{"id":34520,"date":"2024-12-08T05:56:32","date_gmt":"2024-12-07T21:56:32","guid":{"rendered":"https:\/\/17aitech.com\/?p=34520"},"modified":"2024-12-08T05:56:32","modified_gmt":"2024-12-07T21:56:32","slug":"%e8%8b%b1%e4%bc%9f%e8%be%be%e5%8f%88%e8%b5%9a%e5%88%b0%e4%ba%86%ef%bc%81flashattention3%e6%9d%a5%e4%ba%86%ef%bc%9ah100%e5%88%a9%e7%94%a8%e7%8e%87%e9%a3%99%e5%8d%87%e8%87%b375","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=34520","title":{"rendered":"\u82f1\u4f1f\u8fbe\u53c8\u8d5a\u5230\u4e86\uff01FlashAttention3\u6765\u4e86\uff1aH100\u5229\u7528\u7387\u98d9\u5347\u81f375%"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-07-12-6\" target=\"_blank\">\u82f1\u4f1f\u8fbe\u53c8\u8d5a\u5230\u4e86\uff01FlashAttention3\u6765\u4e86\uff1aH100\u5229\u7528\u7387\u98d9\u5347\u81f375%<\/a><\/p>\n<blockquote data-author-name=\"\" data-content-utf8-length=\"32\" data-source-title=\"\" data-type=\"2\" data-url=\"\">\n<section>\n<section>\n<p>740 TFLOPS\uff01\u8fc4\u4eca\u6700\u5f3a FlashAttention \u6765\u4e86\u3002<\/p>\n<\/section>\n<\/section>\n<\/blockquote>\n<p>\u968f\u7740\u5927\u578b<mark data-type=\"tech_tasks\" data-id=\"fee178e8-ee20-42fc-8f8f-1d41c1f34e8f\">\u8bed\u8a00\u6a21\u578b<\/mark>\uff08LLM\uff09\u52a0\u901f\u843d\u5730\uff0c\u6269\u5c55\u6a21\u578b\u4e0a\u4e0b\u6587\u7a97\u53e3\u53d8\u5f97\u8d8a\u6765\u8d8a\u91cd\u8981\u3002\u7136\u800c\uff0cTransformer \u67b6\u6784\u7684\u6838\u5fc3 \u2014\u2014 \u6ce8\u610f\u529b\u5c42\u7684<mark data-type=\"concepts\" data-id=\"324e28ee-fa13-4e36-9968-ae71059f5948\">\u65f6\u95f4\u590d\u6742\u5ea6<\/mark>\u548c\u7a7a\u95f4\u590d\u6742\u5ea6\u4e0e\u8f93\u5165\u5e8f\u5217\u957f\u5ea6\u7684\u5e73\u65b9\u6210\u6b63\u6bd4\u3002\u8fd9\u4f7f\u5f97\u6269\u5c55\u6a21\u578b\u4e0a\u4e0b\u6587\u7a97\u53e3\u5b58\u5728\u6311\u6218\u3002<\/p>\n<p>2022 \u5e74\uff0c\u4e00\u79cd\u5feb\u901f\u3001\u5185\u5b58\u9ad8\u6548\u7684\u6ce8\u610f\u529b\u7b97\u6cd5 \u2014\u2014FlashAttention \u95ee\u4e16\uff0c\u8be5\u7b97\u6cd5\u65e0\u9700\u4efb\u4f55\u8fd1\u4f3c\u5373\u53ef\u52a0\u901f\u6ce8\u610f\u529b\u5e76\u51cf\u5c11\u5185\u5b58\u5360\u7528\u3002<\/p>\n<p>FlashAttention \u5bf9\u6ce8\u610f\u529b\u8ba1\u7b97\u8fdb\u884c\u91cd\u65b0\u6392\u5e8f\u7684\u7b97\u6cd5\uff0c\u5e76\u5229\u7528 tiling \u548c\u91cd\u8ba1\u7b97\u6765\u663e\u8457\u52a0\u5feb\u8ba1\u7b97\u901f\u5ea6\uff0c\u5c06\u5185\u5b58\u4f7f\u7528\u91cf\u4ece\u5e8f\u5217\u957f\u5ea6\u7684\u4e8c\u6b21\u51cf\u5c11\u5230\u7ebf\u6027\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-6e583197a942409fc6b04c1a126acdf8.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-6e583197a942409fc6b04c1a126acdf8.png\"><\/a><\/section>\n<p>2023 \u5e74\uff0c\u7814\u7a76\u56e2\u961f\u5ba3\u5e03\u63a8\u51fa FlashAttention-2\uff0c\u5728\u7b97\u6cd5\u3001\u5e76\u884c\u5316\u548c\u5de5\u4f5c\u5206\u533a\u7b49\u65b9\u9762\u6709\u4e86\u663e\u8457\u6539\u8fdb\u3002<\/p>\n<p>\u73b0\u5728\uff0c\u6765\u81ea Meta\u3001\u82f1\u4f1f\u8fbe\u3001Together AI \u7b49\u673a\u6784\u7684\u7814\u7a76\u8005\u5ba3\u5e03\u63a8\u51fa FlashAttention-3\uff0c\u5b83\u91c7\u7528\u4e86\u52a0\u901f <mark data-type=\"institutions\" data-id=\"3799db24-4c55-4eac-94ba-69dfcf700573\">Hopper<\/mark> GPU \u6ce8\u610f\u529b\u7684\u4e09\u79cd\u4e3b\u8981\u6280\u672f\uff1a<\/p>\n<ul>\n<li>\n<p>\u901a\u8fc7 warp-specialization \u91cd\u53e0\u6574\u4f53\u8ba1\u7b97\u548c\u6570\u636e\u79fb\u52a8\uff1b<\/p>\n<\/li>\n<li>\n<p>\u4ea4\u9519<mark data-type=\"tech_tasks\" data-id=\"70d1a967-90f5-47be-b22f-1e2c746f4914\">\u5206\u5757<\/mark> matmul \u548c softmax \u8fd0\u7b97\uff1b<\/p>\n<\/li>\n<li>\n<p>\u5229\u7528\u786c\u4ef6\u652f\u6301 FP8 \u4f4e\u7cbe\u5ea6\u7684\u4e0d\u8fde\u8d2f\u5904\u7406\u3002<\/p>\n<\/li>\n<\/ul>\n<p>FlashAttention-3 \u7684\u901f\u5ea6\u662f FlashAttention-2 \u7684 1.5-2.0 \u500d\uff0c\u9ad8\u8fbe 740 TFLOPS\uff0c\u5373 H100 \u7406\u8bba\u6700\u5927 FLOPS \u5229\u7528\u7387\u4e3a 75%\u3002\u4f7f\u7528 FP8\uff0cFlashAttention-3 \u7684\u901f\u5ea6\u66f4\u662f\u63a5\u8fd1 1.2 PFLOPS\u3002<\/p>\n<p>FlashAttention-3 \u7684\u6539\u8fdb\u5c06\u5e26\u6765\uff1a<\/p>\n<ul>\n<li>\n<p>\u66f4\u9ad8\u6548\u7684 GPU \u5229\u7528\u7387\uff1aH100 \u7406\u8bba\u6700\u5927\u00a0FLOPS \u5229\u7528\u7387\u4e3a 75%\uff0c\u800c\u4e4b\u524d\u4ec5\u4e3a 35%\u3002\u8fd9\u4f7f\u5f97 LLM \u7684\u8bad\u7ec3\u548c\u8fd0\u884c\u901f\u5ea6\u6bd4\u4ee5\u524d\u7684\u7248\u672c\u5feb\u5f97\u591a\u3002<\/p>\n<\/li>\n<li>\n<p>\u8f83\u4f4e\u7cbe\u5ea6\u4e0b\u66f4\u597d\u7684\u6027\u80fd\uff1aFlashAttention-3 \u53ef\u4ee5\u5728\u4fdd\u6301\u7cbe\u5ea6\u7684\u540c\u65f6\u4f7f\u7528\u8f83\u4f4e\u7cbe\u5ea6\u7684\u6570\u5b57 (FP8)\u3002\u8fd9\u53ef\u4ee5\u5b9e\u73b0\u66f4\u5feb\u7684\u5904\u7406\u901f\u5ea6\u5e76\u53ef\u80fd\u964d\u4f4e\u5185\u5b58\u4f7f\u7528\u91cf\uff0c\u4ece\u800c\u4e3a\u8fd0\u884c\u5927\u89c4\u6a21<mark data-type=\"concepts\" data-id=\"2d28aa9c-942d-471d-bd96-8bfefb7144e0\">\u4eba\u5de5\u667a\u80fd<\/mark>\u64cd\u4f5c\u7684\u5ba2\u6237\u8282\u7701\u6210\u672c\u5e76\u63d0\u9ad8\u6548\u7387\u3002<\/p>\n<\/li>\n<li>\n<p>\u80fd\u591f\u5728 LLM \u4e2d\u4f7f\u7528\u66f4\u957f\u7684\u4e0a\u4e0b\u6587\uff1a\u901a\u8fc7\u52a0\u901f<mark data-type=\"tech_methods\" data-id=\"60bee267-89dc-4606-ba24-6b3f7d1f6189\">\u6ce8\u610f\u529b\u673a\u5236<\/mark>\uff0cFlashAttention-3 \u4f7f AI \u6a21\u578b\u80fd\u591f\u66f4\u6709\u6548\u5730\u5904\u7406\u66f4\u957f\u7684\u6587\u672c\u7247\u6bb5\u3002\u8fd9\u4f7f\u5f97\u5e94\u7528\u7a0b\u5e8f\u80fd\u591f\u7406\u89e3\u5e76\u751f\u6210\u66f4\u957f\u3001\u66f4\u590d\u6742\u7684\u5185\u5bb9\u800c\u4e0d\u4f1a\u51cf\u6162\u901f\u5ea6\u3002<\/p>\n<\/li>\n<\/ul>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-6d0a55ac522d63cd991f9d262a46a62f.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-6d0a55ac522d63cd991f9d262a46a62f.png\"><\/a><\/section>\n<p>\u8bba\u6587\u6807\u9898\uff1aFlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision<\/p>\n<p>\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/tridao.me\/publications\/flash3\/flash3.pdf<\/p>\n<p>\u8bba\u6587\u4f5c\u8005\u4e4b\u4e00 \u3001FlashAttention1-3 \u7248\u672c\u7684\u53c2\u4e0e\u8005 Tri Dao \u8868\u793a\uff1aFlashAttention \u88ab\u5e7f\u6cdb\u7528\u4e8e\u52a0\u901f Transformers\uff0c\u5df2\u7ecf\u4f7f\u6ce8\u610f\u529b\u901f\u5ea6\u63d0\u9ad8\u4e86 4-8 \u500d\uff0c\u4f46\u5c1a\u672a\u5229\u7528\u73b0\u4ee3 GPU\u3002\u56e0\u800c\u4ed6\u4eec\u53d1\u5e03\u4e86 FlashAttention-3\uff1a\u5728 FP16 \u4e0a\u901f\u5ea6\u63d0\u9ad8\u4e86 1.5-2 \u500d\uff0c\u5728 H100 \u4e0a\u9ad8\u8fbe 740 TFLOPS\uff0875% \u5b9e\u7528\u6027\uff09\uff0cFP8 \u63a5\u8fd1 1.2 PFLOPS\uff01<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-b36ace61023ebec480e57cebaeed5dcb.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-b36ace61023ebec480e57cebaeed5dcb.png\"><\/a><\/section>\n<p><strong><mark data-type=\"institutions\" data-id=\"3799db24-4c55-4eac-94ba-69dfcf700573\">Hopper<\/mark> GPU \u786c\u4ef6\u7279\u6027\uff1aWGMMA\u3001TMA\u3001FP8<\/strong><\/p>\n<p>\u867d\u7136 FlashAttention-2 \u5728 Ampere (A100) GPU \u4e0a\u53ef\u4ee5\u5b9e\u73b0 70% \u7684\u7406\u8bba\u6700\u5927 FLOPS\uff0c\u4f46\u5b83\u5c1a\u672a\u5229\u7528 <mark data-type=\"institutions\" data-id=\"3799db24-4c55-4eac-94ba-69dfcf700573\">Hopper<\/mark> GPU \u4e0a\u7684\u65b0\u529f\u80fd\u6765\u6700\u5927\u9650\u5ea6\u5730\u63d0\u9ad8\u6027\u80fd\u3002\u63a5\u4e0b\u6765\u6587\u7ae0\u63cf\u8ff0\u4e86\u4e00\u4e9b\u65b0\u7684 <mark data-type=\"institutions\" data-id=\"3799db24-4c55-4eac-94ba-69dfcf700573\">Hopper<\/mark> \u7279\u5b9a\u529f\u80fd\uff0c\u4ee5\u53ca\u5b83\u4eec\u4e3a\u4f55\u5982\u6b64\u91cd\u8981\u3002<\/p>\n<p>\u9996\u5148\u662f WGMMA\uff08Warpgroup Matrix Multiply-Accumulate\uff09\uff0c\u8be5\u529f\u80fd\u5229\u7528\u4e86 <mark data-type=\"institutions\" data-id=\"3799db24-4c55-4eac-94ba-69dfcf700573\">Hopper<\/mark> \u67b6\u6784\u4e0a\u65b0\u7684<mark data-type=\"concepts\" data-id=\"64cc669a-900a-449c-ac73-f71d0e91fc86\">\u5f20\u91cf<\/mark>\u5185\u6838\uff0c\u6bd4 Ampere \u67b6\u6784\u5177\u6709\u66f4\u9ad8\u7684\u541e\u5410\u91cf\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-aeac5c550f0054f7606d1e882541d776.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-aeac5c550f0054f7606d1e882541d776.png\"><\/a><\/section>\n<p>\u7136\u540e\u662f TMA\uff08Tensor Memory Accelerator\uff09\uff0c\u8fd9\u662f\u4e00\u4e2a\u7279\u6b8a\u7684\u786c\u4ef6\u5355\u5143\uff0c\u53ef\u4ee5\u52a0\u901f\u5168\u5c40\u5185\u5b58\u548c\u5171\u4eab\u5185\u5b58\u4e4b\u95f4\u7684\u6570\u636e\u4f20\u8f93\uff0c\u7528\u4e8e\u5904\u7406\u6240\u6709\u7d22\u5f15\u8ba1\u7b97\u548c\u8fb9\u754c\u5916\u9884\u6d4b\u3002\u8fd9\u6837\u4e00\u6765\u5bc4\u5b58\u5668\u5c31\u91ca\u653e\u4e86\uff0c\u5bc4\u5b58\u5668\u662f\u589e\u52a0 tile \u5927\u5c0f\u548c\u6548\u7387\u7684\u5b9d\u8d35\u8d44\u6e90\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-1549f670ff89ad7a55afb3f727571d6d.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-1549f670ff89ad7a55afb3f727571d6d.png\"><\/a><\/section>\n<p>\u4f4e\u7cbe\u5ea6 FP8\uff0c\u8ba9 Tensor Core \u541e\u5410\u91cf\u7ffb\u4e86\u4e00\u500d\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-f378b875b201073a5216b4203b1b1628.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-f378b875b201073a5216b4203b1b1628.png\"><\/a><\/section>\n<p>FlashAttention-3 \u5145\u5206\u5229\u7528\u4e86 <mark data-type=\"institutions\" data-id=\"3799db24-4c55-4eac-94ba-69dfcf700573\">Hopper<\/mark> \u67b6\u6784\u7684\u6240\u6709\u8fd9\u4e9b\u65b0\u529f\u80fd\u3002<\/p>\n<p><strong>\u5f02\u6b65\uff1aGEMM \u548c Softmax \u91cd\u53e0<\/strong><\/p>\n<p><mark data-type=\"tech_methods\" data-id=\"60bee267-89dc-4606-ba24-6b3f7d1f6189\">\u6ce8\u610f\u529b\u673a\u5236<\/mark>\u4e3b\u8981\u6709\u4e24\u4e2a\u64cd\u4f5c\uff0cGEMM \u548c softmax\u3002\u4e3a\u4ec0\u4e48\u8981\u5c06\u5b83\u4eec\u91cd\u53e0\uff1f<\/p>\n<p>\u95ee\u9898\u5728\u4e8e\u5728\u73b0\u4ee3\u52a0\u901f\u5668\u4e0a\uff0c\u975e\u77e9\u9635\u4e58\u6cd5\uff08matmul\uff09\u8fd0\u7b97\u6bd4\u77e9\u9635\u4e58\u6cd5\u8fd0\u7b97\u6162\u3002\u7279\u6b8a\u51fd\u6570\u5982\u6307\u6570\u8fd0\u7b97\uff08\u5982 softmax \u51fd\u6570\uff09\u7684\u541e\u5410\u91cf\u751a\u81f3\u4f4e\u4e8e\u6d6e\u70b9\u4e58\u52a0\u64cd\u4f5c\uff1b\u8fd9\u4e9b\u8fd0\u7b97\u662f\u7531\u591a\u529f\u80fd\u5355\u5143\u5904\u7406\u7684\uff0c\u8fd9\u662f\u4e00\u4e2a\u4e0e\u6d6e\u70b9\u4e58\u52a0\u6216\u77e9\u9635\u4e58\u52a0\u4e0d\u540c\u7684\u5355\u5143\u3002<\/p>\n<p>\u7406\u60f3\u60c5\u51b5\u4e0b\uff0c\u7814\u7a76\u8005\u5e0c\u671b\u77e9\u9635\u4e58\u6cd5\u548c softmax \u80fd\u591f\u5e76\u884c\u64cd\u4f5c\u3002\u5f53 Tensor Cores \u5fd9\u4e8e\u77e9\u9635\u4e58\u6cd5\u65f6\uff0c\u591a\u529f\u80fd\u5355\u5143\u5e94\u5f53\u5728\u8ba1\u7b97\u6307\u6570\u8fd0\u7b97\uff01\u00a0<\/p>\n<p><strong>Inter-warpgroup \u91cd\u53e0<\/strong><\/p>\n<p>\u91cd\u53e0 GEMM \u548c softmax \u6700\u7b80\u5355\u7684\u65b9\u6cd5\u662f\u4ec0\u4e48\u90fd\u4e0d\u505a\uff0cwarp <mark data-type=\"tech_methods\" data-id=\"1ef8dd91-7c80-46f2-811e-0722d3d333d4\">\u8c03\u5ea6<\/mark>\u7a0b\u5e8f\u4f1a\u514d\u8d39\u5b8c\u6210\u90e8\u5206\u91cd\u53e0\u3002\u4e0b\u56fe\u8bf4\u660e\u4e86 pingpong <mark data-type=\"tech_methods\" data-id=\"1ef8dd91-7c80-46f2-811e-0722d3d333d4\">\u8c03\u5ea6<\/mark>\uff0c\u5176\u4e2d\u76f8\u540c\u7684\u989c\u8272\u8868\u793a\u76f8\u540c\u7684\u8fed\u4ee3\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-6ab9760d3a0ac8476811e2989dc991f0.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-6ab9760d3a0ac8476811e2989dc991f0.png\"><\/a><\/section>\n<p><strong>Intra-warpgroup \u91cd\u53e0<\/strong><\/p>\n<p>\u5373\u4f7f\u5728\u4e00\u4e2a warpgroup \u4e2d\uff0c\u7814\u7a76\u8005\u4e5f\u53ef\u4ee5\u5728\u8fd0\u884c\u8be5 warpgroup \u7684 GEMM \u65f6\u8fd0\u884c softmax \u7684\u67d0\u4e9b\u90e8\u5206\u3002\u5982\u56fe\u6240\u793a\uff0c\u76f8\u540c\u7684\u989c\u8272\u8868\u793a\u76f8\u540c\u7684\u8fed\u4ee3\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-b249a7f9a66902b31e1a9b83e7ed4a5d.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-b249a7f9a66902b31e1a9b83e7ed4a5d.png\"><\/a><\/section>\n<p>\u8fd9\u79cd pipeline \u6d41\u7a0b\u53ef\u4ee5\u5c06 FP16 \u6ce8\u610f\u529b\u524d\u5411\u4f20\u64ad\u7684\u541e\u5410\u91cf\u4ece\u5927\u7ea6 620 TFLOPS \u63d0\u9ad8\u5230 640-660 TFLOPS\uff0c\u4f46\u4ee3\u4ef7\u662f\u66f4\u9ad8\u7684\u5bc4\u5b58\u5668\u538b\u529b\uff0c\u56e0\u800c\u9700\u8981\u66f4\u591a\u7684\u5bc4\u5b58\u5668\u6765\u540c\u65f6\u4fdd\u5b58 GEMM \u7684\u7d2f\u52a0\u5668\u4ee5\u53ca Softmax \u7684\u8f93\u5165 \/ \u8f93\u51fa\u3002<\/p>\n<p><strong>\u4f4e\u7cbe\u5ea6\uff1a\u4f7f\u7528\u975e\u76f8\u5e72\u5904\u7406\u51cf\u5c11<mark data-type=\"tech_tasks\" data-id=\"e1abaa20-0000-4ff4-ad42-f97ca7fca4b1\">\u91cf\u5316<\/mark>\u8bef\u5dee<\/strong><\/p>\n<p>\u6fc0\u6d3b LLM \u53ef\u80fd\u5b58\u5728\u4e00\u4e9b\u6781\u7aef\u503c\uff0c\u5bfc\u81f4<mark data-type=\"tech_tasks\" data-id=\"e1abaa20-0000-4ff4-ad42-f97ca7fca4b1\">\u91cf\u5316<\/mark>\u56f0\u96be\uff0c\u4ece\u800c\u4ea7\u751f\u8f83\u5927\u7684<mark data-type=\"tech_tasks\" data-id=\"e1abaa20-0000-4ff4-ad42-f97ca7fca4b1\">\u91cf\u5316<\/mark>\u8bef\u5dee\u3002\u672c\u6587\u91c7\u7528\u975e\u76f8\u5e72\u5904\u7406\uff08incoherent processing\uff09\uff0c\u8be5\u6280\u672f\u901a\u8fc7\u5c06<mark data-type=\"concepts\" data-id=\"bf740558-f0f7-41a8-87a0-e695a97563b3\">\u67e5\u8be2<\/mark>\u548c\u952e\u4e0e\u4e00\u4e2a\u968f\u673a\u6b63\u4ea4\u77e9\u9635\u76f8\u4e58\u6765\u300c\u5206\u6563\uff08spread out\uff09\u300d\u6781\u7aef\u503c\uff0c\u4ece\u800c\u51cf\u5c11<mark data-type=\"tech_tasks\" data-id=\"e1abaa20-0000-4ff4-ad42-f97ca7fca4b1\">\u91cf\u5316<\/mark>\u8bef\u5dee\u3002\u7279\u522b\u5730\uff0c\u8be5\u7814\u7a76\u4f7f\u7528\u4e86 Hadamard \u53d8\u6362\uff0c\u5b83\u53ef\u4ee5\u5728\u6bcf\u4e2a\u6ce8\u610f\u529b\u5934\u4e2d\u4ee5 O (d log d) \u7684<mark data-type=\"concepts\" data-id=\"324e28ee-fa13-4e36-9968-ae71059f5948\">\u65f6\u95f4\u590d\u6742\u5ea6<\/mark>\u5b8c\u6210\uff0c\u800c\u4e0d\u662f O (d^2)\uff0c\u5176\u4e2d d \u662f\u5934\u90e8\u7ef4\u5ea6\u3002<\/p>\n<p>\u7814\u7a76\u8005\u53d1\u73b0\u975e\u76f8\u5e72\u5904\u7406\u53ef\u4ee5\u5c06<mark data-type=\"tech_tasks\" data-id=\"e1abaa20-0000-4ff4-ad42-f97ca7fca4b1\">\u91cf\u5316<\/mark>\u8bef\u5dee\u51cf\u5c11\u5f88\u591a\uff0c\u5177\u4f53\u7684\u6570\u503c\u8bef\u5dee\u6bd4\u8f83\u89c1\u4e0b\u8868\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-7fe26969a56cc76b09a45eb7bf97cc52.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-7fe26969a56cc76b09a45eb7bf97cc52.png\"><\/a><\/section>\n<p><strong>\u5b9e\u9a8c<\/strong><\/p>\n<p>\u6587\u4e2d\u5c55\u793a\u4e86 FlashAttention-3 \u7684\u4e00\u4e9b\u7ed3\u679c\uff0c\u5e76\u5c06\u5176\u4e0e FlashAttention-2 \u4ee5\u53ca Triton \u548c cuDNN \u4e2d\u7684\u5b9e\u73b0\u8fdb\u884c\u4e86\u6bd4\u8f83\uff08\u4e24\u8005\u90fd\u5df2\u7ecf\u4f7f\u7528\u4e86 <mark data-type=\"institutions\" data-id=\"3799db24-4c55-4eac-94ba-69dfcf700573\">Hopper<\/mark> GPU \u7684\u65b0\u786c\u4ef6\u529f\u80fd\uff09\u3002<\/p>\n<p>\u5728 FP16 \u7cbe\u5ea6\u4e0b\uff0cFlashAttention-3 \u7684\u901f\u5ea6\u662f FlashAttention-2 \u7684 1.5-2.0 \u500d\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-06beeb2f9ac75d313abd9199ca80a12b.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-06beeb2f9ac75d313abd9199ca80a12b.png\"><\/a><\/section>\n<p>\u5bf9\u4e8e FP8\uff0cFlashAttention-3 \u63a5\u8fd1 1.2 PFLOPS\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-5f3a6a83537c15ef7fb31d3cfd8ef2a0.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-5f3a6a83537c15ef7fb31d3cfd8ef2a0.png\"><\/a><\/section>\n<p>\u6269\u5c55\u9605\u8bfb\uff1a<\/p>\n<p><a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650848429&amp;idx=4&amp;sn=4665869919c379023b1bdb29568cdb2c&amp;chksm=84e578d3b392f1c59b3e4c9b986a522f6534e894800e12f5df52964d1ec45f6c067a73b53bd9&amp;scene=21#wechat_redirect\" target=\"_blank\">\u65af\u5766\u798f\u63d0\u51fa\u65b0\u578bAttention\u7b97\u6cd5\uff01\u63d0\u901f2-4\u500d\uff0cBERT\u5355\u8282\u70b9\u8bad\u7ec3\u6700\u5feb<\/a>\u00a0<\/p>\n<p><a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650884509&amp;idx=3&amp;sn=65476dbc71ca235155734ed6cf52197d&amp;chksm=84e48de3b39304f54bc222ce6da480ef5ddd8874b5254697eecf1583aef635eb916818901dab&amp;scene=21#wechat_redirect\" target=\"_blank\">\u6bd4\u6807\u51c6Attention\u63d0\u901f5-9\u500d\uff0c\u5927\u6a21\u578b\u90fd\u5728\u7528\u7684FlashAttention v2\u6765\u4e86<\/a><\/p>\n<p><sup>\u53c2\u8003\u94fe\u63a5\uff1a<\/sup><\/p>\n<p><em><sup>https:\/\/tridao.me\/blog\/2024\/flash3\/<\/sup><\/em><\/p>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-07-12-6\" target=\"_blank\">\u82f1\u4f1f\u8fbe\u53c8\u8d5a\u5230\u4e86\uff01FlashAttention3\u6765\u4e86\uff1aH100\u5229\u7528\u7387\u98d9\u5347\u81f375%<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u82f1\u4f1f\u8fbe\u53c8\u8d5a\u5230\u4e86\uff01Fla [&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],"class_list":["post-34520","post","type-post","status-publish","format-standard","hentry","category-news","tag-68"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.4 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u82f1\u4f1f\u8fbe\u53c8\u8d5a\u5230\u4e86\uff01FlashAttention3\u6765\u4e86\uff1aH100\u5229\u7528\u7387\u98d9\u5347\u81f375% - \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=34520\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=34520\",\"url\":\"https:\/\/17aitech.com\/?p=34520\",\"name\":\"\u82f1\u4f1f\u8fbe\u53c8\u8d5a\u5230\u4e86\uff01FlashAttention3\u6765\u4e86\uff1aH100\u5229\u7528\u7387\u98d9\u5347\u81f375% - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=34520#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=34520#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-6e583197a942409fc6b04c1a126acdf8.png\",\"datePublished\":\"2024-12-07T21:56:32+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=34520#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=34520\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=34520#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-6e583197a942409fc6b04c1a126acdf8.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-6e583197a942409fc6b04c1a126acdf8.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=34520#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u82f1\u4f1f\u8fbe\u53c8\u8d5a\u5230\u4e86\uff01FlashAttention3\u6765\u4e86\uff1aH100\u5229\u7528\u7387\u98d9\u5347\u81f375%\"}]},{\"@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":"\u82f1\u4f1f\u8fbe\u53c8\u8d5a\u5230\u4e86\uff01FlashAttention3\u6765\u4e86\uff1aH100\u5229\u7528\u7387\u98d9\u5347\u81f375% - \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=34520","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=34520","url":"https:\/\/17aitech.com\/?p=34520","name":"\u82f1\u4f1f\u8fbe\u53c8\u8d5a\u5230\u4e86\uff01FlashAttention3\u6765\u4e86\uff1aH100\u5229\u7528\u7387\u98d9\u5347\u81f375% - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=34520#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=34520#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-6e583197a942409fc6b04c1a126acdf8.png","datePublished":"2024-12-07T21:56:32+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=34520#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=34520"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=34520#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-6e583197a942409fc6b04c1a126acdf8.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-6e583197a942409fc6b04c1a126acdf8.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=34520#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u82f1\u4f1f\u8fbe\u53c8\u8d5a\u5230\u4e86\uff01FlashAttention3\u6765\u4e86\uff1aH100\u5229\u7528\u7387\u98d9\u5347\u81f375%"}]},{"@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\/34520","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=34520"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/34520\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=34520"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=34520"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=34520"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}