{"id":1184798,"date":"2026-07-15T19:17:00","date_gmt":"2026-07-16T02:17:00","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/?post_type=msr-blog-post&#038;p=1184798"},"modified":"2026-09-03T03:21:06","modified_gmt":"2026-09-03T10:21:06","slug":"slidesparse%ef%bc%9a%e6%8b%93%e5%b1%95%e7%bb%93%e6%9e%84%e5%8c%96%e7%a8%80%e7%96%8f%e8%be%b9%e7%95%8c","status":"publish","type":"msr-blog-post","link":"https:\/\/www.microsoft.com\/en-us\/research\/articles\/slidesparse%ef%bc%9a%e6%8b%93%e5%b1%95%e7%bb%93%e6%9e%84%e5%8c%96%e7%a8%80%e7%96%8f%e8%be%b9%e7%95%8c\/","title":{"rendered":"SlideSparse\uff1a\u62d3\u5c55\u7ed3\u6784\u5316\u7a00\u758f\u8fb9\u754c"},"content":{"rendered":"\n<p class=\"wp-block-paragraph\">\u7ed3\u6784\u5316\u7a00\u758f\u88ab\u8ba4\u4e3a\u662f\u63d0\u5347\u5927\u6a21\u578b\u63a8\u7406\u6548\u7387\u7684\u91cd\u8981\u65b9\u5411\u4e4b\u4e00\u3002\u4e3a\u4e86\u652f\u6301\u8fd9\u7c7b\u8ba1\u7b97\u6a21\u5f0f\uff0c\u5982\u4eca\u7684GPU&nbsp;\u4e2d\u5f15\u5165\u4e86\u7a00\u758f\u5f20\u91cf\u6838\uff08Sparse Tensor Core\uff09\uff0c\u80fd\u591f\u4e3a&nbsp;2:4&nbsp;\u7ed3\u6784\u5316\u7a00\u758f\u63d0\u4f9b\u6700\u9ad8&nbsp;2&nbsp;\u500d\u7684\u77e9\u9635\u8fd0\u7b97\u541e\u5410\u3002\u5176\u4e2d\uff0c2:4&nbsp;\u7a00\u758f\u8981\u6c42\u6bcf&nbsp;4&nbsp;\u4e2a\u8fde\u7eed\u6743\u91cd\u4e2d\u6070\u6709\u4e24\u4e2a\u4e3a\u96f6\u3002\u5c3d\u7ba1\u8fd9\u4e00\u673a\u5236\u5177\u5907\u663e\u8457\u7684\u52a0\u901f\u6f5c\u529b\uff0c\u4f46&nbsp;50%&nbsp;\u7684\u526a\u679d\u7387\u5bf9\u4e8e\u5927\u8bed\u8a00\u6a21\u578b\u800c\u8a00\u5f80\u5f80\u8fc7\u4e8e\u6fc0\u8fdb\uff0c\u5373\u4f7f\u7ecf\u8fc7\u5fae\u8c03\uff0c\u4ecd\u53ef\u80fd\u5e26\u6765\u660e\u663e\u7684\u7cbe\u5ea6\u635f\u5931\u548c\u80fd\u529b\u9000\u5316\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u800c&nbsp;6:8&nbsp;\u7b49\u6e29\u548c\u7a00\u758f\u6a21\u5f0f\u5219\u80fd\u591f\u5728\u4fdd\u6301\u6a21\u578b\u80fd\u529b\u7684\u540c\u65f6\u5b9e\u73b0\u6709\u6548\u538b\u7f29\u3002\u4f46\u662f\uff0c6:8&nbsp;\u867d\u7136\u4fdd\u6301\u4e86\u6a21\u578b\u80fd\u529b\uff0c\u5374\u65e0\u6cd5\u83b7\u5f97\u786c\u4ef6\u52a0\u901f\u3002\u5f53\u524d\u7a00\u758f\u5f20\u91cf\u6838\u4ec5\u652f\u6301&nbsp;2:4&nbsp;\u8fd9\u4e00\u79cd\u683c\u5f0f\uff0c\u4e0d\u7b26\u5408\u8be5\u683c\u5f0f\u7684\u7a00\u758f\u6743\u91cd\u53ea\u80fd\u56de\u9000\u5230\u7a20\u5bc6\u8ba1\u7b97\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c6:8\u6a21\u578b\u4e2d&nbsp;25%&nbsp;\u7684\u96f6\u5143\u7d20\u5728\u63a8\u7406\u65f6\u5e76\u672a\u5e26\u6765\u4efb\u4f55\u8ba1\u7b97\u6216\u5e26\u5bbd\u4e0a\u7684\u8282\u7701\u2014\u2014\u7b97\u6cd5\u5c42\u9762\u7684\u7a00\u758f\u4f18\u52bf\u65e0\u6cd5\u8f6c\u5316\u4e3a\u90e8\u7f72\u4e0a\u7684\u5b9e\u9645\u6536\u76ca\u3002\u5728\u91cf\u5316\u7cbe\u5ea6\u5df2\u7ecf\u8986\u76d6\u4e86\u4ece&nbsp;16 bit&nbsp;\u52304 bit&nbsp;\u7684\u4e30\u5bcc\u9009\u62e9\u7684\u540c\u65f6\uff0c\u7a00\u758f\u65b9\u5411\u4e0a\uff0c50%&nbsp;\u4e0e&nbsp;0%&nbsp;\u4e4b\u95f4\u5b58\u5728\u4e00\u5927\u7247\u5c1a\u672a\u88ab\u786c\u4ef6\u8986\u76d6\u7684\u7a7a\u767d\u5730\u5e26\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd1\u65e5\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u63d0\u51fa\u4e86\u00a0SlideSparse\uff0c\u9996\u6b21\u5728\u00a0NVIDIA GPU\u00a0\u4e0a\u4f7f\u00a06:8\u30014:6\u30018:10\u00a0\u7b49\u6e29\u548c\u7a00\u758f\u6a21\u5f0f\u4e5f\u80fd\u5229\u7528\u7a00\u758f\u5f20\u91cf\u6838\u52a0\u901f\uff0c\u4e3a\u8fd9\u4e00\u957f\u671f\u5b58\u5728\u7684\u6280\u672f\u7a7a\u767d\u63d0\u4f9b\u4e86\u89e3\u51b3\u65b9\u6848\u3002<strong>\u76f8\u5173\u8bba\u6587\u5df2\u88abICML 2026\u00a0\u63a5\u6536\u3002<\/strong><\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"643\" height=\"268\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/07\/slide-1.png\" alt=\"diagram,chart\" class=\"wp-image-1184862\" style=\"aspect-ratio:2.3993378450752227;width:770px;height:auto\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/07\/slide-1.png 643w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/07\/slide-1-300x125.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/07\/slide-1-240x100.png 240w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/07\/slide-1-640x268.png 640w\" sizes=\"auto, (max-width: 643px) 100vw, 643px\" \/><figcaption class=\"wp-element-caption\">\u56fe1\uff1aSlideSparse\u00a0\u5c06\u00a06:8\u00a0\u7a00\u758f\u6743\u91cd\u8f6c\u6362\u4e3a\u00a02:4\u00a0\u517c\u5bb9\u683c\u5f0f\uff08\u5de6\uff09\uff0cA100\u4e0a\u4e0d\u540c\u6a21\u578b\u7684\u7aef\u5230\u7aef\u52a0\u901f\uff08\u53f3\uff09\u3002<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\"><strong><em>\u8bba\u6587\u4fe1\u606f\u5df2\u6574\u7406\u81f3\u6587\u672b\uff0c\u6b22\u8fce\u70b9\u51fb\u76f8\u5173\u94fe\u63a5\uff0c\u4e86\u89e3\u66f4\u591a\u6280\u672f\u8be6\u60c5\u3002<\/em><\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n\n\n\n<h2 id=\"\u6ed1\u52a8\u7a97\u53e3\u5206\u89e3-\u8ba9\u6e29\u548c\u7a00\u758f\u9002\u914d\u7a00\u758f\u5f20\u91cf\u6838\" class=\"wp-block-heading\">\u6ed1\u52a8\u7a97\u53e3\u5206\u89e3\uff1a\u8ba9\u6e29\u548c\u7a00\u758f\u9002\u914d\u7a00\u758f\u5f20\u91cf\u6838<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">SlideSparse&nbsp;\u7684\u57fa\u672c\u601d\u8def\u662f\u5c06\u4e0d\u6ee1\u8db32:4&nbsp;\u7ea6\u675f\u7684\u7a00\u758f\u5757\uff0c\u5206\u89e3\u4e3a\u82e5\u5e72\u6ee1\u8db3&nbsp;2:4&nbsp;\u7ea6\u675f\u7684\u91cd\u53e0\u5b50\u5757\u3002\u7814\u7a76\u5458\u4eec\u5c06\u8fd9\u4e00\u7b56\u7565\u79f0\u4e3a&#8221;\u8ba1\u7b97\u5957\u5229&#8221;\uff08Computational Arbitrage\uff09\uff0c\u901a\u8fc7\u9002\u5ea6\u7684\u6570\u636e\u81a8\u80c0\u6362\u53d6\u786c\u4ef6\u5c42\u9762\u7684\u52a0\u901f\u80fd\u529b\uff0c\u4ece\u800c\u91ca\u653e\u7a00\u758f\u5f20\u91cf\u6838\u7684\u8ba1\u7b97\u6f5c\u529b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ee5&nbsp;6:8&nbsp;\u7a00\u758f\u4e3a\u4f8b\uff0c\u4e00\u4e2a\u957f\u5ea6\u4e3a&nbsp;8&nbsp;\u7684\u6743\u91cd\u5757\u4e2d\u5305\u542b6&nbsp;\u4e2a\u975e\u96f6\u5143\u7d20\u548c\u4e24\u4e2a\u96f6\u5143\u7d20\u3002\u7531\u4e8e\u4e24\u4e2a\u96f6\u5143\u7d20\u7684\u4f4d\u7f6e\u53ef\u4ee5\u4efb\u610f\u5206\u5e03\uff0c\u56e0\u6b64\u65e0\u6cd5\u76f4\u63a5\u62c6\u5206\u4e3a\u4e24\u4e2a\u7b26\u5408&nbsp;2:4&nbsp;\u7ea6\u675f\u7684\u5b50\u5757\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5bf9\u6b64\uff0cSlideSparse\u91c7\u7528\u6ed1\u52a8\u7a97\u53e3\u7b56\u7565\uff1a\u4f7f\u7528&nbsp;3&nbsp;\u4e2a\u5927\u5c0f\u4e3a&nbsp;4\u3001\u6b65\u957f\u4e3a&nbsp;2&nbsp;\u7684\u7a97\u53e3\u4f9d\u6b21\u6ed1\u8fc7\u8fd98&nbsp;\u4e2a\u5143\u7d20\uff08\u7b2c\u4e00\u4e2a\u7a97\u53e3\u8986\u76d6\u4f4d\u7f6e&nbsp;0-3\uff0c\u7b2c\u4e8c\u4e2a\u8986\u76d6&nbsp;2-5\uff0c\u7b2c\u4e09\u4e2a\u8986\u76d6&nbsp;4-7\uff09\u3002\u76f8\u90bb\u7a97\u53e3\u4e4b\u95f4\u5171\u4eab\u4e24\u4e2a\u4f4d\u7f6e\u3002\u5f53\u67d0\u4e2a\u7a97\u53e3\u5185\u7684\u975e\u96f6\u5143\u7d20\u8d85\u8fc7\u4e24\u4e2a\u65f6\uff0c\u591a\u51fa\u7684\u975e\u96f6\u5143\u7d20\u6070\u597d\u4f4d\u4e8e\u4e0e\u4e0b\u4e00\u4e2a\u7a97\u53e3\u7684\u91cd\u53e0\u533a\u57df\u5185\uff0c\u5219\u53ef\u4ee5\u88ab\u4e0b\u4e00\u4e2a\u7a97\u53e3\u5bb9\u7eb3\u3002\u8fd9\u4e00\u8fc7\u7a0b\u672c\u8d28\u4e0a\u5229\u7528\u4e86\u7ecf\u5178\u7684\u201c\u62bd\u5c49\u539f\u7406\u201d\uff0c\u53ef\u786e\u4fdd\u6240\u6709\u975e\u96f6\u5143\u7d20\u5747\u80fd\u88ab\u5408\u7406\u5206\u914d\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"670\" height=\"357\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/07\/slide-2.jpg\" alt=\"diagram\" class=\"wp-image-1184863\" style=\"aspect-ratio:1.8767613679640491;width:770px;height:auto\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/07\/slide-2.jpg 670w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/07\/slide-2-300x160.jpg 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/07\/slide-2-240x128.jpg 240w\" sizes=\"auto, (max-width: 670px) 100vw, 670px\" \/><figcaption class=\"wp-element-caption\">\u56fe2\uff1aSlideSparse\u00a0\u5bf9\u00a06:8\u00a0\u7a00\u758f\u6a21\u5f0f\u8fdb\u884c\u6ed1\u52a8\u7a97\u53e3\u5206\u89e3\u30023\u00a0\u4e2a\u6b65\u957f\u4e3a\u00a02\u00a0\u7684\u91cd\u53e0\u7a97\u53e3\u8986\u76d6\u5168\u90e8\u00a08\u00a0\u4e2a\u4f4d\u7f6e\u3002\u975e\u96f6\u5143\u7d20\u8d85\u51fa\u5355\u4e2a\u7a97\u53e3\u5bb9\u91cf\u65f6\uff0c\u53ef\u901a\u8fc7\u91cd\u53e0\u533a\u57df\u5206\u914d\u7ed9\u76f8\u90bb\u7a97\u53e3\u3002<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u7ecf\u8fc7\u53d8\u6362\u540e\uff0c\u539f\u672c\u76848&nbsp;\u4e2a\u5143\u7d20\u6269\u5c55\u4e3a&nbsp;12&nbsp;\u4e2a\uff0c\u7ef4\u5ea6\u81a8\u80c01.5&nbsp;\u500d\u3002\u4f46\u4e0e\u6b64\u540c\u65f6\uff0c\u6240\u6709\u751f\u6210\u7684\u5b50\u5757\u5747\u6ee1\u8db3&nbsp;2:4&nbsp;\u7ea6\u675f\uff0c\u53ef\u4ee5\u76f4\u63a5\u7531\u7a00\u758f\u5f20\u91cf\u6838\u6267\u884c\u3002\u7531\u4e8e\u786c\u4ef6\u80fd\u591f\u63d0\u4f9b\u4e24\u500d\u541e\u5410\uff0c\u56e0\u6b64\u6700\u7ec8\u53ef\u83b7\u5f97\u7ea61.33\u500d\u7684\u7406\u8bba\u52a0\u901f\u6bd4\u3002\u8bba\u6587\u8fdb\u4e00\u6b65\u8bc1\u660e\uff0c\u5bf9\u4e8e\u4efb\u610f&nbsp;(2N\u22122):2N&nbsp;\u7a00\u758f\u6a21\u5f0f\uff0c\u91c7\u7528N\u22121&nbsp;\u4e2a\u6ed1\u52a8\u7a97\u53e3\u65e2\u662f\u5145\u5206\u6761\u4ef6\u4e5f\u662f\u5fc5\u8981\u6761\u4ef6\uff0c\u56e0\u6b64\u8fd9\u4e00\u81a8\u80c0\u56e0\u5b50\u5df2\u7ecf\u8fbe\u5230\u7406\u8bba\u6700\u4f18\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u6743\u91cd\u4fa7\u7684\u6ed1\u52a8\u53d8\u6362\u5728\u79bb\u7ebf\u9636\u6bb5\u5b8c\u6210\uff0c\u4e0d\u4f1a\u589e\u52a0\u63a8\u7406\u5f00\u9500\u3002\u8f93\u5165\u4fa7\u6240\u9700\u7684\u5bf9\u5e94\u91cd\u6392\u64cd\u4f5c\uff08Activation Lifting\uff09\uff0c\u672c\u8d28\u4e0a\u53ea\u662f\u7d22\u5f15\u6620\u5c04\u3002\u7814\u7a76\u5458\u4eec\u5c06\u5176\u878d\u5408\u81f3\u63a8\u7406\u8fc7\u7a0b\u4e2d\u5df2\u6709\u7684\u9010token&nbsp;\u91cf\u5316&nbsp;kernel&nbsp;\u5185\uff0c\u5b9e\u73b0\u4e86\u6781\u4f4e\u7684\u989d\u5916\u5f00\u9500\u3002\u800c\u6574\u4e2a\u7cfb\u7edf\u90fd\u96c6\u6210\u4e8evLLM&nbsp;\u63a8\u7406\u6846\u67b6\u4e2d\uff0c\u4ec5\u9700\u5c11\u91cf\u4fee\u6539\u5373\u53ef\u542f\u7528\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n\n\n\n<h2 id=\"\u5c06\u7a00\u758f\u6269\u5c55\u4e3a\u65b0\u7684\u63a8\u7406\u4f18\u5316\u7ef4\u5ea6\" class=\"wp-block-heading\">\u5c06\u7a00\u758f\u6269\u5c55\u4e3a\u65b0\u7684\u63a8\u7406\u4f18\u5316\u7ef4\u5ea6<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fc7\u53bb\u51e0\u5e74\uff0c\u5927\u6a21\u578b\u63a8\u7406\u52a0\u901f\u4e3b\u8981\u6cbf\u7740\u91cf\u5316\u65b9\u5411\u6301\u7eed\u6f14\u8fdb\uff1a\u4ece&nbsp;16 bit\u30018 bit\uff0c\u52304 bit&nbsp;\u751a\u81f3\u66f4\u4f4e\u7cbe\u5ea6\uff0c\u5404\u7c7b\u538b\u7f29\u914d\u7f6e\u5747\u9010\u6b65\u83b7\u5f97\u786c\u4ef6\u652f\u6301\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u7a00\u758f\u4f18\u5316\u7684\u53d1\u5c55\u957f\u671f\u53d7\u5230\u786c\u4ef6\u7ea6\u675f\u3002\u76ee\u524d\u4e3b\u6d41&nbsp;GPU&nbsp;\u4ec5\u652f\u6301&nbsp;2:4&nbsp;\u7a00\u758f\u683c\u5f0f\uff0c\u4f7f\u5f97\u5f00\u53d1\u8005\u4e0d\u5f97\u4e0d\u5728\u201c\u65e0\u7a00\u758f\u201d\u548c\u201c50%\u526a\u679d\u201d\u4e4b\u95f4\u8fdb\u884c\u4e8c\u9009\u4e00\u3002\u4f4625%\u300133%&nbsp;\u7b49\u6e29\u548c\u7a00\u758f\u6bd4\u4f8b\u5f80\u5f80\u80fd\u591f\u66f4\u597d\u5730\u517c\u987e\u6a21\u578b\u6027\u80fd\u4e0e\u538b\u7f29\u7387\uff0c\u56e0\u6b64\u4e00\u76f4\u88ab\u8ba4\u4e3a\u662f\u6781\u5177\u6f5c\u529b\u7684\u8bbe\u8ba1\u7a7a\u95f4\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">SlideSparse\u00a0\u7684\u51fa\u73b0\u6539\u53d8\u4e86\u8fd9\u4e00\u5c40\u9762\u3002\u901a\u8fc7\u5c06\u66f4\u591a\u7a00\u758f\u6a21\u5f0f\u6620\u5c04\u5230\u7a00\u758f\u5f20\u91cf\u6838\u4e0a\u6267\u884c\uff0cSlideSparse\u5c06\u7a00\u758f\u4f18\u5316\u4ece\u8fc7\u53bb\u7684\u4e24\u4e2a\u6781\u7aef\u72b6\u6001\u6269\u5c55\u4e3a\u4e00\u7cfb\u5217\u53ef\u7075\u6d3b\u9009\u62e9\u7684\u914d\u7f6e\uff0c\u4f7f\u7a00\u758f\u538b\u7f29\u771f\u6b63\u6210\u4e3a\u4e0e\u91cf\u5316\u538b\u7f29\u5e76\u5217\u7684\u91cd\u8981\u63a8\u7406\u4f18\u5316\u7ef4\u5ea6\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"908\" height=\"406\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/07\/slide-3.jpg\" alt=\"chart\" class=\"wp-image-1184864\" style=\"width:770px\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/07\/slide-3.jpg 908w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/07\/slide-3-300x134.jpg 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/07\/slide-3-768x343.jpg 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/07\/slide-3-240x107.jpg 240w\" sizes=\"auto, (max-width: 908px) 100vw, 908px\" \/><figcaption class=\"wp-element-caption\">\u56fe3\uff1a\u63a8\u7406\u52a0\u901f\u7684\u4e8c\u7ef4\u7a7a\u95f4\u3002\u6a2a\u8f74\u8868\u793a\u91cf\u5316\u7cbe\u5ea6\uff08\u6700\u9ad88\u00d7\u00a0\u52a0\u901f\uff09\uff0c\u7eb5\u8f74\u8868\u793a\u7a00\u758f\u5ea6\uff08\u6700\u9ad8\u00a02\u00d7\u00a0\u52a0\u901f\uff09\u3002\u7070\u8272\u70b9\u6807\u8bb0\u5df2\u6709\u786c\u4ef6\u652f\u6301\uff0c\u7eff\u8272\u70b9\u4e3aSlideSparse\u00a0\u65b0\u589e\u652f\u6301\u7684\u4e2d\u95f4\u7a00\u758f\u7ea7\u522b\u3002<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u91cf\u5316\u4e0e\u7a00\u758f\u672c\u8d28\u4e0a\u662f\u4e24\u4e2a\u76f8\u4e92\u72ec\u7acb\u7684\u4f18\u5316\u65b9\u5411\uff0c\u53ef\u4ee5\u7ec4\u5408\u4f7f\u7528\u3002\u5728\u672a\u6765\u7684\u5927\u6a21\u578b\u90e8\u7f72\u8fc7\u7a0b\u4e2d\uff0c\u5f00\u53d1\u8005\u4e0d\u4ec5\u53ef\u4ee5\u8c03\u6574\u8ba1\u7b97\u7cbe\u5ea6\uff0c\u8fd8\u80fd\u591f\u6839\u636e\u4efb\u52a1\u9700\u6c42\u7075\u6d3b\u9009\u62e9\u4e0d\u540c\u7a00\u758f\u6bd4\u4f8b\uff0c\u5b9e\u73b0\u66f4\u52a0\u7ec6\u7c92\u5ea6\u7684\u6027\u80fd-\u6548\u7387\u6743\u8861\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n\n\n\n<h2 id=\"\u8de8\u5e73\u53f0\u9a8c\u8bc1-slidesparse-\u7684\u52a0\u901f\u80fd\u529b\" class=\"wp-block-heading\">\u8de8\u5e73\u53f0\u9a8c\u8bc1\u00a0SlideSparse\u00a0\u7684\u52a0\u901f\u80fd\u529b<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u9a8c\u8bc1\u65b9\u6848\u7684\u901a\u7528\u6027\uff0c\u7814\u7a76\u5458\u4eec\u5bf9SlideSparse&nbsp;\u5728&nbsp;6&nbsp;\u6b3eGPU\uff08A100\u3001H100\u3001B200\u3001RTX 4090\u3001RTX 5080\u3001DGX Spark\uff09\u30015&nbsp;\u79cd\u7cbe\u5ea6\uff08FP4\u3001INT8\u3001FP8\u3001BF16\u3001FP16\uff09\u30015&nbsp;\u4e2a\u6a21\u578b\uff08Llama3.2-1B\/3B\u3001Qwen2.5-7B\/14B\u3001BitNet b1.58 (2B)\uff09\u4e0a\u8fdb\u884c\u4e86\u8bc4\u6d4b\uff0c\u8986\u76d6\u6570\u636e\u4e2d\u5fc3\u3001\u6d88\u8d39\u7ea7\u548c\u5d4c\u5165\u5f0f\u4e09\u7c7b\u5e73\u53f0\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"962\" height=\"344\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/07\/slide-4.png\" alt=\"chart, bar chart\" class=\"wp-image-1184865\" style=\"width:770px\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/07\/slide-4.png 962w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/07\/slide-4-300x107.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/07\/slide-4-768x275.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/07\/slide-4-240x86.png 240w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/07\/slide-4-960x344.png 960w\" sizes=\"auto, (max-width: 962px) 100vw, 962px\" \/><figcaption class=\"wp-element-caption\">\u56fe4\uff1aSlideSparse\u00a0\u5728\u4e0d\u540c\u00a0GPU\u5e73\u53f0\u4e0a\u7684\u7aef\u5230\u7aef\u63a8\u7406\u52a0\u901f\u8868\u73b0\u3002\u4e0a\u884c\u4e3a\u00a0Decode\u00a0\u9636\u6bb5\uff0c\u4e0b\u884c\u4e3aPrefill\u00a0\u9636\u6bb5\u3002<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u5728\u8ba1\u7b97\u5bc6\u96c6\u7684&nbsp;Prefill&nbsp;\u9636\u6bb5\uff0c6:8&nbsp;\u7a00\u758f\u80fd\u591f\u5145\u5206\u53d1\u6325\u7a00\u758f\u5f20\u91cf\u6838\u7684\u8ba1\u7b97\u80fd\u529b\u3002\u5728A100&nbsp;\u4e0a\uff0cQwen2.5-7B\u6a21\u578b\u5b9e\u73b0\u4e86&nbsp;1.33\u500d\u7684\u7aef\u5230\u7aef\u52a0\u901f\uff0c\u4e0e\u7406\u8bba\u4e0a\u9650\u9ad8\u5ea6\u4e00\u81f4\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u6d88\u8d39\u7ea7&nbsp;GPU&nbsp;\u4e0a\uff0cSlideSparse&nbsp;\u540c\u6837\u5c55\u73b0\u51fa\u7a33\u5b9a\u6536\u76ca\u3002\u4f8b\u5982\uff0c\u5728RTX 4090&nbsp;\u7684&nbsp;FP8&nbsp;\u914d\u7f6e\u4e0b\uff0c6:8&nbsp;\u7a00\u758f\u53ef\u5b9e\u73b0\u7ea6&nbsp;1.18\uff5e1.19\u500d\u7684\u52a0\u901f\u3002Decode\u9636\u6bb5\u7531\u4e8e\u53d7\u5230\u8bbf\u5b58\u74f6\u9888\u9650\u5236\uff0c\u52a0\u901f\u5e45\u5ea6\u76f8\u5bf9\u8f83\u5c0f\uff0c\u4f46\u4ecd\u53ef\u83b7\u5f97&nbsp;1.07\uff5e1.21\u500d\u7684\u6027\u80fd\u63d0\u5347\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u5b9e\u9645\u670d\u52a1\u573a\u666f\u6d4b\u8bd5\u4e2d\uff0cQwen2.5-14B\u6a21\u578b\u7684\u9996&nbsp;token&nbsp;\u5ef6\u8fdf\u7531390 ms&nbsp;\u964d\u4f4e\u81f3&nbsp;291 ms\uff0c\u8fdb\u4e00\u6b65\u9a8c\u8bc1\u4e86SlideSparse&nbsp;\u7684\u5b9e\u9645\u90e8\u7f72\u4ef7\u503c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7814\u7a76\u5458\u4eec\u8fd8\u5bf9\u7cfb\u7edf\u6548\u7387\u8fdb\u884c\u4e86\u6df1\u5165\u5206\u6790\u3002\u901a\u8fc7\u5bf9\u6bd4\u5b9e\u6d4b\u7ed3\u679c\u4e0e\u7406\u8bba\u9884\u671f\u53d1\u73b0\uff0c\u591a\u6570\u914d\u7f6e\u4e0b\u7684\u6574\u4f53\u6548\u7387\u5747\u8d85\u8fc7&nbsp;100%\uff0c\u8868\u660e\u878d\u5408&nbsp;kernel&nbsp;\u7684\u8bbe\u8ba1\u6709\u6548\u6d88\u9664\u4e86\u6ed1\u52a8\u53d8\u6362\u5e26\u6765\u7684\u989d\u5916\u5f00\u9500\uff0c\u5e76\u5728\u90e8\u5206\u786c\u4ef6\u5e73\u53f0\u4e0a\u8fdb\u4e00\u6b65\u91ca\u653e\u4e86\u6f5c\u5728\u6027\u80fd\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"962\" height=\"355\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/07\/slide-5.png\" alt=\"chart, bar chart\" class=\"wp-image-1184866\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/07\/slide-5.png 962w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/07\/slide-5-300x111.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/07\/slide-5-768x283.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/07\/slide-5-240x89.png 240w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/07\/slide-5-960x355.png 960w\" sizes=\"auto, (max-width: 962px) 100vw, 962px\" \/><figcaption class=\"wp-element-caption\">\u56fe5\uff1a\u7b97\u6cd5\u6548\u7387\u5206\u6790\u3002\u5728\u6570\u636e\u4e2d\u5fc3\u00a0GPU\u00a0\u4e0a\uff0c\u6548\u7387\u5747\u8d85\u8fc7\u00a0100%\uff0c\u8868\u660e\u878d\u5408\u00a0kernel\u00a0\u6709\u6548\u6d88\u9664\u4e86\u989d\u5916\u5f00\u9500\u3002<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n\n\n\n<h2 id=\"\u91cd\u65b0\u91ca\u653e\u6e29\u548c\u7a00\u758f\u7684\u90e8\u7f72\u6f5c\u529b\" class=\"wp-block-heading\">\u91cd\u65b0\u91ca\u653e\u6e29\u548c\u7a00\u758f\u7684\u90e8\u7f72\u6f5c\u529b<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u7a00\u758f\u5f20\u91cf\u6838\u662f\u73b0\u4ee3&nbsp;GPU&nbsp;\u4e2d\u91cd\u8981\u7684\u786c\u4ef6\u52a0\u901f\u80fd\u529b\uff0c\u4f46\u5176\u5e94\u7528\u957f\u671f\u53d7\u9650\u4e8e&nbsp;2:4&nbsp;\u7ed3\u6784\u5316\u7a00\u758f\u8fd9\u4e00\u5355\u4e00\u683c\u5f0f\u3002\u5bf9\u4e8e\u80fd\u591f\u66f4\u597d\u517c\u987e\u6a21\u578b\u6027\u80fd\u4e0e\u538b\u7f29\u7387\u7684\u6e29\u548c\u7a00\u758f\u914d\u7f6e\u800c\u8a00\uff0c\u786c\u4ef6\u652f\u6301\u59cb\u7ec8\u5b58\u5728\u7f3a\u53e3\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">SlideSparse&nbsp;\u901a\u8fc7\u521b\u65b0\u6027\u7684\u6ed1\u52a8\u7a97\u53e3\u5206\u89e3\u673a\u5236\uff0c\u5c06\u7a00\u758f\u5f20\u91cf\u6838\u7684\u9002\u7528\u8303\u56f4\u6269\u5c55\u81f3\u6574\u4e2a(2N\u22122):2N&nbsp;\u7a00\u758f\u5bb6\u65cf\uff0c\u4e3a\u6e29\u548c\u7a00\u758f\u63d0\u4f9b\u4e86\u5207\u5b9e\u53ef\u884c\u7684\u786c\u4ef6\u52a0\u901f\u8def\u5f84\u3002\u8be5\u5de5\u4f5c\u4e0d\u4ec5\u62d3\u5bbd\u4e86\u7ed3\u6784\u5316\u7a00\u758f\u7684\u5e94\u7528\u8fb9\u754c\uff0c\u4e5f\u4f7f\u7a00\u758f\u538b\u7f29\u4e0e\u91cf\u5316\u538b\u7f29\u4e00\u6837\uff0c\u6210\u4e3a\u5927\u6a21\u578b\u63a8\u7406\u4f18\u5316\u4e2d\u53ef\u7075\u6d3b\u8c03\u8282\u7684\u91cd\u8981\u7ef4\u5ea6\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u76f8\u5173\u4ee3\u7801\u73b0\u5df2\u5f00\u6e90\uff0c\u5e76\u96c6\u6210\u4e8e&nbsp;vLLM\uff0c\u6b22\u8fce\u5c1d\u8bd5\u3001\u4ea4\u6d41\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\" id=\"slidesparse-fast-and-flexible-2n-2-2n-structured-sparsity-icml-2026\"><strong>SlideSparse: Fast and Flexible (2N-2):2N Structured Sparsity (ICML 2026)<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bba\u6587\u94fe\u63a5\uff1a<a class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" rel=\"noopener noreferrer\" target=\"_blank\" href=\"https:\/\/arxiv.org\/abs\/2603.05232\">https:\/\/arxiv.org\/abs\/2603.05232<span class=\"sr-only\"> (opens in new tab)<\/span><\/a><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ee3\u7801\u94fe\u63a5\uff1a<a class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" rel=\"noopener noreferrer\" target=\"_blank\" href=\"https:\/\/github.com\/bcacdwk\/vllmbench\">https:\/\/github.com\/bcacdwk\/vllmbench<span class=\"sr-only\"> (opens in new tab)<\/span><\/a><\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u7ed3\u6784\u5316\u7a00\u758f\u88ab\u8ba4\u4e3a\u662f\u63d0\u5347\u5927\u6a21\u578b\u63a8\u7406\u6548\u7387\u7684\u91cd\u8981\u65b9\u5411\u4e4b\u4e00\u3002\u4e3a\u4e86\u652f\u6301\u8fd9\u7c7b\u8ba1\u7b97\u6a21\u5f0f\uff0c\u5982\u4eca\u7684GPU&nbsp;\u4e2d\u5f15\u5165\u4e86\u7a00\u758f\u5f20\u91cf\u6838\uff08Sparse Tensor Core\uff09\uff0c\u80fd\u591f\u4e3a&nbsp;2:4&nbsp;\u7ed3\u6784\u5316\u7a00\u758f\u63d0\u4f9b\u6700\u9ad8&nbsp;2&nbsp;\u500d\u7684\u77e9\u9635\u8fd0\u7b97\u541e\u5410\u3002\u5176\u4e2d\uff0c2:4&nbsp;\u7a00\u758f\u8981\u6c42\u6bcf&nbsp;4&nbsp;\u4e2a\u8fde\u7eed\u6743\u91cd\u4e2d\u6070\u6709\u4e24\u4e2a\u4e3a\u96f6\u3002\u5c3d\u7ba1\u8fd9\u4e00\u673a\u5236\u5177\u5907\u663e\u8457\u7684\u52a0\u901f\u6f5c\u529b\uff0c\u4f46&nbsp;50%&nbsp;\u7684\u526a\u679d\u7387\u5bf9\u4e8e\u5927\u8bed\u8a00\u6a21\u578b\u800c\u8a00\u5f80\u5f80\u8fc7\u4e8e\u6fc0\u8fdb\uff0c\u5373\u4f7f\u7ecf\u8fc7\u5fae\u8c03\uff0c\u4ecd\u53ef\u80fd\u5e26\u6765\u660e\u663e\u7684\u7cbe\u5ea6\u635f\u5931\u548c\u80fd\u529b\u9000\u5316\u3002 \u800c&nbsp;6:8&nbsp;\u7b49\u6e29\u548c\u7a00\u758f\u6a21\u5f0f\u5219\u80fd\u591f\u5728\u4fdd\u6301\u6a21\u578b\u80fd\u529b\u7684\u540c\u65f6\u5b9e\u73b0\u6709\u6548\u538b\u7f29\u3002\u4f46\u662f\uff0c6:8&nbsp;\u867d\u7136\u4fdd\u6301\u4e86\u6a21\u578b\u80fd\u529b\uff0c\u5374\u65e0\u6cd5\u83b7\u5f97\u786c\u4ef6\u52a0\u901f\u3002\u5f53\u524d\u7a00\u758f\u5f20\u91cf\u6838\u4ec5\u652f\u6301&nbsp;2:4&nbsp;\u8fd9\u4e00\u79cd\u683c\u5f0f\uff0c\u4e0d\u7b26\u5408\u8be5\u683c\u5f0f\u7684\u7a00\u758f\u6743\u91cd\u53ea\u80fd\u56de\u9000\u5230\u7a20\u5bc6\u8ba1\u7b97\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c6:8\u6a21\u578b\u4e2d&nbsp;25%&nbsp;\u7684\u96f6\u5143\u7d20\u5728\u63a8\u7406\u65f6\u5e76\u672a\u5e26\u6765\u4efb\u4f55\u8ba1\u7b97\u6216\u5e26\u5bbd\u4e0a\u7684\u8282\u7701\u2014\u2014\u7b97\u6cd5\u5c42\u9762\u7684\u7a00\u758f\u4f18\u52bf\u65e0\u6cd5\u8f6c\u5316\u4e3a\u90e8\u7f72\u4e0a\u7684\u5b9e\u9645\u6536\u76ca\u3002\u5728\u91cf\u5316\u7cbe\u5ea6\u5df2\u7ecf\u8986\u76d6\u4e86\u4ece&nbsp;16 bit&nbsp;\u52304 bit&nbsp;\u7684\u4e30\u5bcc\u9009\u62e9\u7684\u540c\u65f6\uff0c\u7a00\u758f\u65b9\u5411\u4e0a\uff0c50%&nbsp;\u4e0e&nbsp;0%&nbsp;\u4e4b\u95f4\u5b58\u5728\u4e00\u5927\u7247\u5c1a\u672a\u88ab\u786c\u4ef6\u8986\u76d6\u7684\u7a7a\u767d\u5730\u5e26\u3002 \u8fd1\u65e5\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u63d0\u51fa\u4e86\u00a0SlideSparse\uff0c\u9996\u6b21\u5728\u00a0NVIDIA GPU\u00a0\u4e0a\u4f7f\u00a06:8\u30014:6\u30018:10\u00a0\u7b49\u6e29\u548c\u7a00\u758f\u6a21\u5f0f\u4e5f\u80fd\u5229\u7528\u7a00\u758f\u5f20\u91cf\u6838\u52a0\u901f\uff0c\u4e3a\u8fd9\u4e00\u957f\u671f\u5b58\u5728\u7684\u6280\u672f\u7a7a\u767d\u63d0\u4f9b\u4e86\u89e3\u51b3\u65b9\u6848\u3002\u76f8\u5173\u8bba\u6587\u5df2\u88abICML 2026\u00a0\u63a5\u6536\u3002 \u8bba\u6587\u4fe1\u606f\u5df2\u6574\u7406\u81f3\u6587\u672b\uff0c\u6b22\u8fce\u70b9\u51fb\u76f8\u5173\u94fe\u63a5\uff0c\u4e86\u89e3\u66f4\u591a\u6280\u672f\u8be6\u60c5\u3002 SlideSparse&nbsp;\u7684\u57fa\u672c\u601d\u8def\u662f\u5c06\u4e0d\u6ee1\u8db32:4&nbsp;\u7ea6\u675f\u7684\u7a00\u758f\u5757\uff0c\u5206\u89e3\u4e3a\u82e5\u5e72\u6ee1\u8db3&nbsp;2:4&nbsp;\u7ea6\u675f\u7684\u91cd\u53e0\u5b50\u5757\u3002\u7814\u7a76\u5458\u4eec\u5c06\u8fd9\u4e00\u7b56\u7565\u79f0\u4e3a&#8221;\u8ba1\u7b97\u5957\u5229&#8221;\uff08Computational Arbitrage\uff09\uff0c\u901a\u8fc7\u9002\u5ea6\u7684\u6570\u636e\u81a8\u80c0\u6362\u53d6\u786c\u4ef6\u5c42\u9762\u7684\u52a0\u901f\u80fd\u529b\uff0c\u4ece\u800c\u91ca\u653e\u7a00\u758f\u5f20\u91cf\u6838\u7684\u8ba1\u7b97\u6f5c\u529b\u3002 \u4ee5&nbsp;6:8&nbsp;\u7a00\u758f\u4e3a\u4f8b\uff0c\u4e00\u4e2a\u957f\u5ea6\u4e3a&nbsp;8&nbsp;\u7684\u6743\u91cd\u5757\u4e2d\u5305\u542b6&nbsp;\u4e2a\u975e\u96f6\u5143\u7d20\u548c\u4e24\u4e2a\u96f6\u5143\u7d20\u3002\u7531\u4e8e\u4e24\u4e2a\u96f6\u5143\u7d20\u7684\u4f4d\u7f6e\u53ef\u4ee5\u4efb\u610f\u5206\u5e03\uff0c\u56e0\u6b64\u65e0\u6cd5\u76f4\u63a5\u62c6\u5206\u4e3a\u4e24\u4e2a\u7b26\u5408&nbsp;2:4&nbsp;\u7ea6\u675f\u7684\u5b50\u5757\u3002 \u5bf9\u6b64\uff0cSlideSparse\u91c7\u7528\u6ed1\u52a8\u7a97\u53e3\u7b56\u7565\uff1a\u4f7f\u7528&nbsp;3&nbsp;\u4e2a\u5927\u5c0f\u4e3a&nbsp;4\u3001\u6b65\u957f\u4e3a&nbsp;2&nbsp;\u7684\u7a97\u53e3\u4f9d\u6b21\u6ed1\u8fc7\u8fd98&nbsp;\u4e2a\u5143\u7d20\uff08\u7b2c\u4e00\u4e2a\u7a97\u53e3\u8986\u76d6\u4f4d\u7f6e&nbsp;0-3\uff0c\u7b2c\u4e8c\u4e2a\u8986\u76d6&nbsp;2-5\uff0c\u7b2c\u4e09\u4e2a\u8986\u76d6&nbsp;4-7\uff09\u3002\u76f8\u90bb\u7a97\u53e3\u4e4b\u95f4\u5171\u4eab\u4e24\u4e2a\u4f4d\u7f6e\u3002\u5f53\u67d0\u4e2a\u7a97\u53e3\u5185\u7684\u975e\u96f6\u5143\u7d20\u8d85\u8fc7\u4e24\u4e2a\u65f6\uff0c\u591a\u51fa\u7684\u975e\u96f6\u5143\u7d20\u6070\u597d\u4f4d\u4e8e\u4e0e\u4e0b\u4e00\u4e2a\u7a97\u53e3\u7684\u91cd\u53e0\u533a\u57df\u5185\uff0c\u5219\u53ef\u4ee5\u88ab\u4e0b\u4e00\u4e2a\u7a97\u53e3\u5bb9\u7eb3\u3002\u8fd9\u4e00\u8fc7\u7a0b\u672c\u8d28\u4e0a\u5229\u7528\u4e86\u7ecf\u5178\u7684\u201c\u62bd\u5c49\u539f\u7406\u201d\uff0c\u53ef\u786e\u4fdd\u6240\u6709\u975e\u96f6\u5143\u7d20\u5747\u80fd\u88ab\u5408\u7406\u5206\u914d\u3002 \u7ecf\u8fc7\u53d8\u6362\u540e\uff0c\u539f\u672c\u76848&nbsp;\u4e2a\u5143\u7d20\u6269\u5c55\u4e3a&nbsp;12&nbsp;\u4e2a\uff0c\u7ef4\u5ea6\u81a8\u80c01.5&nbsp;\u500d\u3002\u4f46\u4e0e\u6b64\u540c\u65f6\uff0c\u6240\u6709\u751f\u6210\u7684\u5b50\u5757\u5747\u6ee1\u8db3&nbsp;2:4&nbsp;\u7ea6\u675f\uff0c\u53ef\u4ee5\u76f4\u63a5\u7531\u7a00\u758f\u5f20\u91cf\u6838\u6267\u884c\u3002\u7531\u4e8e\u786c\u4ef6\u80fd\u591f\u63d0\u4f9b\u4e24\u500d\u541e\u5410\uff0c\u56e0\u6b64\u6700\u7ec8\u53ef\u83b7\u5f97\u7ea61.33\u500d\u7684\u7406\u8bba\u52a0\u901f\u6bd4\u3002\u8bba\u6587\u8fdb\u4e00\u6b65\u8bc1\u660e\uff0c\u5bf9\u4e8e\u4efb\u610f&nbsp;(2N\u22122):2N&nbsp;\u7a00\u758f\u6a21\u5f0f\uff0c\u91c7\u7528N\u22121&nbsp;\u4e2a\u6ed1\u52a8\u7a97\u53e3\u65e2\u662f\u5145\u5206\u6761\u4ef6\u4e5f\u662f\u5fc5\u8981\u6761\u4ef6\uff0c\u56e0\u6b64\u8fd9\u4e00\u81a8\u80c0\u56e0\u5b50\u5df2\u7ecf\u8fbe\u5230\u7406\u8bba\u6700\u4f18\u3002 \u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u6743\u91cd\u4fa7\u7684\u6ed1\u52a8\u53d8\u6362\u5728\u79bb\u7ebf\u9636\u6bb5\u5b8c\u6210\uff0c\u4e0d\u4f1a\u589e\u52a0\u63a8\u7406\u5f00\u9500\u3002\u8f93\u5165\u4fa7\u6240\u9700\u7684\u5bf9\u5e94\u91cd\u6392\u64cd\u4f5c\uff08Activation Lifting\uff09\uff0c\u672c\u8d28\u4e0a\u53ea\u662f\u7d22\u5f15\u6620\u5c04\u3002\u7814\u7a76\u5458\u4eec\u5c06\u5176\u878d\u5408\u81f3\u63a8\u7406\u8fc7\u7a0b\u4e2d\u5df2\u6709\u7684\u9010token&nbsp;\u91cf\u5316&nbsp;kernel&nbsp;\u5185\uff0c\u5b9e\u73b0\u4e86\u6781\u4f4e\u7684\u989d\u5916\u5f00\u9500\u3002\u800c\u6574\u4e2a\u7cfb\u7edf\u90fd\u96c6\u6210\u4e8evLLM&nbsp;\u63a8\u7406\u6846\u67b6\u4e2d\uff0c\u4ec5\u9700\u5c11\u91cf\u4fee\u6539\u5373\u53ef\u542f\u7528\u3002 \u8fc7\u53bb\u51e0\u5e74\uff0c\u5927\u6a21\u578b\u63a8\u7406\u52a0\u901f\u4e3b\u8981\u6cbf\u7740\u91cf\u5316\u65b9\u5411\u6301\u7eed\u6f14\u8fdb\uff1a\u4ece&nbsp;16 bit\u30018 bit\uff0c\u52304 bit&nbsp;\u751a\u81f3\u66f4\u4f4e\u7cbe\u5ea6\uff0c\u5404\u7c7b\u538b\u7f29\u914d\u7f6e\u5747\u9010\u6b65\u83b7\u5f97\u786c\u4ef6\u652f\u6301\u3002 \u76f8\u6bd4\u4e4b\u4e0b\uff0c\u7a00\u758f\u4f18\u5316\u7684\u53d1\u5c55\u957f\u671f\u53d7\u5230\u786c\u4ef6\u7ea6\u675f\u3002\u76ee\u524d\u4e3b\u6d41&nbsp;GPU&nbsp;\u4ec5\u652f\u6301&nbsp;2:4&nbsp;\u7a00\u758f\u683c\u5f0f\uff0c\u4f7f\u5f97\u5f00\u53d1\u8005\u4e0d\u5f97\u4e0d\u5728\u201c\u65e0\u7a00\u758f\u201d\u548c\u201c50%\u526a\u679d\u201d\u4e4b\u95f4\u8fdb\u884c\u4e8c\u9009\u4e00\u3002\u4f4625%\u300133%&nbsp;\u7b49\u6e29\u548c\u7a00\u758f\u6bd4\u4f8b\u5f80\u5f80\u80fd\u591f\u66f4\u597d\u5730\u517c\u987e\u6a21\u578b\u6027\u80fd\u4e0e\u538b\u7f29\u7387\uff0c\u56e0\u6b64\u4e00\u76f4\u88ab\u8ba4\u4e3a\u662f\u6781\u5177\u6f5c\u529b\u7684\u8bbe\u8ba1\u7a7a\u95f4\u3002 SlideSparse\u00a0\u7684\u51fa\u73b0\u6539\u53d8\u4e86\u8fd9\u4e00\u5c40\u9762\u3002\u901a\u8fc7\u5c06\u66f4\u591a\u7a00\u758f\u6a21\u5f0f\u6620\u5c04\u5230\u7a00\u758f\u5f20\u91cf\u6838\u4e0a\u6267\u884c\uff0cSlideSparse\u5c06\u7a00\u758f\u4f18\u5316\u4ece\u8fc7\u53bb\u7684\u4e24\u4e2a\u6781\u7aef\u72b6\u6001\u6269\u5c55\u4e3a\u4e00\u7cfb\u5217\u53ef\u7075\u6d3b\u9009\u62e9\u7684\u914d\u7f6e\uff0c\u4f7f\u7a00\u758f\u538b\u7f29\u771f\u6b63\u6210\u4e3a\u4e0e\u91cf\u5316\u538b\u7f29\u5e76\u5217\u7684\u91cd\u8981\u63a8\u7406\u4f18\u5316\u7ef4\u5ea6\u3002 \u91cf\u5316\u4e0e\u7a00\u758f\u672c\u8d28\u4e0a\u662f\u4e24\u4e2a\u76f8\u4e92\u72ec\u7acb\u7684\u4f18\u5316\u65b9\u5411\uff0c\u53ef\u4ee5\u7ec4\u5408\u4f7f\u7528\u3002\u5728\u672a\u6765\u7684\u5927\u6a21\u578b\u90e8\u7f72\u8fc7\u7a0b\u4e2d\uff0c\u5f00\u53d1\u8005\u4e0d\u4ec5\u53ef\u4ee5\u8c03\u6574\u8ba1\u7b97\u7cbe\u5ea6\uff0c\u8fd8\u80fd\u591f\u6839\u636e\u4efb\u52a1\u9700\u6c42\u7075\u6d3b\u9009\u62e9\u4e0d\u540c\u7a00\u758f\u6bd4\u4f8b\uff0c\u5b9e\u73b0\u66f4\u52a0\u7ec6\u7c92\u5ea6\u7684\u6027\u80fd-\u6548\u7387\u6743\u8861\u3002 \u4e3a\u9a8c\u8bc1\u65b9\u6848\u7684\u901a\u7528\u6027\uff0c\u7814\u7a76\u5458\u4eec\u5bf9SlideSparse&nbsp;\u5728&nbsp;6&nbsp;\u6b3eGPU\uff08A100\u3001H100\u3001B200\u3001RTX 4090\u3001RTX 5080\u3001DGX Spark\uff09\u30015&nbsp;\u79cd\u7cbe\u5ea6\uff08FP4\u3001INT8\u3001FP8\u3001BF16\u3001FP16\uff09\u30015&nbsp;\u4e2a\u6a21\u578b\uff08Llama3.2-1B\/3B\u3001Qwen2.5-7B\/14B\u3001BitNet b1.58 (2B)\uff09\u4e0a\u8fdb\u884c\u4e86\u8bc4\u6d4b\uff0c\u8986\u76d6\u6570\u636e\u4e2d\u5fc3\u3001\u6d88\u8d39\u7ea7\u548c\u5d4c\u5165\u5f0f\u4e09\u7c7b\u5e73\u53f0\u3002 \u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u5728\u8ba1\u7b97\u5bc6\u96c6\u7684&nbsp;Prefill&nbsp;\u9636\u6bb5\uff0c6:8&nbsp;\u7a00\u758f\u80fd\u591f\u5145\u5206\u53d1\u6325\u7a00\u758f\u5f20\u91cf\u6838\u7684\u8ba1\u7b97\u80fd\u529b\u3002\u5728A100&nbsp;\u4e0a\uff0cQwen2.5-7B\u6a21\u578b\u5b9e\u73b0\u4e86&nbsp;1.33\u500d\u7684\u7aef\u5230\u7aef\u52a0\u901f\uff0c\u4e0e\u7406\u8bba\u4e0a\u9650\u9ad8\u5ea6\u4e00\u81f4\u3002 \u5728\u6d88\u8d39\u7ea7&nbsp;GPU&nbsp;\u4e0a\uff0cSlideSparse&nbsp;\u540c\u6837\u5c55\u73b0\u51fa\u7a33\u5b9a\u6536\u76ca\u3002\u4f8b\u5982\uff0c\u5728RTX 4090&nbsp;\u7684&nbsp;FP8&nbsp;\u914d\u7f6e\u4e0b\uff0c6:8&nbsp;\u7a00\u758f\u53ef\u5b9e\u73b0\u7ea6&nbsp;1.18\uff5e1.19\u500d\u7684\u52a0\u901f\u3002Decode\u9636\u6bb5\u7531\u4e8e\u53d7\u5230\u8bbf\u5b58\u74f6\u9888\u9650\u5236\uff0c\u52a0\u901f\u5e45\u5ea6\u76f8\u5bf9\u8f83\u5c0f\uff0c\u4f46\u4ecd\u53ef\u83b7\u5f97&nbsp;1.07\uff5e1.21\u500d\u7684\u6027\u80fd\u63d0\u5347\u3002 \u5728\u5b9e\u9645\u670d\u52a1\u573a\u666f\u6d4b\u8bd5\u4e2d\uff0cQwen2.5-14B\u6a21\u578b\u7684\u9996&nbsp;token&nbsp;\u5ef6\u8fdf\u7531390 ms&nbsp;\u964d\u4f4e\u81f3&nbsp;291 ms\uff0c\u8fdb\u4e00\u6b65\u9a8c\u8bc1\u4e86SlideSparse&nbsp;\u7684\u5b9e\u9645\u90e8\u7f72\u4ef7\u503c\u3002 \u7814\u7a76\u5458\u4eec\u8fd8\u5bf9\u7cfb\u7edf\u6548\u7387\u8fdb\u884c\u4e86\u6df1\u5165\u5206\u6790\u3002\u901a\u8fc7\u5bf9\u6bd4\u5b9e\u6d4b\u7ed3\u679c\u4e0e\u7406\u8bba\u9884\u671f\u53d1\u73b0\uff0c\u591a\u6570\u914d\u7f6e\u4e0b\u7684\u6574\u4f53\u6548\u7387\u5747\u8d85\u8fc7&nbsp;100%\uff0c\u8868\u660e\u878d\u5408&nbsp;kernel&nbsp;\u7684\u8bbe\u8ba1\u6709\u6548\u6d88\u9664\u4e86\u6ed1\u52a8\u53d8\u6362\u5e26\u6765\u7684\u989d\u5916\u5f00\u9500\uff0c\u5e76\u5728\u90e8\u5206\u786c\u4ef6\u5e73\u53f0\u4e0a\u8fdb\u4e00\u6b65\u91ca\u653e\u4e86\u6f5c\u5728\u6027\u80fd\u3002 \u7a00\u758f\u5f20\u91cf\u6838\u662f\u73b0\u4ee3&nbsp;GPU&nbsp;\u4e2d\u91cd\u8981\u7684\u786c\u4ef6\u52a0\u901f\u80fd\u529b\uff0c\u4f46\u5176\u5e94\u7528\u957f\u671f\u53d7\u9650\u4e8e&nbsp;2:4&nbsp;\u7ed3\u6784\u5316\u7a00\u758f\u8fd9\u4e00\u5355\u4e00\u683c\u5f0f\u3002\u5bf9\u4e8e\u80fd\u591f\u66f4\u597d\u517c\u987e\u6a21\u578b\u6027\u80fd\u4e0e\u538b\u7f29\u7387\u7684\u6e29\u548c\u7a00\u758f\u914d\u7f6e\u800c\u8a00\uff0c\u786c\u4ef6\u652f\u6301\u59cb\u7ec8\u5b58\u5728\u7f3a\u53e3\u3002 SlideSparse&nbsp;\u901a\u8fc7\u521b\u65b0\u6027\u7684\u6ed1\u52a8\u7a97\u53e3\u5206\u89e3\u673a\u5236\uff0c\u5c06\u7a00\u758f\u5f20\u91cf\u6838\u7684\u9002\u7528\u8303\u56f4\u6269\u5c55\u81f3\u6574\u4e2a(2N\u22122):2N&nbsp;\u7a00\u758f\u5bb6\u65cf\uff0c\u4e3a\u6e29\u548c\u7a00\u758f\u63d0\u4f9b\u4e86\u5207\u5b9e\u53ef\u884c\u7684\u786c\u4ef6\u52a0\u901f\u8def\u5f84\u3002\u8be5\u5de5\u4f5c\u4e0d\u4ec5\u62d3\u5bbd\u4e86\u7ed3\u6784\u5316\u7a00\u758f\u7684\u5e94\u7528\u8fb9\u754c\uff0c\u4e5f\u4f7f\u7a00\u758f\u538b\u7f29\u4e0e\u91cf\u5316\u538b\u7f29\u4e00\u6837\uff0c\u6210\u4e3a\u5927\u6a21\u578b\u63a8\u7406\u4f18\u5316\u4e2d\u53ef\u7075\u6d3b\u8c03\u8282\u7684\u91cd\u8981\u7ef4\u5ea6\u3002 \u76f8\u5173\u4ee3\u7801\u73b0\u5df2\u5f00\u6e90\uff0c\u5e76\u96c6\u6210\u4e8e&nbsp;vLLM\uff0c\u6b22\u8fce\u5c1d\u8bd5\u3001\u4ea4\u6d41\u3002 SlideSparse: Fast and Flexible (2N-2):2N Structured Sparsity (ICML 2026) \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2603.05232 (opens in new tab) \u4ee3\u7801\u94fe\u63a5\uff1ahttps:\/\/github.com\/bcacdwk\/vllmbench [&hellip;]<\/p>\n","protected":false},"author":44093,"featured_media":1184860,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"_classifai_error":"","msr-content-parent":1012650,"msr_hide_image_in_river":0,"footnotes":""},"research-area":[13561,13556,13552,13547],"msr-locale":[268881],"msr-post-option":[],"class_list":["post-1184798","msr-blog-post","type-msr-blog-post","status-publish","has-post-thumbnail","hentry","msr-research-area-algorithms","msr-research-area-artificial-intelligence","msr-research-area-hardware-devices","msr-research-area-systems-and-networking","msr-locale-zh_cn"],"msr_assoc_parent":{"id":1012650,"type":"lab"},"_links":{"self":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1184798","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post"}],"about":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-blog-post"}],"author":[{"embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/users\/44093"}],"version-history":[{"count":3,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1184798\/revisions"}],"predecessor-version":[{"id":1184867,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1184798\/revisions\/1184867"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media\/1184860"}],"wp:attachment":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1184798"}],"wp:term":[{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1184798"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1184798"},{"taxonomy":"msr-post-option","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-post-option?post=1184798"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}