0c\u5b83\u5728\u53ef\u7528\u6027\u3001\u5185\u5b58\u5f00\u9500\u3001AI \u7f16\u8bd1\u5668\u5806\u6808\u6784\u5efa\u7b49\u65b9\u9762\u5177\u6709\u4e00\u5b9a\u7684\u4f18\u52bf\uff0c\u5e76\u6301\u7eed\u5f97\u5230\u53d1\u5c55\u3002<\/p>\n<p>\u8fd1\u65e5\uff0cPyTorch \u5b98\u5ba3\u8981\u505a\u300c\u65e0\u82f1\u4f1f\u8fbe CUDA \u53c2\u4e0e\u7684\u5927\u6a21\u578b\u63a8\u7406\u300d\u3002\u5728\u8c08\u5230\u4e3a\u4ec0\u4e48\u8981 100% \u4f7f\u7528 Triton \u8fdb\u884c\u63a2\u7d22\u65f6\uff0cPyTorch \u8868\u793a\uff1a\u300cTriton \u63d0\u4f9b\u4e86\u4e00\u6761\u9014\u5f84\uff0c\u4f7f\u5927\u6a21\u578b \u80fd\u591f\u5728\u4e0d\u540c\u7c7b\u578b\u7684 GPU \u4e0a\u8fd0\u884c\uff0c\u5305\u62ec\u82f1\u4f1f\u8fbe\u3001<mark data-type=\"institutions\" data-id=\"59257b4b-235a-4864-a37b-956cede3debf\">AMD<\/mark>\u3001<mark data-type=\"institutions\" data-id=\"3437dab4-c82a-453e-8dd9-46a27001ab4f\">\u82f1\u7279\u5c14<\/mark>\u548c\u5176\u4ed6\u57fa\u4e8e GPU \u7684\u52a0\u901f\u5668\u3002<\/p>\n<p>\u6b64\u5916 Triton \u8fd8\u5728 Python \u4e2d\u4e3a GPU \u7f16\u7a0b\u63d0\u4f9b\u4e86\u66f4\u9ad8\u7684\u62bd\u8c61\u5c42\uff0c\u4f7f\u5f97\u4f7f\u7528 PyTorch \u80fd\u591f\u6bd4\u4f7f\u7528\u4f9b\u5e94\u5546\u7279\u5b9a\u7684 API \u66f4\u5feb\u5730\u7f16\u5199\u9ad8\u6027\u80fd\u5185\u6838\u3002\u300d<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-9d63e2ec014c249f81a2c2bde0204917.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-9d63e2ec014c249f81a2c2bde0204917.png\"><\/a><\/p>\n<p>\u5728 PyTorch \u535a\u5ba2\u4e2d\u8ba8\u8bba\u4e86\u4f7f\u7528\u6d41\u884c\u7684 LLM \u6a21\u578b\uff08\u4f8b\u5982 Meta \u7684 Llama3-8B \u548c <mark data-type=\"institutions\" data-id=\"b86983cf-b858-45b0-86a1-b05cb94cde08\">IBM<\/mark> \u7684 Granite-8B Code\uff09\u5b9e\u73b0 FP16 \u63a8\u7406\u7684\u65b9\u6cd5\uff0c\u5176\u4e2d\u8ba1\u7b97\u662f 100% \u4f7f\u7528 OpenAI \u7684 Triton \u8bed\u8a00\u6267\u884c\u7684\u3002<\/p>\n<p>\u5bf9\u4e8e\u4f7f\u7528\u57fa\u4e8e Triton \u5185\u6838\u7684\u6a21\u578b\u751f\u6210\u5355\u4e2a token \u7684\u65f6\u95f4\uff0cPyTorch \u80fd\u591f\u5b9e\u73b0\u5728\u82f1\u4f1f\u8fbe H100 GPU \u4e0a Llama \u548c Granite \u7684 CUDA \u5185\u6838\u4e3b\u5bfc\u5de5\u4f5c\u6d41\u7a0b\u7684 0.76-0.78 \u500d\u6027\u80fd\uff0c\u4ee5\u53ca\u5728\u82f1\u4f1f\u8fbe A100 GPU \u4e0a\u7684 0.62-0.82 \u500d\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-37cebe650136bfd5767dc620d481d527.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-37cebe650136bfd5767dc620d481d527.png\"><\/a><\/p>\n<p><em><sup>\u56fe 1. \u5728\u82f1\u4f1f\u8fbe H100 \u548c A100 \u4e0a\uff0cLlama3-8B \u548c Granite-8B \u7684 Triton \u548c CUDA \u53d8\u4f53\u7684\u63a8\u7406\u541e\u5410\u91cf\u6bd4\u8f83\u3002\u8bbe\u7f6e\uff1a\u6279\u5927\u5c0f = 2\uff0c\u8f93\u5165\u5e8f\u5217\u957f\u5ea6 = 512\uff0c\u8f93\u51fa\u5e8f\u5217\u957f\u5ea6 = 256<\/sup><br \/><\/em><\/p>\n<p>\u4e5f\u8bb8\u544a\u522b\u82f1\u4f1f\u8fbe\u7684\u65f6\u5019\u771f\u8981\u6765\u4e86\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-a6f4c046126d1d6f9647bea292e6236b.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-a6f4c046126d1d6f9647bea292e6236b.png\"><\/a><\/p>\n<p><strong>Transformer \u5757\u7684\u7ec4\u6210<\/strong><\/p>\n<p>PyTorch \u56e2\u961f\u9996\u5148\u5bf9\u57fa\u4e8e Transformer \u7684\u6a21\u578b\u4e2d\u53d1\u751f\u7684\u8ba1\u7b97\u8fdb\u884c\u7ec6\u5206\u3002\u4e0b\u56fe\u663e\u793a\u4e86\u5178\u578b Transformer \u5757\u7684\u300c\u5185\u6838\uff08kernel\uff09\u300d\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-86fc03b3a2dbd8935d12796f575de610.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-86fc03b3a2dbd8935d12796f575de610.png\"><\/a><\/p>\n<p><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u56fe 2<\/sup><\/em><\/p>\n<p>Llama3 \u67b6\u6784\u7684\u6838\u5fc3\u64cd\u4f5c\u603b\u7ed3\u5982\u4e0b\uff1a<\/p>\n<ul>\n<li>\n<p>\u5747\u65b9\u6839\u5f52\u4e00\u5316\uff08RMSNorm\uff09<\/p>\n<\/li>\n<li>\n<p>\u77e9\u9635\u4e58\u6cd5\uff1aFused QKV<\/p>\n<\/li>\n<li>\n<p>RoPE<\/p>\n<\/li>\n<li>\n<p>\u6ce8\u610f\u529b<\/p>\n<\/li>\n<li>\n<p>\u77e9\u9635\u4e58\u6cd5\uff1a\u8f93\u51fa\u6295\u5f71<\/p>\n<\/li>\n<li>\n<p>RMSNorm<\/p>\n<\/li>\n<li>\n<p>\u77e9\u9635\u4e58\u6cd5\uff1aFused Gate + Up Projection<\/p>\n<\/li>\n<li>\n<p><mark data-type=\"concepts\" data-id=\"1697e627-30e7-48a6-b799-39e2338ffab5\">\u6fc0\u6d3b\u51fd\u6570<\/mark>\uff1aSiLU<\/p>\n<\/li>\n<li>\n<p>\u70b9\u4e58\uff08Element Wise Multiplication\uff09<\/p>\n<\/li>\n<li>\n<p>\u77e9\u9635\u4e58\u6cd5\uff1aDown Projection<\/p>\n<\/li>\n<\/ul>\n<p>\u8fd9\u4e9b\u64cd\u4f5c\u4e2d\u7684\u6bcf\u4e00\u4e2a\u90fd\u662f\u901a\u8fc7\u5728 GPU \u4e0a\u6267\u884c\u4e00\u4e2a\uff08\u6216\u591a\u4e2a\uff09\u5185\u6838\u6765\u8ba1\u7b97\u7684\u3002\u867d\u7136\u6bcf\u4e2a\u5185\u6838\u7684\u7ec6\u8282\u5728\u4e0d\u540c\u7684 Transformer \u6a21\u578b\u4e2d\u53ef\u80fd\u6709\u6240\u4e0d\u540c\uff0c\u4f46\u6838\u5fc3\u64cd\u4f5c\u4fdd\u6301\u4e0d\u53d8\u3002\u4f8b\u5982\uff0c<mark data-type=\"institutions\" data-id=\"b86983cf-b858-45b0-86a1-b05cb94cde08\">IBM<\/mark> \u7684 Granite 8B Code \u6a21\u578b\u5728 MLP \u5c42\u4e2d\u4f7f\u7528\u504f\u7f6e\uff0c\u4e0e Llama3 \u4e0d\u540c\u3002\u6b64\u7c7b\u66f4\u6539\u786e\u5b9e\u9700\u8981\u5bf9\u5185\u6838\u8fdb\u884c\u4fee\u6539\u3002\u5178\u578b\u7684\u6a21\u578b\u662f\u8fd9\u4e9b Transformer \u5757\u7684<mark data-type=\"tech_methods\" data-id=\"efb128f3-0b7c-45a8-a6ae-065fa93c6556\">\u5806\u53e0<\/mark>\uff0c\u8fd9\u4e9b Transformer \u5757\u901a\u8fc7\u5d4c\u5165\u5c42\u8fde\u63a5\u5728\u4e00\u8d77\u3002<\/p>\n<p><strong>\u6a21\u578b\u63a8\u7406<\/strong><\/p>\n<p>\u5178\u578b\u7684\u6a21\u578b\u67b6\u6784\u4ee3\u7801\u4e0e PyTorch \u542f\u52a8\u7684 python model.py \u6587\u4ef6\u5171\u4eab\u3002\u5728\u9ed8\u8ba4\u7684 PyTorch Eager Execution \u6a21\u5f0f\u4e0b\uff0c\u8fd9\u4e9b\u5185\u6838\u90fd\u662f\u4f7f\u7528 CUDA \u6267\u884c\u7684\u3002\u4e3a\u4e86\u5b9e\u73b0 100% Triton \u8fdb\u884c\u7aef\u5230\u7aef Llama3-8B \u548c Granite-8B \u63a8\u7406\uff0c\u9700\u8981\u7f16\u5199\u548c\u96c6\u6210\u624b\u5199 Triton \u5185\u6838\u4ee5\u53ca\u5229\u7528 torch.compile\uff08\u751f\u6210 Triton \u64cd\u4f5c\uff09\u3002\u9996\u5148\uff0cPyTorch \u7528\u7f16\u8bd1\u5668\u751f\u6210\u7684 Triton \u5185\u6838\u66ff\u6362\u8f83\u5c0f\u7684\u64cd\u4f5c\uff0c\u5176\u6b21\uff0cPyTorch \u7528\u624b\u5199\u7684 Triton \u5185\u6838\u66ff\u6362\u66f4\u6602\u8d35\u548c\u590d\u6742\u7684\u8ba1\u7b97\uff08\u4f8b\u5982\u77e9\u9635\u4e58\u6cd5\u548c\u95ea\u5b58\u6ce8\u610f\u529b\uff09\u3002<\/p>\n<p>Torch.compile \u81ea\u52a8\u4e3a RMSNorm\u3001RoPE\u3001SiLU \u548c\u70b9\u4e58\u751f\u6210 Triton \u5185\u6838\u3002\u4f7f\u7528 Nsight Systems \u7b49\u5de5\u5177\uff0c\u53ef\u4ee5\u89c2\u5bdf\u5230\u8fd9\u4e9b\u751f\u6210\u7684\u5185\u6838\uff0c\u5b83\u4eec\u5728\u77e9\u9635\u4e58\u6cd5\u548c\u6ce8\u610f\u529b\u4e4b\u95f4\u8868\u73b0\u4e3a\u5fae\u5c0f\u7684\u6df1\u7eff\u8272\u5185\u6838\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-613431f7202ebbdde37a4f38b800577d.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-613431f7202ebbdde37a4f38b800577d.png\"><\/a><\/p>\n<p><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u56fe 3. \u4f7f\u7528 torch.compile \u8ddf\u8e2a Llama3-8B\uff0c\u663e\u793a\u7528\u4e8e\u77e9\u9635\u4e58\u6cd5\u548c\u95ea\u5b58\u6ce8\u610f\u529b\u7684 CUDA \u5185\u6838\u3002<\/sup><\/em><\/p>\n<p>\u5bf9\u4e8e\u4e0a\u9762\u7684\u8ddf\u8e2a\uff0cPyTorch \u56e2\u961f\u6ce8\u610f\u5230\uff0c\u5728 Llama3-8B \u6837\u5f0f\u6a21\u578b\u4e2d\uff0c\u5360 E2E \u5ef6\u8fdf 80% \u7684\u4e24\u4e2a\u4e3b\u8981\u64cd\u4f5c\u662f\u77e9\u9635\u4e58\u6cd5\u548c\u6ce8\u610f\u529b\u5185\u6838\uff0c\u5e76\u4e14\u4e24\u8005\u4ecd\u7136\u662f CUDA \u5185\u6838\u3002\u56e0\u6b64\uff0c\u4e3a\u4e86\u5f25\u8865\u5269\u4f59\u7684\u5dee\u8ddd\uff0cPyTorch \u56e2\u961f\u7528\u624b\u5199\u7684 Triton \u5185\u6838\u66ff\u6362\u4e86 matmul \u548c\u6ce8\u610f\u529b\u5185\u6838\u3002<\/p>\n<p><strong>Triton SplitK GEMM \u5185\u6838<\/strong><\/p>\n<p>\u5bf9\u4e8e\u7ebf\u6027\u5c42\u4e2d\u7684\u77e9\u9635\u4e58\u6cd5\uff0cPyTorch \u56e2\u961f\u7f16\u5199\u4e86\u4e00\u4e2a\u81ea\u5b9a\u4e49 FP16 Triton GEMM\uff08\u901a\u7528\u77e9\u9635 &#8211; \u77e9\u9635\u4e58\u6cd5\uff09\u5185\u6838\uff0c\u8be5\u5185\u6838\u5229\u7528\u4e86 SplitK \u5de5\u4f5c\u5206\u89e3\u3002<\/p>\n<p><strong>GEMM \u5185\u6838\u8c03\u4f18<\/strong><\/p>\n<p>\u4e3a\u4e86\u5b9e\u73b0\u6700\u4f73\u6027\u80fd\uff0cPyTorch \u56e2\u961f\u4f7f\u7528\u7a77\u4e3e\u641c\u7d22\u65b9\u6cd5\u6765\u8c03\u6574 SplitK GEMM \u5185\u6838\u3002Granite-8B \u548c Llama3-8B \u5177\u6709\u5982\u4e0b\u5f62\u72b6\u7684\u7ebf\u6027\u5c42\uff1a<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-952c1079251f3b10fa565c1802944a76.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-952c1079251f3b10fa565c1802944a76.png\"><\/a><\/p>\n<p><em>\u56fe 4. Granite-8B \u548c Llama3-8B \u7ebf\u6027\u5c42<mark data-type=\"concepts\" data-id=\"149a12cf-10c2-4555-9899-cc6dee319ef5\">\u6743\u91cd<\/mark>\u77e9\u9635\u5f62\u72b6\u3002<br \/><\/em><\/p>\n<p>\u6bcf\u4e2a\u7ebf\u6027\u5c42\u90fd\u6709\u4e0d\u540c\u7684<mark data-type=\"concepts\" data-id=\"149a12cf-10c2-4555-9899-cc6dee319ef5\">\u6743\u91cd<\/mark>\u77e9\u9635\u5f62\u72b6\u3002\u56e0\u6b64\uff0c\u4e3a\u4e86\u83b7\u5f97\u6700\u4f73\u6027\u80fd\uff0c\u5fc5\u987b\u9488\u5bf9\u6bcf\u4e2a\u5f62\u72b6\u8f6e\u5ed3\u8c03\u6574 Triton \u5185\u6838\u3002\u5728\u5bf9\u6bcf\u4e2a\u7ebf\u6027\u5c42\u8fdb\u884c\u8c03\u6574\u540e\uff0cPyTorch \u80fd\u591f\u5728 Llama3-8B \u548c Granite-8B \u4e0a\u5b9e\u73b0\u76f8\u5bf9\u4e8e\u672a\u8c03\u6574\u7684 Triton \u5185\u6838 1.20 \u500d\u7684 E2E \u52a0\u901f\u3002<\/p>\n<p><strong>Flash Attention \u5185\u6838<\/strong><\/p>\n<p>PyTorch \u56e2\u961f\u4f7f\u7528\u4e0d\u540c\u7684\u914d\u7f6e\uff0c\u5bf9\u73b0\u6709 Triton flash attention \u5185\u6838\u8fdb\u884c\u4e86\u8bc4\u4f30\uff0c\u5305\u62ec<\/p>\n<ul>\n<li>\n<p><mark data-type=\"institutions\" data-id=\"59257b4b-235a-4864-a37b-956cede3debf\">AMD<\/mark> Flash<\/p>\n<\/li>\n<li>\n<p>OpenAI Flash<\/p>\n<\/li>\n<li>\n<p>Dao AI Lab Flash<\/p>\n<\/li>\n<li>\n<p>XFormers Flash<\/p>\n<\/li>\n<li>\n<p>PyTorch FlexAttention<\/p>\n<\/li>\n<\/ul>\n<p>PyTorch \u56e2\u961f\u5206\u522b\u5728 eager \u6a21\u5f0f\u548c\u7f16\u8bd1\u6a21\u5f0f\u4e0b\u8bc4\u4f30\u4e86\u6bcf\u4e2a\u5185\u6838\u7684<mark data-type=\"tech_tasks\" data-id=\"f02bb483-1c90-4010-849a-f2d7e3f0ad59\">\u6587\u672c\u751f\u6210<\/mark>\u8d28\u91cf\u3002\u4e0b\u56fe 5 \u4e3a\u4e0d\u540c Flash Attention \u5185\u6838\u7684\u6bd4\u8f83\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-0d6f338a34f2acca2a245484370dd6fd.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-0d6f338a34f2acca2a245484370dd6fd.png\"><\/a><\/p>\n<p>\u4e0a\u56fe\u603b\u7ed3\u4e86 PyTorch \u89c2\u5bdf\u5230\u7684\u5f00\u7bb1\u5373\u7528\u60c5\u51b5\uff0c\u5e76\u9884\u8ba1\u5185\u6838 2 \u5230 5 \u53ef\u4ee5\u5728\u4fee\u6539\u540e\u6ee1\u8db3\u4e0a\u8ff0\u6807\u51c6\u3002\u4e0d\u8fc7\u8fd9\u4e5f\u8868\u660e\uff0c\u62e5\u6709\u4e00\u4e2a\u53ef\u7528\u4e8e<mark data-type=\"concepts\" data-id=\"308c3a45-0fee-4ec6-858e-85b15f440fc0\">\u57fa\u51c6<\/mark>\u6d4b\u8bd5\u7684\u5185\u6838\u901a\u5e38\u53ea\u662f\u5c06\u5b83\u7528\u4f5c\u7aef\u5230\u7aef\u751f\u4ea7\u5185\u6838\u7684\u5f00\u59cb\u3002<\/p>\n<p>PyTorch \u56e2\u961f\u9009\u62e9\u5728\u540e\u7eed\u6d4b\u8bd5\u4e2d\u4f7f\u7528 <mark data-type=\"institutions\" data-id=\"59257b4b-235a-4864-a37b-956cede3debf\">AMD<\/mark> flash attention \u5185\u6838\uff0c\u5b83\u901a\u8fc7 torch.compile \u8fdb\u884c\u7f16\u8bd1\uff0c\u5e76\u5728 eager \u548c\u7f16\u8bd1\u6a21\u5f0f\u4e0b\u4ea7\u751f\u6e05\u6670\u7684\u8f93\u51fa\u3002<\/p>\n<p>\u4e3a\u4e86\u6ee1\u8db3 torch.compile \u4e0e <mark data-type=\"institutions\" data-id=\"59257b4b-235a-4864-a37b-956cede3debf\">AMD<\/mark> flash attention \u5185\u6838\u7684\u517c\u5bb9\u6027\uff0cPyTorch \u56e2\u961f\u5fc5\u987b\u5c06\u5b83\u5b9a\u4e49\u4e3a torch \u81ea\u5b9a\u4e49\u7b97\u5b50\u3002\u5e76\u4e14\u5c01\u88c5\u66f4\u590d\u6742\u7684 flash attention \u5185\u6838\u9075\u5faa\u4ee5\u4e0b\u4e24\u4e2a\u6b65\u9aa4\uff1a<\/p>\n<p>\u4e00\u662f\u5c06\u51fd\u6570\u5c01\u88c5\u4e3a\u4e00\u4e2a PyTorch \u81ea\u5b9a\u4e49\u7b97\u5b50\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-759c3e1f60bdf75afe60f01fb30d7703.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-759c3e1f60bdf75afe60f01fb30d7703.png\"><\/a><\/p>\n<p>\u4e8c\u662f\u5411\u8be5\u7b97\u5b50\u6dfb\u52a0\u4e00\u4e2a FakeTensor \u5185\u6838\uff0c\u5e76\u5728\u7ed9\u5b9a flash \u8f93\u5165<mark data-type=\"concepts\" data-id=\"64cc669a-900a-449c-ac73-f71d0e91fc86\">\u5f20\u91cf<\/mark>\u7684\u5f62\u72b6\uff08q\u3001k \u548c v\uff09\u65f6\uff0c\u8ba1\u7b97 flash \u5185\u6838\u7684\u8f93\u51fa\u5f62\u72b6\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-1c975eed0eb7d6e016c17d72662990c4.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-1c975eed0eb7d6e016c17d72662990c4.png\"><\/a><\/p>\n<p>\u5728\u5c06 Triton flash \u5185\u6838\u5b9a\u4e49\u4e3a\u4e00\u4e2a\u81ea\u5b9a\u4e49 op \u540e\uff0cPyTorch \u56e2\u961f\u53ef\u4ee5\u6210\u529f\u5730\u5bf9\u5b83\u8fdb\u884c\u7f16\u8bd1\u4ee5\u5b9e\u73b0\u7aef\u5230\u7aef\u8fd0\u884c\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-2c11ec04b9b55bb1b873ec81739c0dc6.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-2c11ec04b9b55bb1b873ec81739c0dc6.png\"><\/a><\/p>\n<p><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u56fe 6\uff1a\u5728\u4ea4\u6362 Triton matmul \u548c Triton flash attention \u5185\u6838\u540e\uff0c\u4f7f\u7528 torch.compile \u7684 Llama3-8B \u8f68\u8ff9\u3002<\/sup><\/em><\/p>\n<p>\u4ece\u56fe\u4e2d\u53ef\u4ee5\u770b\u5230\uff0c\u5728\u96c6\u6210 SplitK \u77e9\u9635\u4e58\u6cd5\u5185\u6838\u540e\uff0ctorch op \u5c01\u88c5 flash attention \u5185\u6838\uff0c\u7136\u540e\u8fd0\u884c torch.compile\uff0c\u5373\u53ef\u5b9e\u73b0\u4f7f\u7528 100% Triton \u8ba1\u7b97\u5185\u6838\u7684\u524d\u5411\u4f20\u9012\u3002<\/p>\n<p><strong>\u7aef\u5230\u7aef<mark data-type=\"concepts\" data-id=\"308c3a45-0fee-4ec6-858e-85b15f440fc0\">\u57fa\u51c6<\/mark>\u6d4b\u8bd5<\/strong><\/p>\n<p>PyTorch \u56e2\u961f\u5206\u522b\u5bf9\u8fd0\u884c Granite-8B \u548c Llama3-8B \u6a21\u578b\u7684\u82f1\u4f1f\u8fbe H100 \u548c A100\uff08\u5355 GPU\uff09\u8fdb\u884c\u4e86\u7aef\u5230\u7aef\u6d4b\u8bd5\uff0c\u4f7f\u7528\u4e86\u4e24\u79cd\u4e0d\u540c\u7684\u914d\u7f6e\u6765\u6267\u884c<mark data-type=\"concepts\" data-id=\"308c3a45-0fee-4ec6-858e-85b15f440fc0\">\u57fa\u51c6<\/mark>\u6d4b\u8bd5\u3002<\/p>\n<p>\u5176\u4e2d Triton \u5185\u6838\u914d\u7f6e\u4f7f\u7528\u4e86\uff1a<\/p>\n<ul>\n<li>\n<p>Triton SplitK GEMM<\/p>\n<\/li>\n<li>\n<p><mark data-type=\"institutions\" data-id=\"59257b4b-235a-4864-a37b-956cede3debf\">AMD<\/mark> Triton Flash Attention<\/p>\n<\/li>\n<\/ul>\n<p>CUDA \u5185\u6838\u914d\u7f6e\u4f7f\u7528\u4e86<\/p>\n<ul>\n<li>\n<p>cuBLAS GEMM<\/p>\n<\/li>\n<li>\n<p>cuDNN Flash Attention &#8211; Scaled Dot-Product Attention (SDPA)<\/p>\n<\/li>\n<\/ul>\n<p>\u5728\u5178\u578b\u63a8\u7406\u8bbe\u7f6e\u4e0b\uff0c\u4e24\u79cd eager \u548c torch \u7f16\u8bd1\u6a21\u5f0f\u7684\u541e\u5410\u91cf\u548c inter-token \u5ef6\u8fdf\u5982\u4e0b\u56fe\u6240\u793a\u3002<br \/><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-a63619790f3dd95febb48c8ac0546e94.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-a63619790f3dd95febb48c8ac0546e94.png\"><\/a><\/p>\n<p><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u56fe 7\uff1aH100 \u548c A100 \u4e0a Granite-8B \u548c Llama3-8B \u5355 token \u751f\u6210\u5ef6\u8fdf\uff08\u6279\u5927\u5c0f = 2\uff0c\u8f93\u5165\u5e8f\u5217\u957f\u5ea6 = 512\uff0c\u8f93\u51fa\u5e8f\u5217\u957f\u5ea6 = 256\uff09\u3002<\/sup><\/em><\/p>\n<p>\u603b\u7684\u6765\u8bf4\uff0c\u5728 H100 \u4e0a\uff0cTriton \u6a21\u578b\u6700\u9ad8\u53ef\u4ee5\u8fbe\u5230 CUDA \u6a21\u578b\u6027\u80fd\u7684 78%\uff1b\u5728 A100 \u4e0a\u53ef\u4ee5\u8fbe\u5230 82\uff05\u3002\u8fd9\u4e9b\u6027\u80fd\u5dee\u8ddd\u662f\u7531 matmul \u548c flash attention \u7684\u5185\u6838\u5ef6\u8fdf\u9020\u6210\u7684\u3002<\/p>\n<p><strong>\u5fae<mark data-type=\"concepts\" data-id=\"308c3a45-0fee-4ec6-858e-85b15f440fc0\">\u57fa\u51c6<\/mark>\u6d4b\u8bd5<\/strong><\/p>\n<p>\u4e0b\u56fe 8 \u4e3a Triton \u548c CUDA \u5185\u6838\u5ef6\u8fdf\u6bd4\u8f83\uff08\u82f1\u4f1f\u8fbe H100 \u4e0a\u8fd0\u884c Llama3-8B\uff09\u3002\u8f93\u5165\u4e3a\u4e00\u4e2a\u4efb\u610f prompt\uff08\u6279\u5927\u5c0f = 1\uff0cprompt \u5e8f\u5217\u957f\u5ea6 = 44\uff09\uff0c\u4ee5\u89e3\u7801\u5ef6\u8fdf\u65f6\u95f4\u3002<\/p>\n<p>\u6700\u540e\u7ed3\u679c\u663e\u793a\uff0cTriton matmul \u5185\u6838\u6bd4 CUDA \u6162\u4e86 1.2 \u81f3 1.4 \u500d\uff0c\u800c <mark data-type=\"institutions\" data-id=\"59257b4b-235a-4864-a37b-956cede3debf\">AMD<\/mark> Triton Flash Attention \u6bd4 CUDA SDPA \u6162\u4e86 1.6 \u500d\u3002<\/p>\n<p>\u4ee5\u4e0a\u7ed3\u679c\u51f8\u663e\u4e86\u9700\u8981\u8fdb\u4e00\u6b65\u63d0\u5347 GEMM \u548c Flash Attention \u7b49\u6838\u5fc3\u539f\u8bed\u5185\u6838\u7684\u6027\u80fd\u3002\u6700\u8fd1\u7684\u4e00\u4e9b\u5de5\u4f5c\uff08\u5982 FlashAttention-3\u3001FlexAttention\uff09 \u5df2\u7ecf\u63d0\u51fa\u4e86\u66f4\u597d\u5730\u5229\u7528\u5e95\u5c42\u786c\u4ef6\u548c Triton \u7684\u65b9\u6cd5\uff0cPyTorch \u5e0c\u671b\u5728\u5b83\u4eec\u7684\u57fa\u7840\u4e0a\u5b9e\u73b0\u66f4\u5927\u52a0\u901f\u3002\u4e3a\u4e86\u9610\u660e\u8fd9\u4e00\u70b9\uff0cPyTorch \u56e2\u961f\u5c06 FlexAttention \u4e0e SDPA\u3001<mark data-type=\"institutions\" data-id=\"59257b4b-235a-4864-a37b-956cede3debf\">AMD<\/mark>\u2019s Triton Flash \u5185\u6838\u8fdb\u884c\u4e86\u6bd4\u8f83\u3002<\/p>\n<p>PyTorch \u56e2\u961f \u6b63\u52aa\u529b\u9a8c\u8bc1 FlexAttention \u7684\u7aef\u5230\u7aef\u6027\u80fd\u3002\u76ee\u524d\uff0cFlexAttention \u7684\u521d\u59cb\u5fae<mark data-type=\"concepts\" data-id=\"308c3a45-0fee-4ec6-858e-85b15f440fc0\">\u57fa\u51c6<\/mark>\u6d4b\u8bd5\u7ed3\u679c\u8868\u660e\uff0c\u5728<mark data-type=\"concepts\" data-id=\"bf740558-f0f7-41a8-87a0-e695a97563b3\">\u67e5\u8be2<\/mark>\u5411\u91cf\u8f83\u5c0f\u7684\u60c5\u51b5\u4e0b\uff0c\u6709\u671b\u5b9e\u73b0\u66f4\u957f\u7684\u4e0a\u4e0b\u6587\u4ee5\u53ca\u89e3\u7801\u95ee\u9898\u5f62\u72b6\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-936f2fe11b7ce5b4652ee6550062af51.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-936f2fe11b7ce5b4652ee6550062af51.png\"><\/a><\/p>\n<p><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u56fe 9\uff1a\u82f1\u4f1f\u8fbe H100 SXM5 80GB \u4e0a FlexAttention \u5185\u6838<mark data-type=\"concepts\" data-id=\"308c3a45-0fee-4ec6-858e-85b15f440fc0\">\u57fa\u51c6<\/mark>\u6d4b\u8bd5\uff08\u6279\u5927\u5c0f = 1\uff0c\u6700\u5927\u5934\u6570 = 32\uff0c\u5934\u7ef4\u6570 = 128\uff09\u3002<\/sup><\/em><\/p>\n<p><strong>\u672a\u6765\u5de5\u4f5c<\/strong><\/p>\n<p>\u672a\u6765\uff0cPyTorch \u56e2\u961f\u8ba1\u5212\u63a2\u7d22\u8fdb\u4e00\u6b65\u4f18\u5316 matmuls \u7684\u65b9\u6cd5\uff0c\u4ee5\u4fbf\u66f4\u597d\u5730\u5229\u7528\u786c\u4ef6\uff0c\u5e76\u4e3a\u57fa\u4e8e Triton \u7684\u65b9\u6cd5\u5b9e\u73b0\u66f4\u5927\u7684\u52a0\u901f\u3002<\/p>\n<p>\u5bf9\u4e8e flash attention\uff0cPyTorch \u56e2\u961f\u8ba1\u5212\u63a2\u7d22 FlexAttention \u548c FlashAttention-3 \u7b49\u5185\u6838\u4e2d\u4f7f\u7528\u5230\u7684\u6280\u672f\uff0c\u4ee5\u5e2e\u52a9\u8fdb\u4e00\u6b65\u7f29\u5c0f Triton \u4e0e CUDA \u4e4b\u95f4\u7684\u5dee\u8ddd\u3002\u540c\u65f6\u8fd8\u5c06\u63a2\u7d22\u7aef\u5230\u7aef FP8 LLM \u63a8\u7406\u3002<\/p>\n<p><em><sup>\u539f\u6587\u94fe\u63a5\uff1ahttps:\/\/pytorch.org\/blog\/cuda-free-inference-for-llms\/<\/sup><\/em><\/p>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-09-09-9\" target=\"_blank\">\u5929\u4e0b\u82e6\u82f1\u4f1f\u8fbe\u4e45\u77e3\uff01PyTorch\u5b98\u65b9\u514dCUDA\u52a0\u901f\u63a8\u7406\uff0cTriton\u65f6\u4ee3\u8981\u6765\uff1f<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u5929\u4e0b\u82e6\u82f1\u4f1f\u8fbe\u4e45\u77e3\uff01Py [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[27],"tags":[68,55],"class_list":["post-35510","post","type-post","status-publish","format-standard","hentry","category-news","tag-68","tag-55"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.4 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u5929\u4e0b\u82e6\u82f1\u4f1f\u8fbe\u4e45\u77e3\uff01PyTorch\u5b98\u65b9\u514dCUDA\u52a0\u901f\u63a8\u7406\uff0cTriton\u65f6\u4ee3\u8981\u6765\uff1f - \u4e00\u8d77AI\u6280\u672f<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/17aitech.com\/?p=35510\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=35510\",\"url\":\"https:\/\/17aitech.com\/?p=35510\",\"name\":\"\u5929\u4e0b\u82e6\u82f1\u4f1f\u8fbe\u4e45\u77e3\uff01PyTorch\u5b98\u65b9\u514dCUDA\u52a0\u901f\u63a8\u7406\uff0cTriton\u65f6\u4ee3\u8981\u6765\uff1f - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=35510#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=35510#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-9d63e2ec014c249f81a2c2bde0204917.png\",\"datePublished\":\"2024-12-26T01:00:02+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=35510#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=35510\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=35510#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-9d63e2ec014c249f81a2c2bde0204917.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-9d63e2ec014c249f81a2c2bde0204917.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=35510#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u5929\u4e0b\u82e6\u82f1\u4f1f\u8fbe\u4e45\u77e3\uff01PyTorch\u5b98\u65b9\u514dCUDA\u52a0\u901f\u63a8\u7406\uff0cTriton\u65f6\u4ee3\u8981\u6765\uff1f\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/17aitech.com\/#website\",\"url\":\"https:\/\/17aitech.com\/\",\"name\":\"\u4e00\u8d77AI\u6280\u672f\",\"description\":\"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca\",\"alternateName\":\"\u4e00\u8d77AI\u6280\u672f\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/17aitech.com\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\",\"name\":\"AI\u5c0f\u52a9\u624b\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/image\/\",\"url\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"contentUrl\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"caption\":\"AI\u5c0f\u52a9\u624b\"},\"description\":\"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e\",\"url\":\"https:\/\/17aitech.com\/?page_id=33738&user=3\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u5929\u4e0b\u82e6\u82f1\u4f1f\u8fbe\u4e45\u77e3\uff01PyTorch\u5b98\u65b9\u514dCUDA\u52a0\u901f\u63a8\u7406\uff0cTriton\u65f6\u4ee3\u8981\u6765\uff1f - \u4e00\u8d77AI\u6280\u672f","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/17aitech.com\/?p=35510","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=35510","url":"https:\/\/17aitech.com\/?p=35510","name":"\u5929\u4e0b\u82e6\u82f1\u4f1f\u8fbe\u4e45\u77e3\uff01PyTorch\u5b98\u65b9\u514dCUDA\u52a0\u901f\u63a8\u7406\uff0cTriton\u65f6\u4ee3\u8981\u6765\uff1f - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=35510#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=35510#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-9d63e2ec014c249f81a2c2bde0204917.png","datePublished":"2024-12-26T01:00:02+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=35510#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=35510"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=35510#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-9d63e2ec014c249f81a2c2bde0204917.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-9d63e2ec014c249f81a2c2bde0204917.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=35510#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u5929\u4e0b\u82e6\u82f1\u4f1f\u8fbe\u4e45\u77e3\uff01PyTorch\u5b98\u65b9\u514dCUDA\u52a0\u901f\u63a8\u7406\uff0cTriton\u65f6\u4ee3\u8981\u6765\uff1f"}]},{"@type":"WebSite","@id":"https:\/\/17aitech.com\/#website","url":"https:\/\/17aitech.com\/","name":"\u4e00\u8d77AI\u6280\u672f","description":"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca","alternateName":"\u4e00\u8d77AI\u6280\u672f","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/17aitech.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d","name":"AI\u5c0f\u52a9\u624b","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/#\/schema\/person\/image\/","url":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","contentUrl":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","caption":"AI\u5c0f\u52a9\u624b"},"description":"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e","url":"https:\/\/17aitech.com\/?page_id=33738&user=3"}]}},"_links":{"self":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/35510","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=35510"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/35510\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=35510"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=35510"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=35510"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}