{"id":26726,"date":"2025-02-24T12:42:44","date_gmt":"2025-02-24T04:42:44","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=26726"},"modified":"2025-02-24T12:42:44","modified_gmt":"2025-02-24T04:42:44","slug":"flashmla","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/ja\/flashmla\/","title":{"rendered":"FlashMLA\uff1a\u4f18\u5316Hopper GPU\u7684MLA\u89e3\u7801\u5185\u6838\uff08DeepSeek \u5f00\u6e90\u5468\u7b2c\u4e00\u5929\uff09"},"content":{"rendered":"<p>FlashMLA \u662f\u7531 <a href=\"https:\/\/www.kdjingpai.com\/ja\/deepseek-chatshena\/\">DeepSeek<\/a> AI \u5f00\u53d1\u7684\u4e00\u6b3e\u9ad8\u6548 MLA\uff08Multi-head Latent Attention\uff09\u89e3\u7801\u5185\u6838\uff0c\u4e13\u4e3a NVIDIA Hopper \u67b6\u6784 GPU \u4f18\u5316\uff0c\u65e8\u5728\u63d0\u5347\u53d8\u957f\u5e8f\u5217\u5904\u7406\u7684\u6027\u80fd\u3002\u8be5\u9879\u76ee\u5df2\u5728 GitHub \u4e0a\u5f00\u6e90\uff0c\u63d0\u4f9b\u7ed9\u5f00\u53d1\u8005\u514d\u8d39\u4f7f\u7528\u3002\u5b83\u652f\u6301 BF16 \u7cbe\u5ea6\u8ba1\u7b97\u548c\u5206\u9875 KV \u7f13\u5b58\uff08\u5757\u5927\u5c0f\u4e3a 64\uff09\uff0c\u5728 H800 SXM5 \u4e0a\u8868\u73b0\u51fa\u8272\uff0c\u5185\u5b58\u5bc6\u96c6\u578b\u914d\u7f6e\u4e0b\u53ef\u8fbe 3000 GB\/s \u5e26\u5bbd\uff0c\u8ba1\u7b97\u5bc6\u96c6\u578b\u914d\u7f6e\u4e0b\u53ef\u8fbe 580 TFLOPS \u7684\u7b97\u529b\u3002FlashMLA \u7684\u8bbe\u8ba1\u7075\u611f\u6765\u6e90\u4e8e FlashAttention 2&amp;3 \u548c Cutlass \u9879\u76ee\uff0c\u9002\u7528\u4e8e\u751f\u4ea7\u73af\u5883\u5f00\u7bb1\u5373\u7528\u7684\u573a\u666f\u3002DeepSeek AI \u901a\u8fc7\u8fd9\u4e00\u5f00\u6e90\u9879\u76ee\u5c55\u793a\u4e86\u5176\u5728 AI \u6280\u672f\u9886\u57df\u7684\u521b\u65b0\u80fd\u529b\uff0c\u5438\u5f15\u4e86\u5e7f\u6cdb\u5173\u6ce8\u3002<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-26727\" title=\"FlashMLA\uff1a\u4f18\u5316Hopper GPU\u7684MLA\u89e3\u7801\u5185\u6838\uff08DeepSeek \u5f00\u6e90\u5468\u7b2c\u4e00\u5929\uff09-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/875bee4db46b8f4.png\" alt=\"FlashMLA\uff1a\u4f18\u5316Hopper GPU\u7684MLA\u89e3\u7801\u5185\u6838\uff08DeepSeek \u5f00\u6e90\u5468\u7b2c\u4e00\u5929\uff09-1\" width=\"595\" height=\"585\" \/><\/p>\n<p>&nbsp;<\/p>\n<h2>\u529f\u80fd\u5217\u8868<\/h2>\n<ul>\n<li><strong>\u9ad8\u6548 MLA \u89e3\u7801<\/strong>\uff1a\u9488\u5bf9 Hopper GPU \u4f18\u5316\uff0c\u663e\u8457\u63d0\u5347\u53d8\u957f\u5e8f\u5217\u7684\u5904\u7406\u901f\u5ea6\u3002<\/li>\n<li><strong>\u652f\u6301 BF16 \u7cbe\u5ea6<\/strong>\uff1a\u5229\u7528\u534a\u7cbe\u5ea6\u6d6e\u70b9\u8fd0\u7b97\uff0c\u5728\u4fdd\u6301\u7cbe\u5ea6\u7684\u540c\u65f6\u63d0\u5347\u8ba1\u7b97\u6548\u7387\u3002<\/li>\n<li><strong>\u5206\u9875 KV \u7f13\u5b58<\/strong>\uff1a\u91c7\u7528\u5757\u5927\u5c0f\u4e3a 64 \u7684\u5206\u9875\u673a\u5236\uff0c\u6709\u6548\u7ba1\u7406\u5185\u5b58\uff0c\u63d0\u5347\u63a8\u7406\u6027\u80fd\u3002<\/li>\n<li><strong>\u9ad8\u6027\u80fd\u8868\u73b0<\/strong>\uff1a\u5728 H800 GPU \u4e0a\u63d0\u4f9b\u9ad8\u8fbe 3000 GB\/s \u7684\u5185\u5b58\u5e26\u5bbd\u548c 580 TFLOPS \u7684\u8ba1\u7b97\u80fd\u529b\u3002<\/li>\n<li><strong>\u5f00\u6e90\u4ee3\u7801<\/strong>\uff1a\u63d0\u4f9b\u5b8c\u6574\u6e90\u7801\uff0c\u652f\u6301\u5f00\u53d1\u8005\u81ea\u5b9a\u4e49\u4fee\u6539\u548c\u96c6\u6210\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h2>\u4f7f\u7528\u5e2e\u52a9<\/h2>\n<h3>\u5b89\u88c5\u6d41\u7a0b<\/h3>\n<p>FlashMLA \u662f\u4e00\u4e2a\u57fa\u4e8e GitHub \u7684\u5f00\u6e90\u9879\u76ee\uff0c\u4f7f\u7528\u524d\u9700\u786e\u4fdd\u73af\u5883\u6ee1\u8db3\u8981\u6c42\u5e76\u5b8c\u6210\u5b89\u88c5\u3002\u4ee5\u4e0b\u662f\u8be6\u7ec6\u6b65\u9aa4\uff1a<\/p>\n<h4>1. \u73af\u5883\u6e96\u5099<\/h4>\n<ul>\n<li><strong>\u64cd\u4f5c\u7cfb\u7edf<\/strong>\uff1a\u652f\u6301 Linux \u7cfb\u7edf\uff08\u63a8\u8350 Ubuntu 20.04 \u6216\u4ee5\u4e0a\uff09\u3002<\/li>\n<li><strong>\u786c\u4ef6\u8981\u6c42<\/strong>\uff1a\u9700\u8981 NVIDIA Hopper \u67b6\u6784 GPU\uff08\u5982 H800 SXM5\uff09\u3002<\/li>\n<li><strong>\u8f6f\u4ef6\u4f9d\u8d56<\/strong>\uff1a\n<ul>\n<li>CUDA 12.6 \u6216\u4ee5\u4e0a\u7248\u672c\uff08\u5b89\u88c5\u65b9\u6cd5\u53ef\u53c2\u8003 NVIDIA \u5b98\u7f51\uff09\u3002<\/li>\n<li>PyTorch 2.0 \u6216\u4ee5\u4e0a\u7248\u672c\uff08\u63a8\u8350\u901a\u8fc7\u00a0<code>pip install torch<\/code>\u00a0\u5b89\u88c5\uff09\u3002<\/li>\n<li>Python 3.8 \u6216\u4ee5\u4e0a\u7248\u672c\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u68c0\u67e5\u5de5\u5177<\/strong>\uff1a\u786e\u4fdd\u5b89\u88c5 Git\uff0c\u7528\u4e8e\u4ece GitHub \u4e0b\u8f7d\u4ee3\u7801\u3002<\/li>\n<\/ul>\n<h4>2. \u4e0b\u8f7d\u6e90\u7801<\/h4>\n<ol>\n<li>\u6253\u5f00\u7ec8\u7aef\uff0c\u8f93\u5165\u4ee5\u4e0b\u547d\u4ee4\u514b\u9686 FlashMLA \u4ed3\u5e93\uff1a\n<pre><code>git clone https:\/\/github.com\/deepseek-ai\/FlashMLA.git<\/code><\/pre>\n<\/li>\n<\/ol>\n<ol start=\"2\">\n<li>\u8fdb\u5165\u9879\u76ee\u76ee\u5f55\uff1a\n<pre><code>cd FlashMLA\r\n<\/code><\/pre>\n<\/li>\n<\/ol>\n<h4>3. \u5b89\u88c5\u4f9d\u8d56<\/h4>\n<p>\u9879\u76ee\u4f9d\u8d56 PyTorch \u548c CUDA\uff0c\u53ef\u901a\u8fc7\u4ee5\u4e0b\u547d\u4ee4\u5b89\u88c5\uff1a<\/p>\n<pre><code>pip install -r requirements.txt\r\n<\/code><\/pre>\n<p>\u5982\u679c\u6ca1\u6709\u00a0<code>requirements.txt<\/code>\u00a0\u6587\u4ef6\uff0c\u53ef\u76f4\u63a5\u786e\u4fdd PyTorch \u5df2\u5b89\u88c5\uff1a<\/p>\n<pre><code>pip install torch torchvision\r\n<\/code><\/pre>\n<p>\u9a8c\u8bc1 CUDA \u662f\u5426\u53ef\u7528\uff1a<\/p>\n<pre><code>python -c \"import torch; print(torch.cuda.is_available())\"\r\n<\/code><\/pre>\n<p>\u8f93\u51fa\u00a0<code>True<\/code>\u00a0\u8868\u793a\u73af\u5883\u914d\u7f6e\u6210\u529f\u3002<\/p>\n<h4>4. \u7f16\u8bd1\u4e0e\u6d4b\u8bd5<\/h4>\n<p>FlashMLA \u63d0\u4f9b\u9884\u7f16\u8bd1\u7684 CUDA \u63d2\u4ef6\uff0c\u4f46\u9700\u786e\u4fdd\u4e0e\u672c\u5730 CUDA \u7248\u672c\u5339\u914d\uff1a<\/p>\n<ol>\n<li>\u8fdb\u5165\u6e90\u7801\u76ee\u5f55\uff0c\u8fd0\u884c\u7f16\u8bd1\u811a\u672c\uff08\u82e5\u6709\uff09\uff1a\n<pre><code>python setup.py install\r\n<\/code><\/pre>\n<\/li>\n<li>\u6d4b\u8bd5\u5b89\u88c5\u662f\u5426\u6210\u529f\uff0c\u8fd0\u884c\u793a\u4f8b\u4ee3\u7801\uff1a\n<pre><code>python example.py\r\n<\/code><\/pre>\n<\/li>\n<\/ol>\n<p>\u82e5\u65e0\u62a5\u9519\uff0c\u8868\u793a\u5b89\u88c5\u5b8c\u6210\u3002<\/p>\n<h3>\u5982\u4f55\u4f7f\u7528<\/h3>\n<p>FlashMLA \u7684\u6838\u5fc3\u529f\u80fd\u662f\u63d0\u4f9b\u9ad8\u6548\u7684 MLA \u89e3\u7801\u652f\u6301\uff0c\u9002\u7528\u4e8e AI \u6a21\u578b\u63a8\u7406\u4efb\u52a1\u3002\u4ee5\u4e0b\u662f\u5177\u4f53\u64cd\u4f5c\u6b65\u9aa4\uff1a<\/p>\n<h4>\u529f\u80fd 1\uff1a\u52a0\u8f7d\u5e76\u8fd0\u884c FlashMLA<\/h4>\n<ol>\n<li><strong>\u5bfc\u5165\u6a21\u5757<\/strong>\uff1a<br \/>\n\u5728 Python \u811a\u672c\u4e2d\u5f15\u5165 FlashMLA \u6838\u5fc3\u51fd\u6570\uff1a<\/p>\n<pre><code>from flash_mla import get_mla_metadata, flash_mla_with_kvcache\r\n<\/code><\/pre>\n<\/li>\n<li><strong>\u51c6\u5907\u8f93\u5165\u6570\u636e<\/strong>\uff1a\n<ul>\n<li><code>cache_seqlens<\/code>\uff1a\u5b9a\u4e49 KV \u7f13\u5b58\u7684\u5e8f\u5217\u957f\u5ea6\u3002<\/li>\n<li><code>q_i<\/code>\uff1a\u67e5\u8be2\u5f20\u91cf\u3002<\/li>\n<li><code>kvcache_i<\/code>\uff1aKV \u7f13\u5b58\u6570\u636e\u3002<\/li>\n<li><code>block_table<\/code>\uff1a\u5206\u9875\u7f13\u5b58\u7684\u5757\u8868\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u83b7\u53d6\u5143\u6570\u636e<\/strong>\uff1a\n<pre><code>tile_scheduler_metadata, num_splits = get_mla_metadata(cache_seqlens, s_q * h_q \/\/ h_kv, h_kv)\r\n<\/code><\/pre>\n<\/li>\n<li><strong>\u8fd0\u884c\u89e3\u7801<\/strong>\uff1a\n<pre><code>o_i, lse_i = flash_mla_with_kvcache(q_i, kvcache_i, block_table, cache_seqlens, dv, tile_scheduler_metadata, num_splits, causal=True)\r\n<\/code><\/pre>\n<p>\u8f93\u51fa\u00a0<code>o_i<\/code>\u00a0\u4e3a\u89e3\u7801\u7ed3\u679c\uff0c<code>lse_i<\/code>\u00a0\u4e3a\u65e5\u5fd7\u548c\u503c\u3002<\/li>\n<\/ol>\n<h4>\u529f\u80fd 2\uff1a\u4f18\u5316\u53d8\u957f\u5e8f\u5217\u5904\u7406<\/h4>\n<ul>\n<li><strong>\u573a\u666f<\/strong>\uff1a\u5904\u7406\u52a8\u6001\u957f\u5ea6\u7684\u8f93\u5165\u5e8f\u5217\u65f6\uff0cFlashMLA \u901a\u8fc7\u5206\u9875 KV \u7f13\u5b58\u51cf\u5c11\u5185\u5b58\u5360\u7528\u3002<\/li>\n<li><strong>\u64cd\u4f5c<\/strong>\uff1a\n<ol>\n<li>\u914d\u7f6e\u5206\u9875\u53c2\u6570\uff1a\u5757\u5927\u5c0f\u56fa\u5b9a\u4e3a 64\uff0c\u53ef\u901a\u8fc7\u8c03\u6574\u00a0<code>cache_seqlens<\/code>\u00a0\u63a7\u5236\u5e8f\u5217\u957f\u5ea6\u3002<\/li>\n<li>\u8fd0\u884c\u65f6\u6307\u5b9a\u00a0<code>causal=True<\/code>\uff0c\u786e\u4fdd\u56e0\u679c\u6ce8\u610f\u529b\u673a\u5236\u751f\u6548\u3002<\/li>\n<\/ol>\n<\/li>\n<li><strong>\u6548\u679c<\/strong>\uff1a\u5728 H800 \u4e0a\u53ef\u5b9e\u73b0 3000 GB\/s \u7684\u5185\u5b58\u5e26\u5bbd\uff0c\u9002\u5408\u5927\u89c4\u6a21\u63a8\u7406\u4efb\u52a1\u3002<\/li>\n<\/ul>\n<h4>\u529f\u80fd 3\uff1a\u6027\u80fd\u6d4b\u8bd5<\/h4>\n<ul>\n<li><strong>\u6d4b\u8bd5\u65b9\u6cd5<\/strong>\uff1a\n<ol>\n<li>\u7f16\u8f91\u793a\u4f8b\u811a\u672c\uff08\u5982\u00a0<code>example.py<\/code>\uff09\uff0c\u589e\u52a0\u8f93\u5165\u6570\u636e\u89c4\u6a21\u3002<\/li>\n<li>\u4f7f\u7528\u4ee5\u4e0b\u4ee3\u7801\u8bb0\u5f55\u6027\u80fd\uff1a\n<pre><code>import time\r\nstart = time.time()\r\n# \u8fd0\u884c\u89e3\u7801\u4ee3\u7801\r\no_i, lse_i = flash_mla_with_kvcache(...)\r\nprint(f\"\u8017\u65f6: {time.time() - start} \u79d2\")\r\n<\/code><\/pre>\n<\/li>\n<\/ol>\n<\/li>\n<li><strong>\u9884\u671f\u7ed3\u679c<\/strong>\uff1a\u5185\u5b58\u5bc6\u96c6\u578b\u4efb\u52a1\u63a5\u8fd1 3000 GB\/s\uff0c\u8ba1\u7b97\u5bc6\u96c6\u578b\u4efb\u52a1\u63a5\u8fd1 580 TFLOPS\u3002<\/li>\n<\/ul>\n<h3>\u6ce8\u610f\u4e8b\u9879<\/h3>\n<ul>\n<li><strong>\u786c\u4ef6\u517c\u5bb9\u6027<\/strong>\uff1a\u4ec5\u652f\u6301 Hopper GPU\uff0c\u5efa\u8bae\u4f7f\u7528 H800 \u6216\u540c\u7ea7\u522b\u8bbe\u5907\u3002<\/li>\n<li><strong>\u8c03\u8bd5\u6280\u5de7<\/strong>\uff1a\u82e5\u9047\u5230 CUDA \u9519\u8bef\uff0c\u68c0\u67e5\u7248\u672c\u662f\u5426\u5339\u914d\uff0c\u6216\u5728 GitHub Issues \u4e2d\u5bfb\u6c42\u793e\u533a\u652f\u6301\u3002<\/li>\n<li><strong>\u751f\u4ea7\u73af\u5883<\/strong>\uff1a\u76f4\u63a5\u96c6\u6210\u5230\u73b0\u6709\u6a21\u578b\u63a8\u7406\u6d41\u7a0b\u4e2d\uff0c\u786e\u4fdd\u8f93\u5165\u6570\u636e\u683c\u5f0f\u4e0e FlashMLA \u8981\u6c42\u4e00\u81f4\u3002<\/li>\n<\/ul>\n<p>\u901a\u8fc7\u4ee5\u4e0a\u6b65\u9aa4\uff0c\u7528\u6237\u53ef\u5feb\u901f\u4e0a\u624b FlashMLA\uff0c\u4eab\u53d7\u5176\u9ad8\u6548\u89e3\u7801\u5e26\u6765\u7684\u6027\u80fd\u63d0\u5347\u3002\u5b8c\u6574\u4ee3\u7801\u548c\u6587\u6863\u53ef\u5728 GitHub \u4ed3\u5e93\u67e5\u770b\uff0c\u5efa\u8bae\u7ed3\u5408\u5b9e\u9645\u9879\u76ee\u9700\u6c42\u8c03\u6574\u53c2\u6570\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>FlashMLA \u662f\u7531 DeepSeek AI \u5f00\u53d1\u7684\u4e00\u6b3e\u9ad8\u6548 MLA\uff08Multi-head Latent Attention\uff09\u89e3\u7801\u5185\u6838\uff0c\u4e13\u4e3a NVIDIA Hopper \u67b6\u6784 GPU \u4f18\u5316\uff0c\u65e8\u5728\u63d0\u5347\u53d8\u957f\u5e8f\u5217\u5904\u7406\u7684\u6027\u80fd\u3002\u8be5\u9879\u76ee\u5df2\u5728 GitH&#8230;<\/p>\n","protected":false},"author":1,"featured_media":61910,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20],"tags":[230],"class_list":["post-26726","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tool","tag-aikaiyuanxiangmu"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts\/26726","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/comments?post=26726"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts\/26726\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/media\/61910"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/media?parent=26726"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/categories?post=26726"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/tags?post=26726"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}