{"id":35923,"date":"2025-01-03T06:00:11","date_gmt":"2025-01-02T22:00:11","guid":{"rendered":"https:\/\/17aitech.com\/?p=35923"},"modified":"2025-01-03T06:00:11","modified_gmt":"2025-01-02T22:00:11","slug":"%e5%91%8a%e5%88%abcuda%e6%97%a0%e9%9c%80triton%ef%bc%81mirage%e9%9b%b6%e9%97%a8%e6%a7%9b%e7%94%9f%e6%88%90pytorch%e7%ae%97%e5%ad%90%ef%bc%8c%e4%ba%ba%e5%9d%87gpu%e7%bc%96%e7%a8%8b%e5%a4%a7%e5%b8%88","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=35923","title":{"rendered":"\u544a\u522bCUDA\u65e0\u9700Triton\uff01Mirage\u96f6\u95e8\u69db\u751f\u6210PyTorch\u7b97\u5b50\uff0c\u4eba\u5747GPU\u7f16\u7a0b\u5927\u5e08\uff1f"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-10-06\" target=\"_blank\">\u544a\u522bCUDA\u65e0\u9700Triton\uff01Mirage\u96f6\u95e8\u69db\u751f\u6210PyTorch\u7b97\u5b50\uff0c\u4eba\u5747GPU\u7f16\u7a0b\u5927\u5e08\uff1f<\/a><\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-ce394fffbfc5469c03c43a0ff80958ce.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-ce394fffbfc5469c03c43a0ff80958ce.png\"><\/a><\/section>\n<blockquote data-author-name=\"\" data-content-utf8-length=\"166\" data-source-title=\"\" data-type=\"2\" data-url=\"\">\n<section>\n<p>AIxiv\u4e13\u680f\u662f\u673a\u5668\u4e4b\u5fc3\u53d1\u5e03\u5b66\u672f\u3001\u6280\u672f\u5185\u5bb9\u7684\u680f\u76ee\u3002\u8fc7\u53bb\u6570\u5e74\uff0c\u673a\u5668\u4e4b\u5fc3AIxiv\u4e13\u680f\u63a5\u6536\u62a5\u9053\u4e862000\u591a\u7bc7\u5185\u5bb9\uff0c\u8986\u76d6\u5168\u7403\u5404\u5927\u9ad8\u6821\u4e0e\u4f01\u4e1a\u7684\u9876\u7ea7\u5b9e\u9a8c\u5ba4\uff0c\u6709\u6548\u4fc3\u8fdb\u4e86\u5b66\u672f\u4ea4\u6d41\u4e0e\u4f20\u64ad\u3002\u5982\u679c\u60a8\u6709\u4f18\u79c0\u7684\u5de5\u4f5c\u60f3\u8981\u5206\u4eab\uff0c\u6b22\u8fce\u6295\u7a3f\u6216\u8005\u8054\u7cfb\u62a5\u9053\u3002\u6295\u7a3f\u90ae\u7bb1\uff1aliyazhou@jiqizhixin.com\uff1bzhaoyunfeng@jiqizhixin.com<\/p>\n<\/section>\n<\/blockquote>\n<section><\/section>\n<section>\u8fd1\u65e5\uff0c\u6765\u81ea CMU \u7684 Catalyst Group \u56e2\u961f\u53d1\u5e03\u4e86\u4e00\u6b3e PyTorch \u7b97\u5b50\u7f16\u8bd1\u5668 Mirage\uff0c\u7528\u6237\u65e0\u9700\u7f16\u5199\u4efb\u4f55 CUDA \u548c Triton \u4ee3\u7801\u5c31\u53ef\u4ee5\u81ea\u52a8\u751f\u6210 GPU \u5185\u6838\uff0c\u5e76\u53d6\u5f97\u66f4\u4f73\u7684\u6027\u80fd\u3002<\/section>\n<section><\/section>\n<section>\u968f\u7740 GPU \u52a0\u901f\u5668\u7684\u4e0d\u65ad\u53d1\u5c55\u4ee5\u53ca\u4ee5\u5927\u8bed\u8a00\u6a21\u578b\u4e3a\u4ee3\u8868\u7684\u751f\u6210\u5f0f AI \u5e94\u7528\u7684\u4e0d\u65ad\u63a8\u5e7f\uff0c\u901a\u8fc7\u5f00\u53d1\u9ad8\u6027\u80fd GPU \u5185\u6838\u6765\u4f18\u5316 PyTorch \u7a0b\u5e8f\u7684\u8ba1\u7b97\u6548\u7387\u53d8\u5f97\u8d8a\u6765\u8d8a\u91cd\u8981\u3002\u76ee\u524d\uff0c\u8fd9\u9879\u4efb\u52a1\u4e3b\u8981\u7531\u4e13\u95e8\u7684 GPU \u4e13\u5bb6\u6765\u5b8c\u6210\u3002\u5728 NVIDIA CUDA \u6216 AMD ROCm \u4e2d\u7f16\u5199\u9ad8\u6027\u80fd GPU \u5185\u6838\u9700\u8981\u9ad8\u6c34\u5e73\u7684 GPU \u4e13\u4e1a\u77e5\u8bc6\u548c\u5927\u91cf\u7684\u5de5\u7a0b\u5f00\u53d1\u7ecf\u9a8c\u3002\u76ee\u524d\u7684\u673a\u5668\u5b66\u4e60\u7f16\u8bd1\u5668\uff08\u5982 TVM\u3001Triton \u548c Mojo\uff09\u63d0\u4f9b\u4e86\u4e00\u4e9b\u9ad8\u7ea7\u7f16\u7a0b\u63a5\u53e3\uff0c\u4ee5\u7b80\u5316 GPU \u7f16\u7a0b\uff0c\u4f7f\u7528\u6237\u53ef\u4ee5\u4f7f\u7528 Python \u800c\u975e CUDA \u6216 ROCm \u6765\u5b9e\u73b0 GPU \u5185\u6838\u3002<\/section>\n<section><\/section>\n<section>\u7136\u800c\uff0c\u8fd9\u4e9b\u8bed\u8a00\u4ecd\u7136\u4f9d\u8d56\u7528\u6237\u81ea\u884c\u8bbe\u8ba1 GPU \u4f18\u5316\u6280\u672f\u4ee5\u8fbe\u5230\u66f4\u9ad8\u7684\u6027\u80fd\u3002\u4f8b\u5982\uff0c\u5728 <strong>Triton<\/strong> \u4e2d\u5b9e\u73b0\u4e00\u4e2a FlashAttention \u5185\u6838\u5927\u7ea6\u9700\u8981 <strong>700 \u884c Python<\/strong> \u4ee3\u7801\uff08\u5728 <strong>CUDA<\/strong> \u4e2d\u9700\u8981\u5927\u7ea6 <strong>7,000 \u884c C++ \u4ee3\u7801<\/strong>\uff09\u3002\u5728\u8fd9\u4e9b\u7a0b\u5e8f\u4e2d\uff0c\u7528\u6237\u9700\u8981\u624b\u52a8\u5212\u5206\u7ebf\u7a0b\u5757\u4e4b\u95f4\u7684\u5de5\u4f5c\u8d1f\u8f7d\uff0c\u7ec4\u7ec7\u6bcf\u4e2a\u7ebf\u7a0b\u5757\u5185\u7684\u8ba1\u7b97\uff0c\u5e76\u7ba1\u7406\u5b83\u4eec\u4e4b\u95f4\u7684\u540c\u6b65\u4e0e\u901a\u4fe1\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-6b6c285440309085c19d8bb60055449c.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-6b6c285440309085c19d8bb60055449c.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u7528 Triton \u5b9e\u73b0\u7684 FlashAttention \u7b97\u5b50<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>\u80fd\u5426\u5728\u4e0d\u4f7f\u7528 CUDA\/Triton \u7f16\u7a0b\u7684\u60c5\u51b5\u4e0b\u5c31\u83b7\u5f97\u9ad8\u6548\u7684 GPU \u5185\u6838\u5462\uff1f<\/strong>\u57fa\u4e8e\u8fd9\u4e00\u52a8\u673a\uff0c\u6765\u81ea\u5361\u5185\u57fa\u6885\u9686\u5927\u5b66\u7684 Catalyst Group \u56e2\u961f\u53d1\u5e03\u4e86 Mirage \u9879\u76ee\uff0c\u57fa\u4e8e SuperOptimization \u6280\u672f\uff08https:\/\/arxiv.org\/abs\/2405.05751\uff09\uff0c\u4e3a PyTorch \u81ea\u52a8\u751f\u6210\u9ad8\u6548 GPU \u5185\u6838\u7b97\u5b50\u3002\u4f8b\u5982\uff0c\u5bf9\u4e8e\u4e00\u4e2a FlashAttention \u7b97\u5b50\uff0c\u7528\u6237\u53ea\u9700\u7f16\u5199\u51e0\u884c Python \u4ee3\u7801\u6765\u63cf\u8ff0\u6ce8\u610f\u529b\uff08Attention\uff09\u7684\u8ba1\u7b97\u8fc7\u7a0b\u800c\u65e0\u9700\u4e86\u89e3 GPU \u7f16\u7a0b\u7ec6\u8282\uff0c\u5982\u4e0b\u6240\u793a\uff1a<\/section>\n<section>\n<pre data-lang=\"makefile\"><code># Use Mirage to generate GPU kernels for attention<\/code>\r\n<code>import mirage as mi<\/code>\r\n<code>graph = mi.new_kernel_graph ()<\/code>\r\n<code>Q = graph.new_input (dims=(64, 1, 128), dtype=mi.float16)<\/code>\r\n<code>K = graph.new_input (dims=(64, 128, 4096), dtype=mi.float16)<\/code>\r\n<code>V = graph.new_input (dims=(64, 4096, 128), dtype=mi.float16)<\/code>\r\n<code>A = graph.matmul (Q, K)<\/code>\r\n<code>S = graph.softmax (A)<\/code>\r\n<code>O = graph.matmul (S, V)<\/code>\r\n<code>optimized_graph = graph.superoptimize ()<\/code><\/pre>\n<\/section>\n<section>Mirage \u4f1a\u81ea\u52a8\u641c\u7d22\u53ef\u80fd\u7684 Attention GPU \u5185\u6838\u5b9e\u73b0\uff0c\u641c\u7d22\u7a7a\u95f4\u4e0d\u4ec5\u5305\u62ec\u73b0\u6709\u7684\u624b\u52a8\u8bbe\u8ba1\u7684\u6ce8\u610f\u529b\u5185\u6838\uff08\u5982 FlashAttention \u548c FlashDecoding\uff09\uff0c\u8fd8\u5305\u62ec\u5728\u67d0\u4e9b\u573a\u666f\u4e2d\u6bd4\u76ee\u524d\u7684\u624b\u5199\u7248\u672c\u5feb\u591a\u8fbe 3.5 \u500d\u7684\u5176\u4ed6\u5b9e\u73b0\u3002Mirage \u751f\u6210\u7684 GPU \u5185\u6838\u53ef\u4ee5\u76f4\u63a5\u5728 PyTorch \u5f20\u91cf\u4e0a\u64cd\u4f5c\uff0c\u5e76\u53ef\u4ee5\u5728 PyTorch \u7a0b\u5e8f\u4e2d\u76f4\u63a5\u8c03\u7528\u3002<\/section>\n<section>\n<pre data-lang=\"makefile\"><code>import torch<\/code>\r\n<code>input_tensors = [<\/code>\r\n<code>    torch.randn (64, 1, 128, dtype=torch.float16, device='cuda:0'),<\/code>\r\n<code>    torch.randn (64, 128, 4096, dtype=torch.float16, device='cuda:0'),<\/code>\r\n<code>    torch.randn (64, 4096, 128, dtype=torch.float16, device='cuda:0')<\/code>\r\n<code>]<\/code>\r\n<code># Launch the Mirage-generated kernel to perform attention<\/code>\r\n<code>output = optimized_graph (input_tensors)<\/code><\/pre>\n<\/section>\n<section><\/section>\n<section><strong>Why Mirage?<\/strong><\/section>\n<section><\/section>\n<section>\u4e0e\u4f7f\u7528 CUDA\/Triton \u7f16\u7a0b\u76f8\u6bd4\uff0cMirage \u63d0\u4f9b\u4e86\u4e00\u79cd\u65b0\u7684\u7f16\u7a0b\u8303\u5f0f\uff0c\u5305\u542b\u4e09\u4e2a\u4e3b\u8981\u4f18\u52bf\uff1a<\/section>\n<section><\/section>\n<section><strong>\u66f4\u9ad8\u7684\u751f\u4ea7\u529b<\/strong>\uff1a\u968f\u7740 GPU \u67b6\u6784\u65e5\u65b0\u6708\u5f02\uff0c\u73b0\u4ee3 GPU \u7f16\u7a0b\u9700\u8981\u6301\u7eed\u5b66\u4e60\u5927\u91cf\u7684\u4e13\u4e1a\u77e5\u8bc6\u3002Mirage \u7684\u76ee\u6807\u662f\u63d0\u9ad8\u673a\u5668\u5b66\u4e60\u7cfb\u7edf\u5de5\u7a0b\u5e08\u7684\u751f\u4ea7\u529b \u2014\u2014 \u4ed6\u4eec\u53ea\u9700\u5728 PyTorch \u5c42\u9762\u63cf\u8ff0\u6240\u9700\u7684\u8ba1\u7b97\uff0cMirage \u4fbf\u4f1a\u81ea\u52a8\u751f\u6210\u9002\u7528\u4e8e\u5404\u79cd GPU \u67b6\u6784\u7684\u9ad8\u6027\u80fd\u5b9e\u73b0\u3002\u56e0\u6b64\uff0c\u7a0b\u5e8f\u5458\u4e0d\u518d\u9700\u8981\u624b\u52a8\u7f16\u5199 CUDA\/Triton \u6216\u7279\u5b9a\u67b6\u6784\u7684\u4f4e\u7ea7\u4ee3\u7801\u3002<\/section>\n<section><\/section>\n<section><strong>\u66f4\u597d\u7684\u6027\u80fd<\/strong>\uff1a\u76ee\u524d\u624b\u52a8\u8bbe\u8ba1\u7684 GPU \u5185\u6838\u7531\u4e8e\u65e0\u6cd5\u5145\u5206\u63a2\u7d22\u548c\u5229\u7528\u5404\u79cd GPU \u4f18\u5316\u6280\u672f\uff0c\u5f80\u5f80\u53ea\u80fd\u8fbe\u5230\u6b21\u4f18\u6027\u80fd\u3002Mirage \u53ef\u4ee5\u81ea\u52a8\u641c\u7d22\u4e0e\u8f93\u5165\u7684 PyTorch \u7a0b\u5e8f\u529f\u80fd\u7b49\u4ef7\u7684\u6f5c\u5728 GPU \u5b9e\u73b0\uff0c\u63a2\u7d22\u5e76\u6700\u7ec8\u53d1\u73b0\u6027\u80fd\u6700\u4f18\u7684\u5185\u6838\u3002\u5728\u591a\u4e2a LLM\/GenAI \u57fa\u51c6\u6d4b\u8bd5\u4e2d\u7684\u6d4b\u8bd5\u7ed3\u679c\u663e\u793a\uff0cMirage \u751f\u6210\u7684\u5185\u6838\u901a\u5e38\u6bd4 SOTA \u7684\u4e13\u5bb6\u4eba\u5de5\u7f16\u5199\u6216\u7f16\u8bd1\u5668\u751f\u6210\u7684\u66ff\u4ee3\u65b9\u6848\u5feb 1.2 \u81f3 2.5 \u500d\u3002<\/section>\n<section><\/section>\n<section><strong>\u66f4\u5f3a\u7684\u6b63\u786e\u6027<\/strong>\uff1a\u624b\u52a8\u5b9e\u73b0\u7684 CUDA\/Triton GPU \u5185\u6838\u5bb9\u6613\u51fa\u9519\uff0c\u800c\u4e14 GPU \u5185\u6838\u4e2d\u7684\u9519\u8bef\u96be\u4ee5\u8c03\u8bd5\u548c\u5b9a\u4f4d\uff0c\u800c Mirage \u5219\u5229\u7528\u5f62\u5f0f\u5316\u9a8c\u8bc1\uff08Formal Verification\uff09\u6280\u672f\u81ea\u52a8\u9a8c\u8bc1\u751f\u6210\u7684 GPU \u5185\u6838\u7684\u6b63\u786e\u6027\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-08cbc6c66a1a611e05c227f598e63d21.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-08cbc6c66a1a611e05c227f598e63d21.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0LLaMA-3-8B \u548c Chameleon-7B \u7aef\u5230\u7aef\u63a8\u7406\u5ef6\u8fdf\u5bf9\u6bd4\uff08NVIDIA A100\uff0cbatch size=1\uff0ccontext length=4K\uff09\uff0c\u76f8\u6bd4\u4e8e CUDA\/Triton \u7684\u5b9e\u73b0\uff0cMirage \u53ef\u4ee5\u5b9e\u73b0 15-20% \u7684\u52a0\u901f<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>GPU \u67b6\u6784\u4e0e Mirage \u4e2d\u7684<\/strong><strong><strong><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-7c83d7c9061f2d41a5e665527caff1e5.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-7c83d7c9061f2d41a5e665527caff1e5.png\"><\/a><\/strong><\/strong><\/section>\n<section><\/section>\n<section>GPU \u8ba1\u7b97\u7684\u5185\u6838\u51fd\u6570\u4ee5\u5355\u7a0b\u5e8f\u591a\u6570\u636e\uff08SPMD\uff09\u65b9\u5f0f\u5728\u591a\u4e2a\u6d41\u5904\u7406\u5668\uff08SM\uff09\u4e0a\u540c\u65f6\u8fd0\u884c\u3002GPU \u5185\u6838\uff08Kernel\uff09\u501f\u52a9\u7531\u7ebf\u7a0b\u5757\uff08Thread Block\uff09\u7ec4\u6210\u7684\u7f51\u683c\u7ed3\u6784\u6765\u7ec4\u7ec7\u5176\u8ba1\u7b97\uff0c\u6bcf\u4e2a\u7ebf\u7a0b\u5757\u5728\u5355\u4e2a SM \u4e0a\u8fd0\u884c\u3002\u6bcf\u4e2a\u5757\u8fdb\u4e00\u6b65\u5305\u542b\u591a\u4e2a\u7ebf\u7a0b\uff08Thread\uff09\uff0c\u4ee5\u5bf9\u5355\u72ec\u7684\u6570\u636e\u5143\u7d20\u8fdb\u884c\u8ba1\u7b97\u3002GPU \u8fd8\u62e5\u6709\u590d\u6742\u7684\u5185\u5b58\u5c42\u6b21\u7ed3\u6784\uff0c\u4ee5\u652f\u6301\u8fd9\u79cd\u590d\u6742\u7684\u5904\u7406\u7ed3\u6784\u3002\u6bcf\u4e2a\u7ebf\u7a0b\u90fd\u6709\u81ea\u5df1\u7684\u5bc4\u5b58\u5668\u6587\u4ef6\uff08Register File\uff09\uff0c\u4ee5\u4fbf\u5feb\u901f\u8bbf\u95ee\u6570\u636e\u3002\u7ebf\u7a0b\u5757\u5185\u7684\u6240\u6709\u7ebf\u7a0b\u53ef\u4ee5\u8bbf\u95ee\u4e00\u4e2a\u516c\u5171\u7684\u5171\u4eab\u5185\u5b58\uff08Shared Memory\uff09\uff0c\u8fd9\u6709\u52a9\u4e8e\u5b83\u4eec\u4e4b\u95f4\u9ad8\u6548\u7684\u6570\u636e\u4ea4\u6362\u548c\u96c6\u4f53\u64cd\u4f5c\u3002\u6700\u540e\uff0c\u5185\u6838\u5185\u7684\u6240\u6709\u7ebf\u7a0b\u53ef\u4ee5\u8bbf\u95ee\u5206\u914d\u7ed9\u6574\u4e2a GPU \u7684\u5927\u578b\u8bbe\u5907\u5185\u5b58\uff08Device Memory\uff09\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-386d9c16f8428b865324d0450e312c94.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-386d9c16f8428b865324d0450e312c94.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 GPU \u8ba1\u7b97\u67b6\u6784\u548c\u7f16\u7a0b\u62bd\u8c61\u793a\u610f\u56fe<\/sup><\/em><\/section>\n<section><\/section>\n<section>Mirage \u4f7f\u7528<strong><strong><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-3b4011c26a402594b44a1d474edd712a.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-3b4011c26a402594b44a1d474edd712a.png\"><\/a><\/strong><\/strong>\u6765\u63cf\u8ff0 GPU \u5185\u6838\uff0c<strong><strong><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-982d17dbd1fbcef24844169199424b3e.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-982d17dbd1fbcef24844169199424b3e.png\"><\/a><\/strong><\/strong>\u5305\u542b\u591a\u4e2a\u5c42\u6b21\uff0c\u4ee3\u8868\u5185\u6838\u3001\u7ebf\u7a0b\u5757\u548c\u7ebf\u7a0b\u7ea7\u522b\u7684\u8ba1\u7b97\u3002\u5927\u4f53\u4e0a\uff0cKernel Graph\u3001Thread Block Graph \u548c Thread Graph \u5206\u522b\u4ee3\u8868\u6574\u4e2a GPU\u3001\u4e00\u4e2a\u6d41\u5904\u7406\u5668\uff08SM\uff09\u548c\u4e00\u4e2a CUDA\/tensor \u6838\u5fc3\u4e0a\u7684\u8ba1\u7b97\u3002<\/section>\n<section><\/section>\n<section>\u5bf9<strong><strong><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-49a0fc00b7e9cf11d821ce1a04fa5e16.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-49a0fc00b7e9cf11d821ce1a04fa5e16.png\"><\/a><\/strong><\/strong>\u7ec6\u8282\u611f\u5174\u8da3\u7684\u8bfb\u8005\u53ef\u4ee5\u53c2\u8003\uff1a<\/section>\n<ul>\n<li>\n<section>https:\/\/mirage-project.readthedocs.io\/en\/latest\/mugraph.html<\/section>\n<\/li>\n<\/ul>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-87bfc4a88948ba55e5940efe659d6c8c.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-87bfc4a88948ba55e5940efe659d6c8c.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0Mirage \u5de5\u4f5c\u6d41\u793a\u610f\u56fe<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u4e0a\u56fe\u5c55\u793a\u4e86 Mirage \u7684\u5de5\u4f5c\u6d41\u7a0b\uff1a\u5bf9\u4e8e\u8f93\u5165\u7684 PyTorch \u7a0b\u5e8f\uff0cMirage \u7684<strong><strong><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-6465bae91b17565b7baaf2289febca3e.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-6465bae91b17565b7baaf2289febca3e.png\"><\/a><\/strong><\/strong>\u751f\u6210\u5668\u81ea\u52a8\u641c\u7d22\u4e0e\u8f93\u5165\u7a0b\u5e8f\u529f\u80fd\u7b49\u4ef7\u7684\u5176\u4ed6<strong><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-96280a7d1265cb8e21b71e68dfe6173e.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-96280a7d1265cb8e21b71e68dfe6173e.png\"><\/a><\/strong>\uff0c\u641c\u7d22\u7a7a\u95f4\u6db5\u76d6\u4e86\u5185\u6838\u3001\u7ebf\u7a0b\u5757\u548c\u7ebf\u7a0b\u7ea7\u522b\u7684\u5404\u79cd GPU \u4f18\u5316\u3002\u6240\u6709\u751f\u6210\u7684<strong><strong><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-c068c0bbec2ab8326ea23587484bf90d.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-c068c0bbec2ab8326ea23587484bf90d.png\"><\/a><\/strong><\/strong>\u90fd\u88ab\u53d1\u9001\u5230\u7b49\u4ef7\u6027\u9a8c\u8bc1\u5668\uff0c\u8be5\u9a8c\u8bc1\u5668\u81ea\u52a8\u68c0\u67e5\u6bcf\u4e2a<strong><strong><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-c074fd70fca2cbac4d1a7dee69b4bf74.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-c074fd70fca2cbac4d1a7dee69b4bf74.png\"><\/a><\/strong><\/strong>\u662f\u5426\u4e0e\u6240\u9700\u7a0b\u5e8f\u7b49\u4ef7\u3002\u6700\u540e\uff0c<strong><strong><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-a6fc16fb31b758957e9571cc68405a6b.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-a6fc16fb31b758957e9571cc68405a6b.png\"><\/a><\/strong><\/strong>\u8f6c\u8bd1\u5668\u5c06\u6240\u6709\u7ecf\u8fc7\u9a8c\u8bc1\u7684<strong><strong><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-427630e97986e03ed89134f5b7107e0f.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-427630e97986e03ed89134f5b7107e0f.png\"><\/a><\/strong><\/strong>\u8f6c\u8bd1\u4e3a CUDA \u5185\u6838\u3002\u6700\u540e\uff0cMirage \u4f1a\u4ece\u4e2d\u8fd4\u56de\u6027\u80fd\u6700\u4f73\u7684 CUDA \u5185\u6838\u3002<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u9879\u76ee\u6210\u5458\uff1aMengdi Wu (CMU), Xinhao Cheng (CMU), Shengyu Liu (PKU), Chuan Shi (PKU), Jianan Ji (CMU), Oded Padon (VMWare), Xupeng Miao (Purdue), Zhihao Jia (CMU)<\/section>\n<\/li>\n<li>\n<section>\u9879\u76ee\u5730\u5740\uff1ahttps:\/\/github.com\/mirage-project\/mirage<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section><strong>\u4e3a\u4ec0\u4e48 Mirage \u751f\u6210\u7684\u5185\u6838\u66f4\u9ad8\u6548\uff1f<\/strong><\/section>\n<section><\/section>\n<section>\u5728\u591a\u4e2a LLM\/GenAI \u57fa\u51c6\u6d4b\u8bd5\u4e2d\u7684\u6d4b\u8bd5\u7ed3\u679c\u663e\u793a\uff0cMirage \u751f\u6210\u7684\u5185\u6838\u901a\u5e38\u6bd4\u73b0\u6709\u7684\u624b\u5199\u6216\u7f16\u8bd1\u5668\u751f\u6210\u7684\u5185\u6838\u5feb 1.2 \u81f3 2.5 \u500d\u3002\u63a5\u4e0b\u6765\uff0c\u672c\u6587\u4ee5 LLM \u4e2d\u7684 Transformer \u67b6\u6784\u4e3a\u4f8b\uff0c\u5c55\u793a\u73b0\u6709\u7cfb\u7edf\u4e2d\u7f3a\u5931\u7684\u51e0\u9879 GPU \u7a0b\u5e8f\u4f18\u5316\u6280\u672f\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-9df26419bab33318645dd066ec2dbbe1.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-9df26419bab33318645dd066ec2dbbe1.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0Transformer \u67b6\u6784\u793a\u610f\u56fe<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>Case 1: Normalization + Linear<\/strong><\/section>\n<section><\/section>\n<section>\u5f52\u4e00\u5316\uff08Normalization\uff09\u64cd\u4f5c\uff0c\u5982 LayerNorm\u3001RMSNorm\u3001GroupNorm \u548c BatchNorm\uff0c\u5e7f\u6cdb\u5e94\u7528\u4e8e\u5f53\u4eca\u7684\u673a\u5668\u5b66\u4e60\u6a21\u578b\u3002\u5f53\u524d\u7684\u673a\u5668\u5b66\u4e60\u7f16\u8bd1\u5668\u901a\u5e38\u5728\u72ec\u7acb\u7684\u5185\u6838\u4e2d\u542f\u52a8\u5f52\u4e00\u5316\u5c42\uff0c\u56e0\u4e3a\u5f52\u4e00\u5316\u6d89\u53ca\u5230\u5f52\u7ea6\u548c\u5e7f\u64ad\uff0c\u96be\u4ee5\u4e0e\u5176\u4ed6\u8ba1\u7b97\u878d\u5408\u3002\u7136\u800c\uff0cMirage \u53d1\u73b0\uff0c\u5927\u591a\u6570\u5f52\u4e00\u5316\u5c42\u53ef\u4ee5\u901a\u8fc7\u8fdb\u884c\u9002\u5f53\u7684\u4ee3\u6570\u53d8\u6362\uff0c\u4e0e\u540e\u7eed\u7684\u7ebf\u6027\u5c42\uff08\u5982 MatMul\uff09\u878d\u5408\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-758f2f4c3157aa82212b63fd17ac1c97.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-758f2f4c3157aa82212b63fd17ac1c97.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 Normalization + Linear \u73b0\u6709\u5185\u6838 v.s. Mirage \u53d1\u73b0\u7684\u5185\u6838<\/sup><\/em><\/section>\n<section><\/section>\n<section>Mirage \u53d1\u73b0\u7684\u81ea\u5b9a\u4e49\u5185\u6838\u5229\u7528\u4e86 RMSNorm \u4e2d\u7684\u9664\u6cd5\u548c MatMul \u4e2d\u7684\u4e58\u6cd5\u7684\u53ef\u4ea4\u6362\u6027\uff0c\u5c06\u9664\u6cd5\u79fb\u5230 MatMul \u4e4b\u540e\u3002\u8fd9\u4e00\u53d8\u6362\u4fdd\u6301\u4e86\u529f\u80fd\u7b49\u4ef7\u6027\uff0c\u540c\u65f6\u907f\u514d\u4e86\u4e2d\u95f4\u5f20\u91cf Y \u7684\u5b9e\u4f8b\u5316\u3002\u8be5\u5185\u6838\u7684\u6027\u80fd\u6bd4\u5355\u72ec\u8fd0\u884c\u8fd9\u4e24\u4e2a\u64cd\u4f5c\u5feb 1.5 \u5230 1.7 \u500d\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-023f00b55cb35d373cca1a493af23976.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-023f00b55cb35d373cca1a493af23976.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0Normalization + Linear \u5185\u6838\u6027\u80fd\u5bf9\u6bd4<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>Case 2: LoRA + Linear<\/strong><\/section>\n<section><\/section>\n<section>LoRA \u5e7f\u6cdb\u7528\u4e8e\u9884\u8bad\u7ec3\u6a21\u578b\u7684\u5fae\u8c03\u573a\u666f\uff0c\u4ee5\u9002\u914d\u5230\u7279\u5b9a\u9886\u57df\u548c\u4efb\u52a1\u3002\u8fd9\u4e9b LoRA \u9002\u914d\u5668\u901a\u5e38\u4f1a\u88ab\u63d2\u5165\u5230\u6a21\u578b\u7684\u7ebf\u6027\u5c42\u4e2d\uff0c\u5f15\u5165\u989d\u5916\u7684\u77e9\u9635\u4e58\u6cd5\u3002\u73b0\u6709\u7cfb\u7edf\u901a\u5e38\u4e3a\u539f\u59cb\u77e9\u9635\u4e58\u6cd5\u548c LoRA \u4e2d\u7684\u4e24\u4e2a\u77e9\u9635\u4e58\u6cd5\u542f\u52a8\u72ec\u7acb\u7684\u5185\u6838\uff0c\u4ece\u800c\u5bfc\u81f4\u8f83\u9ad8\u7684\u5185\u6838\u542f\u52a8\u5f00\u9500\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-e82cc83b104423b3261d2179aac99a2f.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-e82cc83b104423b3261d2179aac99a2f.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 LoRA+Linear \u73b0\u6709\u5185\u6838 v.s. Mirage \u53d1\u73b0\u7684\u5185\u6838<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u5982\u4e0a\u56fe\u6240\u793a\uff0cMirage \u53d1\u73b0\u4e86\u4e00\u4e2a\u5c06\u4e09\u4e2a\u77e9\u9635\u4e58\u6cd5\u548c\u968f\u540e\u7684\u52a0\u6cd5\u878d\u5408\u4e3a\u5355\u4e2a\u5185\u6838\u7684\u5185\u6838\u3002\u8fd9\u662f\u901a\u8fc7\u5c06\u8ba1\u7b97\u91cd\u7ec4\u4e3a\u4e24\u4e2a\u7ebf\u7a0b\u5757\u7ea7\u522b\u7684\u77e9\u9635\u4e58\u6cd5\u5b9e\u73b0\u7684\uff0c\u5229\u7528\u4e86\u4ee5\u4e0b\u4ee3\u6570\u53d8\u6362\uff1aW\u00d7X+B\u00d7A\u00d7X=(W|B)\u00d7(X|(A\u00d7X))\uff0c\u5176\u4e2d\u7684\u4e24\u4e2a\u62fc\u63a5\u64cd\u4f5c\u4e0d\u6d89\u53ca\u4efb\u4f55\u8ba1\u7b97\uff0c\u800c\u662f\u901a\u8fc7\u5728 GPU \u5171\u4eab\u5185\u5b58\u4e2d\u66f4\u65b0\u5f20\u91cf\u504f\u79fb\u91cf\u6765\u5b8c\u6210\u3002Mirage \u53d1\u73b0\u7684\u5185\u6838\u6bd4\u73b0\u6709\u7cfb\u7edf\u4e2d\u4f7f\u7528\u7684\u5185\u6838\u5feb 1.6 \u500d\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-d31614a5735b2422b5eb3b29eab5f815.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-d31614a5735b2422b5eb3b29eab5f815.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 LoRA+Linear \u5185\u6838\u6027\u80fd\u5bf9\u6bd4<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>Case 3: Gated MLP<\/strong><\/section>\n<section><\/section>\n<section>Gated MLP \u5c42\u76ee\u524d\u5728\u8bb8\u591a LLM \u4e2d\u4f7f\u7528\uff08\u5982 LLAMA-2\u3001LLAMA-3 \u53ca\u5176\u53d8\u4f53\uff09\uff0c\u5b83\u7684\u8f93\u5165\u5f20\u91cf X \u4e0e\u4e24\u4e2a\u6743\u91cd\u77e9\u9635\u76f8\u4e58\uff0c\u8f93\u51fa\u7ed3\u679c\u88ab\u7ec4\u5408\u4ee5\u4ea7\u751f\u6700\u7ec8\u7ed3\u679c\u3002Mirage \u53d1\u73b0\u4e86\u4e00\u4e2a\u5185\u6838\uff0c\u8be5\u5185\u6838\u6267\u884c\u4e24\u4e2a\u77e9\u9635\u4e58\u6cd5\u3001SiLU \u6fc0\u6d3b\u4ee5\u53ca\u968f\u540e\u7684\u9010\u5143\u7d20\u4e58\u6cd5\uff0c\u4ece\u800c\u51cf\u5c11\u4e86\u5185\u6838\u542f\u52a8\u5f00\u9500\u548c\u5bf9\u8bbe\u5907\u5185\u5b58\u7684\u8bbf\u95ee\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-596f5777fad7a1dbf828d7a4bb011cc5.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-596f5777fad7a1dbf828d7a4bb011cc5.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0Gated MLP \u73b0\u6709\u5185\u6838 v.s. Mirage \u53d1\u73b0\u7684\u5185\u6838<\/sup><\/em><\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-0b780535653b1fac2f4f023ce1fce111.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-0b780535653b1fac2f4f023ce1fce111.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0Gated MLP \u5185\u6838\u6027\u80fd\u5bf9\u6bd4<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>Case 4: Attention Variants<\/strong><\/section>\n<section><\/section>\n<section>\u5982\u4eca\u7684\u5927\u591a\u6570 LLM \u57fa\u4e8e\u6ce8\u610f\u529b\u53ca\u5176\u53d8\u4f53\uff0c\u867d\u7136\u73b0\u6709\u7cfb\u7edf\u901a\u5e38\u63d0\u4f9b\u9ad8\u5ea6\u4f18\u5316\u7684\u6ce8\u610f\u529b\u5b9e\u73b0\uff0c\u5982 FlashAttention\u3001FlashInfer \u548c FlexAttention\uff0c\u4f46\u652f\u6301\u6ce8\u610f\u529b\u53d8\u4f53\u901a\u5e38\u9700\u8981\u65b0\u7684\u81ea\u5b9a\u4e49\u5185\u6838\u3002\u4e0b\u9762\u7528\u4e24\u4e2a\u4f8b\u5b50\u6765\u5c55\u793a Mirage \u5982\u4f55\u4e3a\u975e\u5e38\u89c4\u6ce8\u610f\u529b\u8ba1\u7b97\u53d1\u73b0\u81ea\u5b9a\u4e49 GPU \u5185\u6838\u3002<\/section>\n<section><\/section>\n<section>Case 4.1: Attention with Query-Key Normalization<\/section>\n<section><\/section>\n<section>\u8bb8\u591a\u6700\u8fd1\u7684 LLM \u67b6\u6784\uff08\u5305\u62ec Chameleon\u3001ViT-22B \u7b49\uff09\u5728 LLaMA \u67b6\u6784\u4e2d\u5f15\u5165\u4e86 QK-Norm \u6765\u7f13\u89e3\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7684\u6570\u503c\u53d1\u6563\u95ee\u9898\u3002QK-Norm \u5728\u6ce8\u610f\u529b\u4e4b\u524d\u5bf9 Query \u548c Key \u5411\u91cf\u5e94\u7528 LayerNorm \u5c42\u3002\u73b0\u6709\u6ce8\u610f\u529b\u5b9e\u73b0\u4e2d\u5e76\u4e0d\u652f\u6301\u8fd9\u4e9b\u989d\u5916\u7684\u5f52\u4e00\u5316\u5c42\uff0c\u5e76\u4e14\u5b83\u4eec\u8fd8\u9700\u8981\u4f5c\u4e3a\u72ec\u7acb\u5185\u6838\u542f\u52a8\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-8ccdfeee0a08df8bb63b33c9da0d5501.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-8ccdfeee0a08df8bb63b33c9da0d5501.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 QK-Norm \u6ce8\u610f\u529b\u73b0\u6709\u5185\u6838 v.s. Mirage \u53d1\u73b0\u7684\u5185\u6838<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u5bf9\u4e8e\u5728\u6ce8\u610f\u529b\u4e4b\u524d\u548c \/ \u6216\u4e4b\u540e\u5f15\u5165\u8ba1\u7b97\u7684\u6ce8\u610f\u529b\u53d8\u4f53\uff0c\u8fd9\u4e9b\u8ba1\u7b97\u53ef\u4ee5\u4e0e\u6ce8\u610f\u529b\u878d\u5408\u4ee5\u63d0\u9ad8 GPU \u6027\u80fd\uff0c\u800c\u8fd9\u9700\u8981\u81ea\u5b9a\u4e49\u5185\u6838\u3002\u5bf9\u4e8e\u5e26\u6709 QK-Norm \u7684\u6ce8\u610f\u529b\uff0cMirage \u53d1\u73b0\u4e86\u4e0a\u8ff0\u5185\u6838\u6765\u878d\u5408\u8ba1\u7b97\uff0c\u4ece\u800c\u907f\u514d\u5728 GPU \u8bbe\u5907\u5185\u5b58\u4e2d\u5b9e\u4f8b\u5316\u4e2d\u95f4\u7ed3\u679c\u3002\u8fd9\u4e2a\u81ea\u5b9a\u4e49\u5185\u6838\u8fd8\u5bf9\u6ce8\u610f\u529b\u8fdb\u884c\u4e86\u73b0\u6709\u7684 GPU \u4f18\u5316\uff0c\u5b9e\u73b0\u4e86 1.7 \u81f3 2.5 \u500d\u7684\u6027\u80fd\u63d0\u5347\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-f8fc67a3a06b13e4be3a8fd754a82331.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-f8fc67a3a06b13e4be3a8fd754a82331.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 QK-Norm \u6ce8\u610f\u529b\u5185\u6838\u6027\u80fd\u5bf9\u6bd4<\/sup><\/em><\/section>\n<section><\/section>\n<section>Case 4.2: Multi-Head Latent Attention<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-65f66e6ee47bb62a14f49818d4cc0ad1.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-65f66e6ee47bb62a14f49818d4cc0ad1.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 MLA \u7684\u73b0\u6709\u5185\u6838 v.s. Mirage \u53d1\u73b0\u7684\u5185\u6838<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u53e6\u4e00\u4e2a\u5e38\u7528\u7684\u6ce8\u610f\u529b\u53d8\u4f53\u662f MLA\uff08Multi-Head Latent Attention\uff09\uff0c\u5b83\u5c06\u6ce8\u610f\u529b\u7684 KV Cache \u538b\u7f29\u4e3a\u4e00\u4e2a\u5411\u91cf\uff0c\u4ee5\u51cf\u5c11\u5b58\u50a8 KV Cache \u7684\u5185\u5b58\u5f00\u9500\u3002\u8fd9\u4e00\u53d8\u5316\u8fd8\u5728\u6ce8\u610f\u529b\u4e4b\u524d\u5f15\u5165\u4e86\u4e24\u4e2a\u7ebf\u6027\u5c42\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002\u4e0e QK-Norm \u7c7b\u4f3c\uff0c\u73b0\u6709\u6ce8\u610f\u529b\u5b9e\u73b0\u4e2d\u5e76\u4e0d\u652f\u6301\u8fd9\u4e9b\u989d\u5916\u7684\u5f52\u4e00\u5316\u5c42\uff0c\u540c\u6837\u9700\u8981\u4f5c\u4e3a\u72ec\u7acb\u5185\u6838\u542f\u52a8\uff0c\u800c Mirage \u53ef\u4ee5\u5c06\u7ebf\u6027\u5c42\u548c\u6ce8\u610f\u529b\u878d\u5408\u4e3a\u4e00\u4e2a\u5355\u72ec\u7684\u81ea\u5b9a\u4e49\u5185\u6838\u3002<\/section>\n<section><\/section>\n<section><strong>\u957f\u671f\u613f\u666f<\/strong><\/section>\n<section><\/section>\n<section>Mirage \u9879\u76ee\u7684\u957f\u671f\u76ee\u6807\u662f\u5e0c\u671b\u80fd\u591f\u8ba9\u672a\u6765\u7684 AI \u5f00\u53d1\u8005\u65e0\u9700\u5b66\u4e60 CUDA \u6216\u8005 Triton \u7b49\u590d\u6742\u7684 GPU \u7f16\u7a0b\u8bed\u8a00\uff0c\u53ea\u9700\u6307\u5b9a\u6240\u9700\u7684\u6570\u5b66\u64cd\u4f5c\uff0c\u5c31\u80fd\u5728 GPU \u4e0a\u8f7b\u677e\u5b9e\u73b0 AI \u6a21\u578b\u3002\u901a\u8fc7\u5229\u7528 Mirage \u7684 SuperOptimization \u6280\u672f\uff0c\u5404\u79cd\u8ba1\u7b97\u4efb\u52a1\u53ef\u4ee5\u81ea\u52a8\u8f6c\u6362\u4e3a\u9ad8\u5ea6\u4f18\u5316\u7684 GPU \u5b9e\u73b0\u3002\u968f\u7740 LLM \u548c\u5176\u4ed6\u751f\u6210\u5f0f AI \u5e94\u7528\u7684\u98de\u901f\u53d1\u5c55\uff0c\u5728\u5404\u79cd\u5b9e\u9645\u90e8\u7f72\u573a\u666f\u90fd\u9700\u8981\u9ad8\u6548\u7684 GPU \u652f\u6301\uff0c\u964d\u4f4e GPU \u7f16\u7a0b\u95e8\u69db\u5e76\u63d0\u9ad8\u7a0b\u5e8f\u6548\u7387\u4e5f\u6108\u53d1\u91cd\u8981\u3002<\/section>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-10-06\" target=\"_blank\">\u544a\u522bCUDA\u65e0\u9700Triton\uff01Mirage\u96f6\u95e8\u69db\u751f\u6210PyTorch\u7b97\u5b50\uff0c\u4eba\u5747GPU\u7f16\u7a0b\u5927\u5e08\uff1f<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u544a\u522bCUDA\u65e0\u9700Tri [&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":[71,55],"class_list":["post-35923","post","type-post","status-publish","format-standard","hentry","category-news","tag-rag","tag-55"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v27.9 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>\u544a\u522bCUDA\u65e0\u9700Triton\uff01Mirage\u96f6\u95e8\u69db\u751f\u6210PyTorch\u7b97\u5b50\uff0c\u4eba\u5747GPU\u7f16\u7a0b\u5927\u5e08\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=35923\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35923#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35923\"},\"author\":{\"name\":\"AI\u5c0f\u52a9\u624b\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\"},\"headline\":\"\u544a\u522bCUDA\u65e0\u9700Triton\uff01Mirage\u96f6\u95e8\u69db\u751f\u6210PyTorch\u7b97\u5b50\uff0c\u4eba\u5747GPU\u7f16\u7a0b\u5927\u5e08\uff1f\",\"datePublished\":\"2025-01-02T22:00:11+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35923\"},\"wordCount\":348,\"commentCount\":0,\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35923#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/10\\\/frc-ce394fffbfc5469c03c43a0ff80958ce.png\",\"keywords\":[\"RAG\",\"\u673a\u5668\u5b66\u4e60\"],\"articleSection\":[\"\u884c\u4e1a\u8d44\u8baf\"],\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=35923#respond\"]}]},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35923\",\"url\":\"https:\\\/\\\/17aitech.com\\\/?p=35923\",\"name\":\"\u544a\u522bCUDA\u65e0\u9700Triton\uff01Mirage\u96f6\u95e8\u69db\u751f\u6210PyTorch\u7b97\u5b50\uff0c\u4eba\u5747GPU\u7f16\u7a0b\u5927\u5e08\uff1f - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35923#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35923#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/10\\\/frc-ce394fffbfc5469c03c43a0ff80958ce.png\",\"datePublished\":\"2025-01-02T22:00:11+00:00\",\"author\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35923#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=35923\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35923#primaryimage\",\"url\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/10\\\/frc-ce394fffbfc5469c03c43a0ff80958ce.png\",\"contentUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/10\\\/frc-ce394fffbfc5469c03c43a0ff80958ce.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35923#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\\\/\\\/17aitech.com\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u544a\u522bCUDA\u65e0\u9700Triton\uff01Mirage\u96f6\u95e8\u69db\u751f\u6210PyTorch\u7b97\u5b50\uff0c\u4eba\u5747GPU\u7f16\u7a0b\u5927\u5e08\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\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/robot_3.png\",\"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":"\u544a\u522bCUDA\u65e0\u9700Triton\uff01Mirage\u96f6\u95e8\u69db\u751f\u6210PyTorch\u7b97\u5b50\uff0c\u4eba\u5747GPU\u7f16\u7a0b\u5927\u5e08\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=35923","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/17aitech.com\/?p=35923#article","isPartOf":{"@id":"https:\/\/17aitech.com\/?p=35923"},"author":{"name":"AI\u5c0f\u52a9\u624b","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"headline":"\u544a\u522bCUDA\u65e0\u9700Triton\uff01Mirage\u96f6\u95e8\u69db\u751f\u6210PyTorch\u7b97\u5b50\uff0c\u4eba\u5747GPU\u7f16\u7a0b\u5927\u5e08\uff1f","datePublished":"2025-01-02T22:00:11+00:00","mainEntityOfPage":{"@id":"https:\/\/17aitech.com\/?p=35923"},"wordCount":348,"commentCount":0,"image":{"@id":"https:\/\/17aitech.com\/?p=35923#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-ce394fffbfc5469c03c43a0ff80958ce.png","keywords":["RAG","\u673a\u5668\u5b66\u4e60"],"articleSection":["\u884c\u4e1a\u8d44\u8baf"],"inLanguage":"zh-Hans","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/17aitech.com\/?p=35923#respond"]}]},{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=35923","url":"https:\/\/17aitech.com\/?p=35923","name":"\u544a\u522bCUDA\u65e0\u9700Triton\uff01Mirage\u96f6\u95e8\u69db\u751f\u6210PyTorch\u7b97\u5b50\uff0c\u4eba\u5747GPU\u7f16\u7a0b\u5927\u5e08\uff1f - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=35923#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=35923#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-ce394fffbfc5469c03c43a0ff80958ce.png","datePublished":"2025-01-02T22:00:11+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=35923#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=35923"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=35923#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-ce394fffbfc5469c03c43a0ff80958ce.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-ce394fffbfc5469c03c43a0ff80958ce.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=35923#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u544a\u522bCUDA\u65e0\u9700Triton\uff01Mirage\u96f6\u95e8\u69db\u751f\u6210PyTorch\u7b97\u5b50\uff0c\u4eba\u5747GPU\u7f16\u7a0b\u5927\u5e08\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":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","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\/35923","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=35923"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/35923\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=35923"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=35923"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=35923"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}