{"id":34443,"date":"2024-12-06T17:56:22","date_gmt":"2024-12-06T09:56:22","guid":{"rendered":"https:\/\/17aitech.com\/?p=34443"},"modified":"2024-12-06T17:56:22","modified_gmt":"2024-12-06T09:56:22","slug":"%e5%8d%95%e5%8d%a1a100%e5%ae%9e%e7%8e%b0%e7%99%be%e4%b8%87token%e6%8e%a8%e7%90%86%ef%bc%8c%e9%80%9f%e5%ba%a6%e5%bf%ab10%e5%80%8d%ef%bc%8c%e8%bf%99%e6%98%af%e5%be%ae%e8%bd%af%e5%ae%98%e6%96%b9%e7%9a%84","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=34443","title":{"rendered":"\u5355\u5361A100\u5b9e\u73b0\u767e\u4e07token\u63a8\u7406\uff0c\u901f\u5ea6\u5feb10\u500d\uff0c\u8fd9\u662f\u5fae\u8f6f\u5b98\u65b9\u7684\u5927\u6a21\u578b\u63a8\u7406\u52a0\u901f"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-07-08-17\" target=\"_blank\">\u5355\u5361A100\u5b9e\u73b0\u767e\u4e07token\u63a8\u7406\uff0c\u901f\u5ea6\u5feb10\u500d\uff0c\u8fd9\u662f\u5fae\u8f6f\u5b98\u65b9\u7684\u5927\u6a21\u578b\u63a8\u7406\u52a0\u901f<\/a><\/p>\n<blockquote data-author-name=\"\" data-content-utf8-length=\"39\" data-source-title=\"\" data-type=\"2\" data-url=\"\">\n<section>\n<p>\u5fae\u8f6f\u7684\u8fd9\u9879\u7814\u7a76\u8ba9\u5f00\u53d1\u8005\u53ef\u4ee5\u5728\u5355\u5361\u673a\u5668\u4e0a\u4ee5 10 \u500d\u7684\u901f\u5ea6\u5904\u7406\u8d85\u8fc7 1M \u7684\u8f93\u5165\u6587\u672c\u3002<\/p>\n<\/section>\n<\/blockquote>\n<section><\/section>\n<section>\u5927\u578b<mark data-type=\"tech_tasks\" data-id=\"fee178e8-ee20-42fc-8f8f-1d41c1f34e8f\">\u8bed\u8a00\u6a21\u578b<\/mark> (LLM) \u5df2\u8fdb\u5165\u957f\u4e0a\u4e0b\u6587\u5904\u7406\u65f6\u4ee3\uff0c\u5176\u652f\u6301\u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\u4ece\u5148\u524d\u7684 128K \u731b\u589e\u5230 10M token \u7ea7\u522b\u3002<\/section>\n<section><\/section>\n<section>\u7136\u800c\uff0c\u7531\u4e8e<mark data-type=\"tech_methods\" data-id=\"60bee267-89dc-4606-ba24-6b3f7d1f6189\">\u6ce8\u610f\u529b\u673a\u5236<\/mark>\u7684\u4e8c\u6b21\u590d\u6742\u5ea6\uff0c\u6a21\u578b\u5904\u7406\u8f93\u5165\u63d0\u793a\uff08\u5373\u9884\u586b\u5145\u9636\u6bb5\uff09\u5e76\u5f00\u59cb\u4ea7\u751f\u7b2c\u4e00\u4e2a token \u53ef\u80fd\u9700\u8981\u51e0\u5206\u949f\u65f6\u95f4\u3002\u5bfc\u81f4\u9996\u4e2a token \u751f\u6210\u7684\u65f6\u95f4\u8fc7\u957f\uff0c\u4ece\u800c\u4e25\u91cd\u5f71\u54cd\u4e86\u7528\u6237\u4f53\u9a8c\uff0c\u8fd9\u4e5f\u6781\u5927\u5730\u9650\u5236\u4e86\u957f\u4e0a\u4e0b\u6587 LLM \u7684\u5e7f\u6cdb\u5e94\u7528\u3002\u00a0<\/section>\n<section><\/section>\n<section>\u4e3e\u4f8b\u6765\u8bf4\uff08\u5982\u56fe 2a \u6240\u793a\uff09\uff0c\u5728\u5355\u53f0\u88c5\u6709 A100 \u7684\u673a\u5668\u4e0a\u4e3a LLaMA-3-8B \u63d0\u4f9b\u670d\u52a1\u65f6\uff0c\u5982\u679c\u63d0\u793a\u6709 30 \u4e07\u4e2a token\uff0c\u6a21\u578b\u9700\u8981 6 \u5206\u949f\u624d\u80fd\u5b8c\u6210\u9884\u586b\u5145\uff08 pre-filling\uff09\u9636\u6bb5\uff0c\u5982\u679c\u63d0\u793a\u589e\u52a0\u5230 100 \u4e07\u4e2a token\uff0c\u8fd9\u4e2a\u6570\u5b57\u5c06\u589e\u52a0\u5230 30 \u5206\u949f\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-6f25f427168d825205c8589b833e2ae8.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-6f25f427168d825205c8589b833e2ae8.png\"><\/a><\/section>\n<section><mark data-type=\"concepts\" data-id=\"2ad7abdc-bbaa-4a33-b3ce-53f4fa019fdd\">\u81ea\u6ce8\u610f\u529b<\/mark>\u8ba1\u7b97\u7684\u5f00\u9500\u5360\u5230\u4e86\u603b\u9884\u586b\u5145\u5ef6\u8fdf\u7684 90% \u4ee5\u4e0a\uff0c\u8fd9\u4f7f\u5176\u6210\u4e3a LLM \u5904\u7406\u957f\u4e0a\u4e0b\u6587\u65f6\u7684\u4e3b\u8981\u74f6\u9888\u3002\u73b0\u6709\u7684\u52a0\u901f\u9884\u586b\u5145\u65b9\u6cd5\u5728\u5e94\u7528\u4e8e\u957f\u4e0a\u4e0b\u6587 LLM \u65f6\u901a\u5e38\u65e0\u6cd5\u4fdd\u6301\u53ef\u63a5\u53d7\u7684\u51c6\u786e\u6027\u6216\u6548\u7387\u3002<\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u89e3\u51b3\u4e0a\u8ff0\u95ee\u9898\uff0c\u6765\u81ea\u5fae\u8f6f\u3001\u8428\u91cc\u5927\u5b66\u7684\u7814\u7a76\u8005\u63d0\u51fa\u4e86\u4e00\u79cd\u65e8\u5728\u52a0\u901f\u957f\u5e8f\u5217\u5904\u7406\u9884\u586b\u5145\u7684\u7a00\u758f\u8ba1\u7b97\u65b9\u6cd5\uff1aMInference\uff08 Milliontokens Inference \uff09\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-c726c458d14ade9cafdbc8b40a516746.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-c726c458d14ade9cafdbc8b40a516746.png\"><\/a><\/section>\n<ul>\n<li>\n<section>\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/pdf\/2407.02490<\/section>\n<\/li>\n<li>\n<section>\u8bba\u6587\u4e3b\u9875\uff1ahttps:\/\/hqjiang.com\/minference.html<\/section>\n<\/li>\n<li>\n<section>\u8bba\u6587\u6807\u9898\uff1aMInference 1.0: Accelerating Pre-filling for Long-Context LLMs via Dynamic Sparse Attention<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>MInference \u53ef\u4ee5\u76f4\u63a5\u5e94\u7528\u4e8e\u73b0\u6709 LLM\uff0c\u65e0\u9700\u5bf9\u9884\u8bad\u7ec3\u8bbe\u7f6e\u8fdb\u884c\u4fee\u6539\u6216\u989d\u5916\u7684\u5fae\u8c03\u3002<\/section>\n<section><\/section>\n<section>\u901a\u8fc7\u5bf9\u5404\u79cd\u4e0b\u6e38\u4efb\u52a1\uff08\u5305\u62ec InfiniteBench\u3001RULER\u3001PG-19 \u548c Needle In A Haystack\uff09\u4ee5\u53ca\u6a21\u578b\uff08\u5305\u62ec LLaMA-3-1M\u3001Yi-200K\u3001GLM-4-1M\u3001Phi-3-128K \u548c Qwen2-128K\uff09\u8fdb\u884c\u8bc4\u4f30\uff0c\u5b9e\u9a8c\u8bc1\u660e MInference \u53ef\u6709\u6548\u5c06 A100 \u4e0a\u7684\u9884\u586b\u5145\u63a8\u7406\u5ef6\u8fdf\u964d\u4f4e\u591a\u8fbe 10 \u500d\uff0c\u540c\u65f6\u4fdd\u6301\u51c6\u786e\u6027\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-a7bdbd403ca8269d9e8139cc5e33c7d7.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-a7bdbd403ca8269d9e8139cc5e33c7d7.png\"><\/a><\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-ef8898d831003ab858f1bc4d79f7eb27.gif\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-ef8898d831003ab858f1bc4d79f7eb27.gif\"><\/a><\/p>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u4f7f\u7528 MInference 1.0 \uff0c\u957f\u4e0a\u4e0b\u6587 LLM\uff08\u5982 LLaMA-3-8B-1M\u3001GLM-4-1M\uff09\u5728\u5355\u4e2a A100 \u4e0a\u7684\u63a8\u7406\u901f\u5ea6\u5b9e\u73b0\u4e86 10 \u500d\u63d0\u5347\uff0c\u5e76\u4e14\u51c6\u786e\u5ea6\u66f4\u9ad8\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>\u65b9\u6cd5\u4ecb\u7ecd<\/strong><\/section>\n<section><\/section>\n<section>\u4f5c\u8005\u63d0\u51fa\u4e86 MInference\uff0c\u8fd9\u4e2a\u540d\u5b57\u53cd\u6620\u4e86\u4ed6\u4eec\u5e0c\u671b\u5728\u4e00\u53f0 A100 \u673a\u5668\u4e0a\u5b9e\u73b0\u767e\u4e07\uff08million\uff09token \u63a8\u7406\u7684\u96c4\u5fc3\u3002<\/section>\n<section><\/section>\n<section>MInference \u662f\u4e00\u79cd\u65e0\u9700\u8bad\u7ec3\u7684\u9ad8\u6548\u65b9\u6cd5\uff0c\u7528\u4e8e\u57fa\u4e8e\u52a8\u6001\u7a00\u758f\u6ce8\u610f\u529b\u7684\u957f\u4e0a\u4e0b\u6587 LLM \u7684\u9884\u586b\u5145\u9636\u6bb5\u3002<\/section>\n<section><\/section>\n<section>\u7814\u7a76\u8005\u8ba4\u4e3a\u6ce8\u610f\u529b\uff0c\u7279\u522b\u662f\u5728\u957f\u4e0a\u4e0b\u6587\u4e2d\uff0c\u662f\u7a00\u758f\u548c\u52a8\u6001\u7684\uff0c\u5373\u5728\u4e0d\u540c\u7684\u8f93\u5165\u4e2d\uff0c\u7a00\u758f\u6a21\u5f0f\u6709\u5f88\u5927\u7684\u4e0d\u540c\u3002\u8fd9\u79cd\u52a8\u6001\u7a00\u758f\u6027\u5448\u73b0\u51fa\u4e09\u79cd\u9002\u7528\u4e8e\u6240\u6709\u8f93\u5165\u7684\u72ec\u7279\u7a7a\u95f4\u805a\u5408\u6a21\u5f0f\uff1aA \u5f62\uff08A-shape\uff09\u3001\u5782\u76f4 &#8211; \u659c\u7ebf\uff08Vertical-Slash\uff09\u548c\u5757\u72b6 &#8211; \u7a00\u758f\uff08Block-Sparse\uff09\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-429f731ba712891728cb10ec7412d246.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-429f731ba712891728cb10ec7412d246.png\"><\/a><\/section>\n<section>MInference \u9996\u5148\u4f7f\u7528\u5185\u6838<mark data-type=\"concepts\" data-id=\"1ccbfcc7-1b56-44d6-b64b-729d855abcb1\">\u611f\u77e5<\/mark>\u7a00\u758f\u6a21\u5f0f\u641c\u7d22\u7b97\u6cd5\u4e3a\u6bcf\u4e2a\u5934\u90e8\u79bb\u7ebf\u786e\u5b9a\u6700\u4f73\u52a8\u6001\u7a00\u758f\u6a21\u5f0f\uff0c\u5982\u7b97\u6cd5 1 \u6240\u793a\u3002\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\uff0c\u5b83\u4f1a\u6839\u636e\u5934\u90e8\u7684\u6a21\u5f0f\u52a8\u6001\u903c\u8fd1\u52a8\u6001\u7a00\u758f\u6307\u6570\uff0c\u5982\u7b97\u6cd5 2\u30013 \u6240\u793a\u3002\u6700\u540e\uff0c\u4f5c\u8005\u4f7f\u7528\u4f18\u5316\u540e\u7684 GPU \u5185\u6838\u6267\u884c\u9ad8\u6548\u7684\u52a8\u6001\u7a00\u758f\u6ce8\u610f\u529b\u8ba1\u7b97\uff0c\u5927\u5927\u51cf\u5c11\u4e86\u957f\u4e0a\u4e0b\u6587 LLM \u7684\u9884\u586b\u5145\u9636\u6bb5\u5ef6\u8fdf\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-c111d08f8f02c6359e8eced4e5bfc500.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-c111d08f8f02c6359e8eced4e5bfc500.png\"><\/a><\/section>\n<section>\u4f8b\u5982\uff0c\u5bf9\u4e8e\u300c\u5782\u76f4 &#8211; \u659c\u7ebf\u300d\u6a21\u5f0f\uff0c\u4f5c\u8005\u9996\u5148\u5229\u7528\u6700\u540e\u4e00\u4e2a Q \u548c K \u4e4b\u95f4\u7684\u6ce8\u610f\u529b\u8ba1\u7b97\u6765\u4f30\u8ba1\u5782\u76f4\u7ebf\u548c\u659c\u7ebf\u7684\u6700\u4f73\u6307\u6570\u3002\u7136\u540e\uff0c\u4ed6\u4eec\u5229\u7528\u52a8\u6001\u7a00\u758f\u7f16\u8bd1\u5668 PIT \u548c Triton \u6784\u5efa\u5782\u76f4 &#8211; \u659c\u7ebf FlashAttention \u5185\u6838\uff0c\u52a0\u901f\u6ce8\u610f\u529b\u8ba1\u7b97\u3002\u5bf9\u4e8e A \u5f62\u3001\u5782\u76f4 &#8211; \u659c\u7ebf\u548c\u5757\u72b6 &#8211; \u7a00\u758f\u6a21\u5f0f\uff0c\u4f5c\u8005\u9996\u5148\u5728\u6ce8\u610f\u529b\u8ba1\u7b97\u4e2d\u4f7f\u7528 Q \u548c K \u7684\u5747\u503c\u6c60\u3002\u5229\u7528\u5747\u503c\u6c60\u548c MatMul \u7684\u4ea4\u6362\u5c5e\u6027\uff0c\u53ef\u4ee5\u4f30\u7b97\u51fa\u5757\u72b6 &#8211; \u7a00\u758f\u6307\u6570\u3002\u7136\u540e\uff0c\u4ed6\u4eec\u4f7f\u7528 Triton \u6784\u5efa\u5757\u7a00\u758f FlashAttention \u5185\u6838\uff0c\u52a0\u901f\u6ce8\u610f\u529b\u8ba1\u7b97\u3002\u6709\u5173\u5185\u6838\u7684\u8be6\u7ec6\u5b9e\u73b0\uff0c\u8bf7\u53c2\u9605\u9644\u5f55 C.4 \u548c\u4ee3\u7801\u3002<\/section>\n<section><\/section>\n<section><strong>\u5728\u957f\u4e0a\u4e0b\u6587<mark data-type=\"concepts\" data-id=\"308c3a45-0fee-4ec6-858e-85b15f440fc0\">\u57fa\u51c6<\/mark>\u4e2d\u7684\u8bc4\u4f30\u7ed3\u679c<br \/><\/strong><\/section>\n<section><\/section>\n<section>\u4f5c\u8005\u5728\u4e00\u7cfb\u5217\u573a\u666f\u4e2d\u6d4b\u8bd5\u4e86 MInference\uff0c\u5305\u62ec QA\u3001\u7f16\u7801\u3001\u57fa\u4e8e\u68c0\u7d22\u7684\u4efb\u52a1\u3001multi-hop QA\u3001\u603b\u7ed3\u548c\u6570\u5b66\u4efb\u52a1\u3002RULER <mark data-type=\"concepts\" data-id=\"308c3a45-0fee-4ec6-858e-85b15f440fc0\">\u57fa\u51c6<\/mark>\u5305\u62ec\u51e0\u4e2a\u590d\u6742\u7684 multi-hop \u6216 multi-needle \u4efb\u52a1\uff0c\u6709\u6548\u5730\u53cd\u6620\u4e86 LLM \u7684\u5b9e\u9645\u4e0a\u4e0b\u6587\u7a97\u53e3\u5927\u5c0f\u3002\u5982\u8868 1 \u6240\u793a\uff0cMInference \u6709\u6548\u5730\u4fdd\u7559\u4e86 LLM \u7684\u5b9e\u9645\u4e0a\u4e0b\u6587\u7a97\u53e3\u5904\u7406\u80fd\u529b\uff0c\u751a\u81f3\u5c06\u5b9e\u9645\u4e0a\u4e0b\u6587\u7a97\u53e3\u5927\u5c0f\u7565\u5fae\u6269\u5c55\u5230 32K\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-756f404d0b879b618aec5cf17af49661.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-756f404d0b879b618aec5cf17af49661.png\"><\/a><\/section>\n<section>\u4f5c\u8005\u8fd8\u4f7f\u7528\u5e73\u5747 token \u957f\u5ea6\u4e3a 214K \u7684 InfiniteBench \u5728\u66f4\u5e7f\u6cdb\u7684\u4efb\u52a1\u4e2d\u6d4b\u8bd5\u4e86 MInference\uff0c\u5982\u8868 2 \u6240\u793a\u3002\u4e0e SoTA \u57fa\u7ebf\u76f8\u6bd4\uff0cMInference \u5728\u6240\u6709\u4efb\u52a1\u4e2d\u90fd\u59cb\u7ec8\u4fdd\u6301\u4e86\u826f\u597d\u7684\u6027\u80fd\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u5728\u66f4\u5177\u6311\u6218\u6027\u7684\u68c0\u7d22\u4efb\u52a1\uff08\u5982 KV \u68c0\u7d22\u4efb\u52a1\uff09\u4e2d\uff0c\u6240\u6709\u57fa\u7ebf\u90fd\u65e0\u6cd5\u505a\u51fa\u51c6\u786e\u9884\u6d4b\uff0c<mark data-type=\"concepts\" data-id=\"8be77eae-12da-4e9e-9a88-b7f5bae98c2e\">\u51c6\u786e\u7387<\/mark>\u4f4e\u4e8e 1.2%\u3002\u4f46\u662f\uff0cMInference \u6210\u529f\u5730\u4fdd\u7559\u4e86\u5904\u7406\u52a8\u6001 KV \u5bf9\u68c0\u7d22\u7684\u80fd\u529b\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-e2e35130439b18d63662cc3d07299aee.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-e2e35130439b18d63662cc3d07299aee.png\"><\/a><\/section>\n<section>\u4e3a\u4e86\u8fdb\u4e00\u6b65\u8bc4\u4f30\u4e0d\u540c\u4e0a\u4e0b\u6587\u957f\u5ea6\u548c\u5173\u952e\u4fe1\u606f\u5728\u63d0\u793a\u4e2d\u4e0d\u540c\u4f4d\u7f6e\u65f6\u7684\u6027\u80fd\uff0c\u4f5c\u8005\u4f7f\u7528\u300c\u5927\u6d77\u635e\u9488\u300d\u4efb\u52a1\u6d4b\u8bd5\u4e86\u5404\u79cd\u6a21\u578b\u548c\u65b9\u6cd5\u3002\u5982\u56fe 1 \u6240\u793a\uff0cMInference \u5728\u4e0d\u540c\u7684\u6a21\u578b\u3001\u4e0a\u4e0b\u6587\u7a97\u53e3\u548c\u63d0\u793a\u4fe1\u606f\u4f4d\u7f6e\u4e0b\u90fd\u8868\u73b0\u826f\u597d\uff0c\u4e0e\u539f\u59cb\u6a21\u578b\u76f8\u6bd4\uff0c\u5176\u6027\u80fd\u4fdd\u6301\u4e0d\u53d8\u751a\u81f3\u7565\u6709\u63d0\u9ad8\u3002\u5728 LLaMA-3-8B \u548c GLM-4-9B-1M \u7684\u60c5\u51b5\u4e0b\uff0cMInference \u5728\u9ad8\u8fbe 1M \u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\u4e2d\u5b9e\u73b0\u4e86\u5b8c\u5168\u7eff\u8272\u7684\u6027\u80fd\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u5373\u4f7f\u5728 70K \u4e0a\u4e0b\u6587\u7a97\u53e3\u4e2d\uff0cStreamingLLM \u548c InfLLM \u5728\u63d0\u793a\u7684\u4e2d\u95f4\u6bb5\u6027\u80fd\u4e5f\u4f1a\u4e0b\u964d\u5230 20% \u4ee5\u4e0b\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-8a4d42b5cada68bd24fa480aec0d956e.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-8a4d42b5cada68bd24fa480aec0d956e.png\"><\/a><\/section>\n<section>\u4f5c\u8005\u8fd8\u4f7f\u7528 PG-19 \u5728<mark data-type=\"tech_tasks\" data-id=\"fee178e8-ee20-42fc-8f8f-1d41c1f34e8f\">\u8bed\u8a00\u6a21\u578b<\/mark>\u4efb\u52a1\u4e2d\u6d4b\u8bd5\u4e86 MInference\uff0c\u5176\u4e2d\u5305\u62ec\u591a\u8fbe 100k \u7684 token\u3002\u5982\u56fe 2 \u6240\u793a\uff0cMInference \u6709\u6548\u5730\u4fdd\u6301\u4e86 LLaMA-3-8B \u548c Yi-9B-200K \u7684\u56f0\u60d1\u5ea6\uff0c\u800c\u6240\u6709\u57fa\u7ebf\u90fd\u51fa\u73b0\u4e86\u4e0d\u540c\u7a0b\u5ea6\u7684\u56f0\u60d1\u5ea6\u4e0b\u964d\u3002\u6b64\u5916\uff0c\u4e0e\u6807\u51c6\u7684 StreamingLLM \u76f8\u6bd4\uff0c\u4f7f\u7528\u81a8\u80c0\u548c\u6b65\u957f\u914d\u7f6e\u7684 StreamingLLM \u66f4\u597d\u5730\u4fdd\u6301\u4e86\u56f0\u60d1\u5ea6\u6027\u80fd\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-417c74942458aa17e6fea78af3e3fcc0.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-417c74942458aa17e6fea78af3e3fcc0.png\"><\/a><\/section>\n<section><strong>\u5ef6\u8fdf\u548c\u5185\u6838\u4e2d\u7684\u7a00\u758f\u6a21\u5f0f\u00a0<\/strong><\/section>\n<section><\/section>\n<section>\u56fe 3 \u5c55\u793a\u4e86\u672c\u6587\u63d0\u51fa\u7684\u4e09\u79cd\u6ce8\u610f\u529b\u6a21\u5f0f\u4ee5\u53ca FlashAttention \u7684\u5fae<mark data-type=\"concepts\" data-id=\"308c3a45-0fee-4ec6-858e-85b15f440fc0\">\u57fa\u51c6<\/mark>\u6d4b\u8bd5\u7ed3\u679c\u3002\u53ef\u4ee5\u770b\u51fa\uff0cVertical-Slash \u662f\u4e09\u79cd\u6a21\u5f0f\u4e2d\u6700\u6162\u7684\uff0c\u4f46\u5728 1M \u4e0a\u4e0b\u6587\u7a97\u53e3\u4e0b\uff0c\u76f8\u6bd4 FlashAttention \u4ecd\u7136\u5b9e\u73b0\u4e86 13 \u500d\u7684\u52a0\u901f\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-9d2fe7cb999edd150c30cdc40d0f9f1b.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-9d2fe7cb999edd150c30cdc40d0f9f1b.png\"><\/a><\/section>\n<section>\u56fe 4 \u5c55\u793a\u4e86 Vertical-Slash \u5934\u90e8\u5185\u6838\u4e2d\u7684\u7a00\u758f\u7d22\u5f15\u3002\u5782\u76f4\u7ebf\u901a\u8fc7 PIT FlashAttention \u4f7f\u7528 1&#215;64 \u5757\u8ba1\u7b97\uff0c\u800c\u659c\u7ebf\u901a\u8fc7\u5757\u7ea7 FlashAttention \u4f7f\u7528 64&#215;64 \u5757\u8ba1\u7b97\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-a1b3ca9c75aac4e031755fb4dc609688.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-a1b3ca9c75aac4e031755fb4dc609688.png\"><\/a><\/section>\n<section><em><sup>\u53c2\u8003\u94fe\u63a5\uff1ahttps:\/\/hqjiang.com\/minference.html<\/sup><\/em><\/section>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-07-08-17\" target=\"_blank\">\u5355\u5361A100\u5b9e\u73b0\u767e\u4e07token\u63a8\u7406\uff0c\u901f\u5ea6\u5feb10\u500d\uff0c\u8fd9\u662f\u5fae\u8f6f\u5b98\u65b9\u7684\u5927\u6a21\u578b\u63a8\u7406\u52a0\u901f<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u5355\u5361A100\u5b9e\u73b0\u767e\u4e07t [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[27],"tags":[68],"class_list":["post-34443","post","type-post","status-publish","format-standard","hentry","category-news","tag-68"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.4 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u5355\u5361A100\u5b9e\u73b0\u767e\u4e07token\u63a8\u7406\uff0c\u901f\u5ea6\u5feb10\u500d\uff0c\u8fd9\u662f\u5fae\u8f6f\u5b98\u65b9\u7684\u5927\u6a21\u578b\u63a8\u7406\u52a0\u901f - \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=34443\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=34443\",\"url\":\"https:\/\/17aitech.com\/?p=34443\",\"name\":\"\u5355\u5361A100\u5b9e\u73b0\u767e\u4e07token\u63a8\u7406\uff0c\u901f\u5ea6\u5feb10\u500d\uff0c\u8fd9\u662f\u5fae\u8f6f\u5b98\u65b9\u7684\u5927\u6a21\u578b\u63a8\u7406\u52a0\u901f - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=34443#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=34443#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-6f25f427168d825205c8589b833e2ae8.png\",\"datePublished\":\"2024-12-06T09:56:22+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=34443#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=34443\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=34443#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-6f25f427168d825205c8589b833e2ae8.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-6f25f427168d825205c8589b833e2ae8.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=34443#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u5355\u5361A100\u5b9e\u73b0\u767e\u4e07token\u63a8\u7406\uff0c\u901f\u5ea6\u5feb10\u500d\uff0c\u8fd9\u662f\u5fae\u8f6f\u5b98\u65b9\u7684\u5927\u6a21\u578b\u63a8\u7406\u52a0\u901f\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/17aitech.com\/#website\",\"url\":\"https:\/\/17aitech.com\/\",\"name\":\"\u4e00\u8d77AI\u6280\u672f\",\"description\":\"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca\",\"alternateName\":\"\u4e00\u8d77AI\u6280\u672f\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/17aitech.com\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\",\"name\":\"AI\u5c0f\u52a9\u624b\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/image\/\",\"url\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"contentUrl\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"caption\":\"AI\u5c0f\u52a9\u624b\"},\"description\":\"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e\",\"url\":\"https:\/\/17aitech.com\/?page_id=33738&user=3\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u5355\u5361A100\u5b9e\u73b0\u767e\u4e07token\u63a8\u7406\uff0c\u901f\u5ea6\u5feb10\u500d\uff0c\u8fd9\u662f\u5fae\u8f6f\u5b98\u65b9\u7684\u5927\u6a21\u578b\u63a8\u7406\u52a0\u901f - \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=34443","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=34443","url":"https:\/\/17aitech.com\/?p=34443","name":"\u5355\u5361A100\u5b9e\u73b0\u767e\u4e07token\u63a8\u7406\uff0c\u901f\u5ea6\u5feb10\u500d\uff0c\u8fd9\u662f\u5fae\u8f6f\u5b98\u65b9\u7684\u5927\u6a21\u578b\u63a8\u7406\u52a0\u901f - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=34443#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=34443#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-6f25f427168d825205c8589b833e2ae8.png","datePublished":"2024-12-06T09:56:22+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=34443#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=34443"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=34443#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-6f25f427168d825205c8589b833e2ae8.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-6f25f427168d825205c8589b833e2ae8.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=34443#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u5355\u5361A100\u5b9e\u73b0\u767e\u4e07token\u63a8\u7406\uff0c\u901f\u5ea6\u5feb10\u500d\uff0c\u8fd9\u662f\u5fae\u8f6f\u5b98\u65b9\u7684\u5927\u6a21\u578b\u63a8\u7406\u52a0\u901f"}]},{"@type":"WebSite","@id":"https:\/\/17aitech.com\/#website","url":"https:\/\/17aitech.com\/","name":"\u4e00\u8d77AI\u6280\u672f","description":"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca","alternateName":"\u4e00\u8d77AI\u6280\u672f","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/17aitech.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d","name":"AI\u5c0f\u52a9\u624b","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/#\/schema\/person\/image\/","url":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","contentUrl":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","caption":"AI\u5c0f\u52a9\u624b"},"description":"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e","url":"https:\/\/17aitech.com\/?page_id=33738&user=3"}]}},"_links":{"self":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/34443","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=34443"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/34443\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=34443"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=34443"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=34443"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}