{"id":2217,"date":"2025-02-24T07:20:32","date_gmt":"2025-02-24T07:20:32","guid":{"rendered":"https:\/\/oimi.me\/?p=2217"},"modified":"2025-02-24T07:20:32","modified_gmt":"2025-02-24T07:20:32","slug":"deepseek%e5%bc%80%e6%ba%90%e5%91%a8%e9%a6%96%e6%97%a5%e6%8e%a8%e5%87%baflashmla%e9%a1%b9%e7%9b%ae-%e5%8f%af%e4%bb%a5%e6%98%be%e8%91%97%e9%99%8d%e4%bd%8e%e5%86%85%e5%ad%98%e5%8d%a0%e7%94%a8%e5%92%8c","status":"publish","type":"post","link":"https:\/\/myya.net\/index.php\/2025\/02\/24\/deepseek%e5%bc%80%e6%ba%90%e5%91%a8%e9%a6%96%e6%97%a5%e6%8e%a8%e5%87%baflashmla%e9%a1%b9%e7%9b%ae-%e5%8f%af%e4%bb%a5%e6%98%be%e8%91%97%e9%99%8d%e4%bd%8e%e5%86%85%e5%ad%98%e5%8d%a0%e7%94%a8%e5%92%8c\/","title":{"rendered":"DeepSeek\u5f00\u6e90\u5468\u9996\u65e5\u63a8\u51faFlashMLA\u9879\u76ee \u53ef\u4ee5\u663e\u8457\u964d\u4f4e\u5185\u5b58\u5360\u7528\u548c\u8ba1\u7b97\u5f00\u9500"},"content":{"rendered":"

\u5f00\u6e90\u4eba\u5de5\u667a\u80fd\u6280\u672f\u5f00\u53d1\u5546 DeepSeek \u4e0a\u5468\u5df2\u7ecf\u9884\u544a\u5c06\u5728\u672c\u5468\u9646\u7eed\u63a8\u51fa 5 \u4e2a\u5df2\u7ecf\u5728\u751f\u4ea7\u73af\u5883\u4e2d\u4f7f\u7528\u7684\u6280\u672f\uff0c\u76ee\u524d\u9996\u4e2a\u9879\u76ee FlashMLA \u5df2\u7ecf\u5728 GitHub \u4e0a\u516c\u5e03\u3002<\/p>\n

FlashMLA \u662f\u4e00\u79cd\u9488\u5bf9 NVIDIA Grace Hopper \u67b6\u6784 GPU \u8bbe\u8ba1\u7684\u9ad8\u6548\u591a\u5c42\u6ce8\u610f\u529b (Multi-Layer Attention\uff0cMLA) \u89e3\u7801\u5185\u6838\uff0c\u8be5\u6280\u672f\u4e0d\u4ec5\u53ef\u4ee5\u4f18\u5316\u53d8\u957f\u5e8f\u5217\u7684\u5904\u7406\u6027\u80fd\uff0c\u8fd8\u53ef\u4ee5\u5c06\u4f4e\u5185\u5b58\u5360\u7528\u548c\u8ba1\u7b97\u5f00\u9500\u3002<\/p>\n

\u6280\u672f\u7684\u5173\u952e\u7279\u70b9\u5305\u62ec\uff1a<\/strong><\/p>\n

BF16 \u652f\u6301\uff1aFlashMLA \u91c7\u7528 BF16 \u7cbe\u5ea6\u683c\u5f0f\uff0c\u517c\u987e FP32 \u7684\u52a8\u6001\u8303\u56f4\u548c FP16 \u7684\u8ba1\u7b97\u6548\u7387\uff0c\u8fd9\u79cd\u8bbe\u8ba1\u53ef\u4ee5\u663e\u8457\u964d\u4f4e\u5185\u5b58\u5360\u7528\u548c\u8ba1\u7b97\u5f00\u9500\uff0c\u7279\u522b\u9002\u5408\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u7684\u63a8\u7406\u9636\u6bb5\u3002<\/p>\n

\u5206\u9875 KV \u7f13\u5b58\u6280\u672f\uff1aPaged KV Cache \u901a\u8fc7\u5757\u5927\u5c0f\u4e3a 64 \u7684\u5206\u9875\u952e\u7f13\u5b58\u7cfb\u7edf\uff0cFlashMLA \u4f18\u5316\u4e86 Transformer \u6a21\u578b\u4e2d\u952e\u503c\u5bf9\u7684\u5b58\u50a8\u548c\u8bbf\u95ee\uff0c\u51cf\u5c11\u5185\u5b58\u788e\u7247\u548c\u5ef6\u8fdf\u7b49\uff0c\u8fd9\u9879\u6280\u672f\u4e3b\u8981\u662f\u548c\u5904\u7406\u53d8\u957f\u5e8f\u5217\uff0c\u786e\u4fdd\u5728\u4e0d\u540c\u8f93\u5165\u957f\u5ea6\u4e0b\u90fd\u80fd\u4fdd\u6301\u9ad8\u6548\u6027\u80fd\u3002<\/p>\n

\u5353\u8d8a\u6027\u80fd\uff1a\u5728 NVIDIA H800 GPU \u4e0a\uff0cFlashMLA \u5b9e\u73b0\u4e86 3000GB \/ \u79d2\u7684\u5185\u5b58\u5e26\u5bbd\u5229\u7528\u7387 (\u5185\u5b58\u9650\u5236\u573a\u666f) \u548c 580TFLOPS \u7684\u8ba1\u7b97\u80fd\u529b (\u8ba1\u7b97\u9650\u5236\u573a\u666f)\uff0c\u6570\u636e\u8868\u660e FlashMLA \u53ef\u4ee5\u5145\u5206\u5229\u7528 Hopper \u67b6\u6784\u7684 HBM \u9ad8\u5e26\u5bbd\u5185\u5b58\u548c\u5e76\u884c\u8ba1\u7b97\u80fd\u529b\u3002<\/p>\n

FlashMLA \u4f18\u5316\u53d8\u957f\u5e8f\u5217\u5e26\u6765\u7684\u4f18\u52bf\uff1a<\/strong><\/p>\n

\u53d8\u957f\u5e8f\u5217\u662f\u81ea\u7136\u8bed\u8a00\u5904\u7406\u3001\u8bed\u97f3\u8bc6\u522b\u3001\u65f6\u95f4\u5e8f\u5217\u5206\u6790\u7b49\u9886\u57df\u9762\u4e34\u7684\u5e38\u89c1\u6280\u672f\u6311\u6218\uff0c\u4f20\u7edf\u6a21\u578b\u5728\u5904\u7406\u4e0d\u56fa\u5b9a\u957f\u5ea6\u7684\u8f93\u5165\u65f6\u5f80\u5f80\u6548\u7387\u4f1a\u6bd4\u8f83\u4f4e\uff0cFlashMLA \u901a\u8fc7\u9488\u5bf9\u6027\u4f18\u5316\u53ef\u4ee5\u63d0\u9ad8\u5927\u578b\u6a21\u578b\u5728\u53d8\u957f\u5e8f\u5217\u573a\u666f\u4e0b\u7684\u63a8\u7406\u901f\u5ea6\uff0c\u56e0\u6b64\u9002\u5408\u7528\u4e8e\u9700\u8981\u5b9e\u65f6\u54cd\u5e94\u548c\u9ad8\u541e\u5410\u91cf\u7684\u5e94\u7528\u3002<\/p>\n

\u4e5f\u5c31\u662f\u8bf4\u501f\u52a9\u8fd9\u9879\u4f18\u52bf\u672a\u6765\u5176\u4ed6\u6a21\u578b\u4e5f\u53ef\u4ee5\u4f18\u5316\u54cd\u5e94\u901f\u5ea6\uff0c\u5c24\u5176\u662f\u5b9e\u65f6\u8bed\u97f3\u6a21\u5f0f\u8fd9\u79cd\u5bf9\u54cd\u5e94\u901f\u5ea6\u6709\u8981\u6c42\u7684\u573a\u666f\uff0cAI \u53ef\u4ee5\u66f4\u5feb\u7684\u56de\u7b54\u800c\u4e0d\u662f\u8ba9\u7528\u6237\u957f\u65f6\u95f4\u7b49\u5f85\u3002<\/p>\n

\u76ee\u524d FlashMLA \u5df2\u7ecf\u5728 GitHub \u4e0a\u5b8c\u5168\u5f00\u6e90\uff0c\u5f00\u53d1\u8005\u53ea\u9700\u8981\u4f7f\u7528\u7b80\u5355\u7684 Python \u547d\u4ee4\u5373\u53ef\u5feb\u901f\u90e8\u7f72\uff0cDeepSeek \u4e5f\u63d0\u4f9b\u4e86\u6d4b\u8bd5\u811a\u672c\u7528\u6765\u9a8c\u8bc1\u6027\u80fd\uff1ahttps:\/\/github.com\/deepseek-ai\/FlashMLA<\/a><\/p>\n

\u8be5\u9879\u76ee\u7684\u5f00\u6e90\u7279\u6027\u8fd8\u501f\u9274\u4e86 FlashAttention 2&3 \u4ee5\u53ca CUTLASS \u9879\u76ee\u7684\u6a21\u5757\u5316\u8bbe\u8ba1\uff0c\u6709\u5174\u8da3\u7684\u5f00\u53d1\u8005\u4e5f\u53ef\u4ee5\u7814\u7a76\u4e0a\u6e38\u9879\u76ee\u7684\u5177\u4f53\u7ec6\u8282\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

\u5f00\u6e90\u4eba\u5de5\u667a\u80fd\u6280\u672f\u5f00\u53d1\u5546 DeepSeek \u4e0a\u5468\u5df2\u7ecf\u9884\u544a\u5c06\u5728\u672c\u5468\u9646\u7eed\u63a8\u51fa 5 \u4e2a\u5df2\u7ecf\u5728\u751f\u4ea7\u73af\u5883\u4e2d\u4f7f\u7528\u7684\u6280\u672f\uff0c\u76ee\u524d […]<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3],"tags":[],"class_list":["post-2217","post","type-post","status-publish","format-standard","hentry","category-ai"],"_links":{"self":[{"href":"https:\/\/myya.net\/index.php\/wp-json\/wp\/v2\/posts\/2217","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/myya.net\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/myya.net\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/myya.net\/index.php\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/myya.net\/index.php\/wp-json\/wp\/v2\/comments?post=2217"}],"version-history":[{"count":0,"href":"https:\/\/myya.net\/index.php\/wp-json\/wp\/v2\/posts\/2217\/revisions"}],"wp:attachment":[{"href":"https:\/\/myya.net\/index.php\/wp-json\/wp\/v2\/media?parent=2217"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/myya.net\/index.php\/wp-json\/wp\/v2\/categories?post=2217"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/myya.net\/index.php\/wp-json\/wp\/v2\/tags?post=2217"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}