{"id":3191,"date":"2026-03-11T16:07:42","date_gmt":"2026-03-11T08:07:42","guid":{"rendered":"https:\/\/www.starverse-ai.com\/guide\/archives\/3191"},"modified":"2026-03-11T16:07:42","modified_gmt":"2026-03-11T08:07:42","slug":"%e6%a8%a1%e5%9e%8b%e5%8e%8b%e7%bc%a9%e6%8e%a8%e7%90%86%e5%8a%a0%e9%80%9f%e5%8f%8cbuff%ef%bc%9a%e6%98%9f%e5%ae%87%e6%99%ba%e7%ae%97%e5%b9%b3%e5%8f%b0%e8%ae%a9%e8%be%b9%e7%bc%98%e5%b0%8f%e6%a8%a1","status":"publish","type":"post","link":"https:\/\/www.starverse-ai.com\/guide\/archives\/3191","title":{"rendered":"\u6a21\u578b\u538b\u7f29+\u63a8\u7406\u52a0\u901f\u53ccbuff\uff1a\u661f\u5b87\u667a\u7b97\u5e73\u53f0\u8ba9\u8fb9\u7f18\u5c0f\u6a21\u578b\u4e5f\u80fd\u8dd1\u5728\u4e91\u7aef\u5927\u663e\u5b58"},"content":{"rendered":"<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/www.starverse-ai.com\/guide\/wp-content\/uploads\/2026\/03\/1773216462_c27f46.png\" alt=\"\u6a21\u578b\u538b\u7f29+\u63a8\u7406\u52a0\u901f\u53ccbuff\uff1a\u661f\u5b87\u667a\u7b97\u5e73\u53f0\u8ba9\u8fb9\u7f18\u5c0f\u6a21\u578b\u4e5f\u80fd\u8dd1\u5728\u4e91\u7aef\u5927\u663e\u5b58\" style=\"display:block; margin:10px auto; max-width:100%; height:auto;\" \/><\/figure>\n<blockquote>\n<p>\u80cc\u666f\uff1aMIT 2024 \u5e74 5 \u6708\u53d1\u5e03\u7684\u300aThe State of TinyML\u300b\u6307\u51fa\uff0c\u84b8\u998f\u3001\u526a\u679d\u3001\u91cf\u5316\u4e09\u4ef6\u5957\u5df2\u80fd\u8ba9 Llama-3-8B \u4f53\u79ef\u7f29\u6c34 75%\uff0c\u4f46\u201c\u6a21\u578b\u4f9d\u7136\u65e0\u6cd5\u5728 16 GB \u8fb9\u7f18\u8bbe\u5907\u8fd0\u884c\uff0c\u4e14\u63a8\u7406\u5ef6\u8fdf\u9ad8\u4e8e 200 ms\u201d\u3002\u6362\u8a00\u4e4b\uff0c\u538b\u7f29\u53ea\u662f\u7b2c\u4e00\u6b65\uff0c\u771f\u6b63\u7684\u74f6\u9888\u662f\u201c\u663e\u5b58\u5899\u201d\u4e0e\u201c\u7b97\u529b\u5899\u201d\u53cc\u91cd\u5939\u51fb\u3002  <\/p>\n<\/blockquote>\n<hr \/>\n<h2>\u6a21\u578b\u538b\u7f29+\u63a8\u7406\u52a0\u901f\u53ccbuff\uff1a\u661f\u5b87\u667a\u7b97\u5e73\u53f0\u8ba9\u8fb9\u7f18\u5c0f\u6a21\u578b\u4e5f\u80fd\u8dd1\u5728\u4e91\u7aef\u5927\u663e\u5b58<\/h2>\n<h3>\u2460 MIT \u6a21\u578b\u538b\u7f29\u6280\u672f\u7efc\u8ff0\uff1a\u84b8\u998f\u3001\u526a\u679d\u3001\u91cf\u5316\u4e09\u4ef6\u5957<\/h3>\n<p>\u77e5\u8bc6\u84b8\u998f\u628a 8B \u8001\u5e08\u7f51\u7edc\u7684\u77e5\u8bc6\u8fc1\u79fb\u5230 2B \u5b66\u751f\u7f51\u7edc\uff1b\u7ed3\u6784\u5316\u526a\u679d\u780d\u6389 40% attention \u5934\uff1bINT8\/INT4 \u91cf\u5316\u518d\u628a\u6743\u91cd\u538b\u6210 1-2 Byte\u3002\u4e09\u6b65\u8d70\u5b8c\uff0c\u7406\u8bba\u4e0a\u6a21\u578b\u53ef\u7626\u8eab 75%\uff0c\u5374\u5e26\u6765\u4e24\u4e2a\u9690\u85cf\u4ee3\u4ef7\uff1a<br \/>\n1. \u6fc0\u6d3b\u503c\u4ecd\u9700 FP16 \u7f13\u5b58\uff0c\u5cf0\u503c\u663e\u5b58\u4e0d\u964d\u53cd\u5347\uff1b<br \/>\n2. \u4f4e\u6bd4\u7279 kernel \u9700\u8981\u4e13\u7528 CUDA \u7b97\u5b50\uff0c\u8fb9\u7f18 NPU \u5f80\u5f80\u4e0d\u652f\u6301\u3002  <\/p>\n<p>\u7ed3\u679c\u2014\u2014\u624b\u673a\u3001\u6811\u8393\u6d3e\u3001Jetson \u4f9d\u65e7\u8dd1\u4e0d\u52a8\uff0c\u5f00\u53d1\u8005\u4e0d\u5f97\u4e0d\u628a\u76ee\u5149\u91cd\u65b0\u6295\u5411\u4e91\u7aef <strong>GPU\u670d\u52a1\u5668\u79df\u7528<\/strong> \u65b9\u6848\u3002<\/p>\n<h3>\u2461 \u73b0\u5b9e\u56f0\u5883\uff1a\u538b\u7f29\u540e\u4ecd\u8d85\u8fb9\u7f18\u8bbe\u5907\u663e\u5b58<\/h3>\n<p>\u4ee5\u538b\u7f29\u7248 Llama-3-8B-Infer \u4e3a\u4f8b\uff0cINT4 \u6743\u91cd 4 GB\uff0c\u4f46 KV-Cache \u5728 8 k \u4e0a\u4e0b\u6587\u4e0b\u9ad8\u8fbe 12 GB\uff0c\u5408\u8ba1 16 GB \u5df2\u8d85 Jetson AGX Orin \u7684 12 GB LPDDR5\u3002\u82e5\u518d\u53e0\u52a0 Speculative Decoding \u7684 drafts tensor\uff0c\u663e\u5b58\u5cf0\u503c\u76f4\u63a5\u98d9\u5230 22 GB\u3002\u8fb9\u7f18\u201c\u5c0f\u76d2\u5b50\u201d\u53ea\u80fd\u964d batch\u3001\u964d\u957f\u5ea6\uff0c\u727a\u7272\u6389\u5b9e\u65f6\u4f53\u9a8c\u3002  <\/p>\n<h3>\u2462 \u661f\u5b87\u667a\u7b97 GPU \u4e91\u4e3b\u673a\u5f39\u6027\u663e\u5b58\u65b9\u6848\uff1a24 GB \u8d77\u6b65\uff0c\u53ef\u6302\u8f7d 80 GB A100<\/h3>\n<p><a href=\"https:\/\/www.starverse-ai.com\">\u661f\u5b87\u667a\u7b97<\/a> \u628a\u201c\u663e\u5b58\u5899\u201d\u53d8\u6210\u201c\u5f39\u6027\u4e91\u76d8\u201d\uff1a<br \/>\n&#8211; <strong>GPU\u4e91\u4e3b\u673a<\/strong> 24 GB RTX 4090 \u5b9e\u4f8b 1.8 \u5143\/\u5c0f\u65f6\u8d77\uff0c\u9002\u5408\u8c03\u8bd5\uff1b<br \/>\n&#8211; \u4e00\u952e\u70ed\u5347\u7ea7 40 GB A100\uff0c\u518d\u6302\u8f7d 80 GB \u663e\u5b58\u6269\u5c55\u5305\uff0c\u603b\u663e\u5b58 120 GB\uff0cINT4 \u6a21\u578b\u53ef\u628a batch \u62c9\u5230 128\uff1b<br \/>\n&#8211; \u4e91\u786c\u76d8\u4e0e\u4e91\u5b58\u50a8\u53cc\u5411\u6253\u901a\uff0c\u6a21\u578b\u6743\u91cd\u6301\u4e45\u5316\u4fdd\u5b58\uff0c\u8de8\u5b9e\u4f8b 30 \u79d2\u5b8c\u6210\u201c\u514b\u9686\u2014\u7eed\u8dd1\u201d\u3002  <\/p>\n<p>\u8fd9\u610f\u5473\u7740\uff0c\u5f00\u53d1\u8005\u65e0\u9700\u91cd\u5199 kernel\uff0c\u4e5f\u65e0\u9700\u91c7\u8d2d\u7269\u7406\u5361\uff0c\u76f4\u63a5\u79df\u7528\u4e91\u7aef\u5927\u663e\u5b58\uff0c\u5c31\u80fd\u628a\u201c\u8fb9\u7f18\u5c0f\u6a21\u578b\u201d\u5f53\u6210\u201c\u4e91\u7aef\u5927\u6a21\u578b\u201d\u6765\u8dd1\u3002<\/p>\n<h3>\u2463 \u6848\u4f8b\uff1a\u538b\u7f29\u7248 Llama-3-8B-Infer \u5728\u5355\u5361\u5b9e\u73b0 5000 tokens\/s<\/h3>\n<p>\u67d0 AIGC \u521b\u4e1a\u56e2\u961f\u5c06 Llama-3-8B \u84b8\u998f\u81f3 2B+INT4\uff0c\u5728\u661f\u5b87\u667a\u7b97 80 GB A100 \u5b9e\u4f8b\u4e0a\u542f\u7528 vLLM + PagedAttention\uff1a<br \/>\n&#8211; \u8f93\u5165\u957f\u5ea6 4 k\uff0c\u8f93\u51fa\u957f\u5ea6 1 k\uff0cbatch=64\uff1b<br \/>\n&#8211; \u5b9e\u6d4b\u9996 token \u5ef6\u8fdf 120 ms\uff0c\u6301\u7eed\u541e\u5410 5000 tokens\/s\uff1b<br \/>\n&#8211; \u5bf9\u6bd4\u672c\u5730 4090 24 GB \u65b9\u6848\uff0c\u541e\u5410\u63d0\u5347 3.2 \u500d\uff0c\u6210\u672c\u53cd\u800c\u4e0b\u964d 42%\uff08\u6309 0.004 \u5143\/1k tokens \u8ba1\uff09\u3002  <\/p>\n<p>\u56e2\u961f\u628a\u7701\u4e0b\u7684 6 \u4e07\u5143\u786c\u4ef6\u9884\u7b97\u6295\u5165\u6570\u636e\u6807\u6ce8\uff0c\u4e24\u5468\u5185\u8fed\u4ee3 4 \u7248\u6a21\u578b\uff0c\u7528\u6237\u7559\u5b58\u63d0\u5347 18%\u3002<\/p>\n<h3>\u2464 \u5f00\u53d1\u8005\u6536\u76ca\uff1a\u6d77\u91cf\u6570\u636e\u96c6\u3001Checkpoint \u65ad\u70b9\u7eed\u8bad\u3001\u4e00\u952e\u5207\u6362\u63a8\u7406<\/h3>\n<ul>\n<li><strong>\u516c\u5171\u8d44\u6e90\u6c60<\/strong>\u5185\u7f6e 3 TB \u4e2d\u6587\u6307\u4ee4\u6570\u636e\u96c6\u3001200+ \u5fae\u8c03\u6a21\u578b\uff0c\u53ef\u76f4\u63a5 <code>cp \/datasets\/Chinese-Instruction-3T .<\/code> \u5f00\u7ec3\uff1b  <\/li>\n<li>\u8bad\u7ec3\u4e2d\u65ad\uff1f\u5e73\u53f0\u81ea\u52a8\u6bcf 30 \u5206\u949f\u751f\u6210 Checkpoint\uff0c\u5199\u5165 <strong>\u4e91\u786c\u76d8<\/strong>\uff0c\u4e0b\u6b21\u5b9e\u4f8b\u5f00\u673a\u81ea\u52a8\u52a0\u8f7d\uff1b  <\/li>\n<li>\u8bad\u7ec3\u2014\u63a8\u7406\u4e00\u952e\u5207\u6362\uff1a\u8bad\u7ec3\u955c\u50cf\u81ea\u5e26 vLLM\u3001TGI\u3001TensorRT-LLM \u73af\u5883\uff0c<code>bash switch_to_infer.sh<\/code> 30 \u79d2\u5b8c\u6210\u670d\u52a1\u5316\uff0c\u65e0\u9700\u91cd\u590d\u914d\u73af\u5883\u3002  <\/li>\n<\/ul>\n<p>\u6b64\u5916\uff0c\u661f\u5b87\u667a\u7b97\u63d0\u4f9b <strong>AI\u5e94\u7528<\/strong> \u5e02\u573a\uff0c\u538b\u7f29\u6a21\u578b\u53ef\u76f4\u63a5\u4e0a\u67b6\uff0c\u6309\u8c03\u7528\u91cf\u5206\u6210\uff0c\u8ba9\u7b97\u6cd5\u56e2\u961f\u591a\u4e00\u6761\u53d8\u73b0\u901a\u9053\u3002<\/p>\n<h3>\u2465 \u547c\u53eb\u884c\u52a8\uff1a\u7acb\u5373\u8bd5\u7528 GPU \u670d\u52a1\u5668\u79df\u7528\uff0c\u4f53\u9a8c AI \u5e94\u7528\u6781\u901f\u4e0a\u7ebf<\/h3>\n<p>\u65b0\u7528\u6237\u6ce8\u518c\u5373\u9001 10 \u5143\u4f53\u9a8c\u91d1\uff0c\u53ef\u96f6\u6210\u672c\u8dd1\u6ee1 5 \u5c0f\u65f6 RTX 4090 <a href=\"https:\/\/www.starverse-ai.com\">GPU\u670d\u52a1\u5668\u79df\u7528<\/a>\u3002<br \/>\n\u73b0\u5728\u5c31\u8bbf\u95ee <a href=\"https:\/\/www.starverse-ai.com\">\u661f\u5b87\u667a\u7b97<\/a>\uff0c\u4e0a\u4f20\u4f60\u7684\u538b\u7f29\u6a21\u578b\uff0c\u611f\u53d7\u201c24 GB \u8d77\u6b65\u300180 GB \u5f39\u6027\u201d\u7684\u53cc\u500d\u5feb\u4e50\u2014\u2014\u8ba9\u8fb9\u7f18\u5c0f\u6a21\u578b\uff0c\u5728\u4e91\u7aef\u5927\u663e\u5b58\u91cc\u8d77\u98de\uff01<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u80cc\u666f\uff1aMIT 2024 \u5e74 5 \u6708\u53d1\u5e03\u7684\u300aThe State&hellip;<\/p>\n","protected":false},"author":2,"featured_media":3190,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-3191","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-zixun"],"views":39,"_links":{"self":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/3191","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/comments?post=3191"}],"version-history":[{"count":0,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/3191\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media\/3190"}],"wp:attachment":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media?parent=3191"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/categories?post=3191"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/tags?post=3191"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}