{"id":169,"date":"2025-06-27T16:33:00","date_gmt":"2025-06-27T08:33:00","guid":{"rendered":"http:\/\/159.223.199.105\/?p=169"},"modified":"2025-09-17T01:11:56","modified_gmt":"2025-09-16T17:11:56","slug":"%e5%a6%82%e4%bd%95%e4%bc%b0%e7%ae%97%e5%a4%a7%e8%af%ad%e8%a8%80%e6%a8%a1%e5%9e%8bllm-%e7%9a%84-gpu-%e5%86%85%e5%ad%98%ef%bc%9a%e7%94%a8%e4%ba%8e%e6%9c%ac%e5%9c%b0%e8%bf%90%e8%a1%8c-llm-%e7%9a%84-gpu","status":"publish","type":"post","link":"http:\/\/www.applicationsofllm.com\/index.php\/2025\/06\/27\/%e5%a6%82%e4%bd%95%e4%bc%b0%e7%ae%97%e5%a4%a7%e8%af%ad%e8%a8%80%e6%a8%a1%e5%9e%8bllm-%e7%9a%84-gpu-%e5%86%85%e5%ad%98%ef%bc%9a%e7%94%a8%e4%ba%8e%e6%9c%ac%e5%9c%b0%e8%bf%90%e8%a1%8c-llm-%e7%9a%84-gpu\/","title":{"rendered":"\u5982\u4f55\u4f30\u7b97\u5927\u8bed\u8a00\u6a21\u578bLLM \u7684 GPU \u5185\u5b58\uff1a\u7528\u4e8e\u672c\u5730\u8fd0\u884c LLM \u7684 GPU \u5185\u5b58"},"content":{"rendered":"\n

\u968f\u7740 DeepSeek\u3001GPT\u3001Llama\u3001Mistral \u7b49\u5927\u89c4\u6a21 LLM \u7684\u5174\u8d77\uff0cAI \u4ece\u4e1a\u8005\u9762\u4e34\u7684\u6700\u5927\u6311\u6218\u4e4b\u4e00\u662f\u786e\u5b9a\u9700\u8981\u591a\u5c11 GPU \u5185\u5b58\u624d\u80fd\u9ad8\u6548\u5730\u670d\u52a1\u4e8e\u8fd9\u4e9b\u6a21\u578b\u3002GPU \u8d44\u6e90\u6602\u8d35\u4e14\u7a00\u7f3a\uff0c\u56e0\u6b64\u4f18\u5316\u5185\u5b58\u5206\u914d\u81f3\u5173\u91cd\u8981\u3002<\/strong><\/p>\n\n\n\n

\u672c\u6307\u5357\u5c06\u5f15\u5bfc\u60a8\u4f7f\u7528\u4e00\u4e2a\u7b80\u5355\u800c\u6709\u6548\u7684\u516c\u5f0f\u6765\u4f30\u7b97\u670d\u52a1 LLM \u6240\u9700\u7684 GPU \u5185\u5b58\u3002\u65e0\u8bba\u60a8\u662f\u90e8\u7f72\u6a21\u578b\u8fdb\u884c\u63a8\u7406\uff0c\u8fd8\u662f\u9488\u5bf9\u7279\u5b9a\u4efb\u52a1\u8fdb\u884c\u5fae\u8c03\uff0c\u8fd9\u4e9b\u77e5\u8bc6\u90fd\u5c06\u5e2e\u52a9\u60a8\u6709\u6548\u5730\u89c4\u5212\u57fa\u7840\u67b6\u6784\u3002<\/p>\n\n\n\n

GPU\u5185\u5b58\u4f30\u7b97\u516c\u5f0f<\/strong>\u4e3a\u4e86\u8ba1\u7b97\u670d\u52a1 LLM \u6240\u9700\u7684 GPU \u5185\u5b58\uff0c\u6211\u4eec\u4f7f\u7528\u4ee5\u4e0b\u516c\u5f0f\uff1a<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u53c2\u6570\u8bf4\u660e\uff1a<\/strong><\/p>\n\n\n\n

M<\/strong>\uff1a\u6240\u9700\u7684 GPU \u5185\u5b58\uff08\u4ee5\u5343\u5146\u5b57\u8282 (GB)\u4e3a\u5355\u4f4d\uff09<\/strong><\/p>\n\n\n\n

P<\/strong>\uff1a\u6a21\u578b\u4e2d\u7684\u53c2\u6570<\/strong>\u6570\u91cf\uff08\u4f8b\u5982\uff0c7B \u6a21\u578b\u6709 70 \u4ebf\u4e2a\u53c2\u6570\uff09<\/p>\n\n\n\n

4B<\/strong>\uff1a\u6bcf\u4e2a\u53c2\u6570 4 \u4e2a\u5b57\u8282<\/strong>\uff08\u5047\u8bbe\u5168\u7cbe\u5ea6 FP32\uff09<\/p>\n\n\n\n

32<\/strong>\uff1a4 \u4e2a\u5b57\u8282\u6709 32 \u4f4d<\/strong><\/p>\n\n\n\n

Q<\/strong>\uff1a\u6bcf\u4e2a\u53c2\u6570\u7528\u4e8e<\/strong>\u6a21\u578b\u5b58\u50a8\u7684\u4f4d\u6570\uff08\u4f8b\u5982\uff0cFP16 = 16 \u4f4d\u3001INT8 = 8 \u4f4d\u7b49\uff09<\/p>\n\n\n\n

1.2<\/strong>\uff1a\u8868\u793a\u6fc0\u6d3b\u5b58\u50a8\u3001\u6ce8\u610f\u952e\u503c\u7f13\u5b58\u7b49\u989d\u5916\u5185\u5b58\u9700\u6c42\u768420% \u5f00\u9500\u3002<\/strong><\/p>\n\n\n\n

\u6b64\u516c\u5f0f\u53ef\u5e2e\u52a9\u60a8\u786e\u5b9a\u5728\u8003\u8651\u4e0d\u540c\u91cf\u5316\u7ea7\u522b\u548c\u5f00\u9500\u7684\u540c\u65f6\u5c06\u6a21\u578b\u52a0\u8f7d\u5230\u5185\u5b58\u4e2d\u9700\u8981\u591a\u5c11 GPU \u5185\u5b58\u3002<\/p>\n\n\n\n

\u793a\u4f8b<\/strong><\/p>\n\n\n\n

\u5047\u8bbe\u60a8\u60f3\u4f30\u7b97FP16 \u7cbe\u5ea6<\/strong>\u7684Llama 70B<\/strong>\u6240\u9700\u7684 GPU \u5185\u5b58\u3002<\/p>\n\n\n\n

\u9274\u4e8e\uff1a<\/strong><\/p>\n\n\n\n

P = 70B<\/strong>\uff08700\u4ebf\u4e2a\u53c2\u6570\uff09<\/p>\n\n\n\n

Q = 16<\/strong>\uff08\u56e0\u4e3a\u6211\u4eec\u4f7f\u7528 FP16 \u7cbe\u5ea6\uff09<\/p>\n\n\n\n

\u5f00\u9500\u7cfb\u6570 = 1.2<\/strong><\/p>\n\n\n\n

\u73b0\u5728\uff0c\u5e94\u7528\u516c\u5f0f\uff1a<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u8f6c\u6362\u4e3aGB\uff1a<\/strong><\/p>\n\n\n\n

\u7531\u4e8e1 GB = 10\u2079\u00a0\u5b57\u8282<\/strong>\uff0c\u6211\u4eec\u9664\u4ee510\u2079<\/strong>\uff1a<\/p>\n\n\n\n

\u56e0\u6b64\uff0c\u8981\u5728FP16<\/strong>\u4e2d\u52a0\u8f7d Llama 70B\uff0c\u60a8\u9700\u8981168GB \u7684 GPU \u5185\u5b58<\/strong>\u3002<\/p>\n\n\n\n

\u91cf\u5316\u4f1a\u53d1\u751f\u4ec0\u4e48\uff1f<\/strong><\/p>\n\n\n\n

\u91cf\u5316\u4f7f\u6211\u4eec\u80fd\u591f\u4ee5\u8f83\u4f4e\u7684\u7cbe\u5ea6\u5b58\u50a8\u6a21\u578b\u6743\u91cd\uff0c\u4ece\u800c\u51cf\u5c11\u5185\u5b58\u9700\u6c42\u3002\u4ee5\u4e0b\u662f Llama 70B \u5728\u4e0d\u540c\u4f4d\u683c\u5f0f\u4e0b\u6240\u9700\u7684\u5185\u5b58\u91cf\uff1a<\/p>\n\n\n\n

\u7cbe\u5ea6 (Q) GPU \u5185\u5b58\u8981\u6c42<\/p>\n\n\n\n

FP32\uff0832 \u4f4d\uff09<\/strong>336 GB<\/p>\n\n\n\n

FP16\uff0816\u4f4d\uff09<\/strong>168 GB<\/p>\n\n\n\n

INT8\uff088\u4f4d\uff09<\/strong>84 GB<\/p>\n\n\n\n

4\u4f4d\u91cf\u5316<\/strong>42GB<\/p>\n\n\n\n

\u5173\u952e\u8981\u70b9\uff1a\u7cbe\u5ea6\u8f83\u4f4e\u7684\u6a21\u578b<\/strong>\u6240\u9700\u7684GPU \u5185\u5b58\u660e\u663e\u8f83\u5c11<\/strong>\u3002<\/p>\n\n\n\n

4 \u4f4d\u91cf\u5316<\/strong>\u975e\u5e38\u8282\u7701\u5185\u5b58\uff0c\u5927\u91cf\u6a21\u578b\u53ef\u4ee5\u9002\u5e94 RTX 4090\uff0824GB VRAM\uff09\u7b49\u6d88\u8d39\u7ea7 GPU\u3002<\/p>\n\n\n\n

FP16 \u662f\u5e73\u8861\u6027\u80fd\u548c\u5185\u5b58\u4f7f\u7528\u60c5\u51b5\u7684\u884c\u4e1a\u6807\u51c6\u3002<\/strong><\/p>\n\n\n\n

\u4f18\u5316\u6a21\u578b\u90e8\u7f72<\/strong><\/p>\n\n\n\n

\u5982\u679c\u4f60\u7684 GPU \u5185\u5b58\u6709\u9650\uff0c\u8fd9\u91cc\u6709\u4e00\u4e9b\u4f18\u5316\u7b56\u7565<\/strong>\uff1a<\/p>\n\n\n\n

\u4f7f\u7528\u91cf\u5316<\/strong>\uff1a\u5c06\u6a21\u578b\u8f6c\u6362\u4e3a 8 \u4f4d\u6216 4 \u4f4d\u4ee5\u51cf\u5c11\u5185\u5b58\u5360\u7528\u3002<\/p>\n\n\n\n

\u5378\u8f7d\u5230 CPU<\/strong>\uff1a\u4e00\u4e9b\u6743\u91cd\u53ef\u4ee5\u5378\u8f7d\u5230 CPU\uff0c\u4ece\u800c\u51cf\u5c11 GPU \u5185\u5b58\u4f7f\u7528\u91cf\u3002<\/p>\n\n\n\n

\u4f7f\u7528\u6a21\u578b\u5e76\u884c\u6027<\/strong>\uff1a\u5728\u591a\u4e2a GPU \u4e4b\u95f4\u5206\u5272\u6a21\u578b\u6743\u91cd\u3002<\/p>\n\n\n\n

\u4f18\u5316KV Cache<\/strong>\uff1a\u51cf\u5c11\u5b58\u50a8\u6ce8\u610f\u952e\u503c\u5bf9\u7684\u6570\u91cf\u3002<\/p>\n\n\n\n

\u5229\u7528\u9ad8\u6548\u7684\u670d\u52a1\u6846\u67b6<\/strong>\uff1a\u4f7f\u7528 vLLM \u6216 TensorRT-LLM \u7b49\u5de5\u5177\u8fdb\u884c\u4f18\u5316\u63a8\u7406\u3002<\/p>\n\n\n\n

\u5c0f\u7ed3<\/strong><\/p>\n\n\n\n

\u8ba1\u7b97\u7528\u4e8e\u670d\u52a1 LLM \u7684 GPU \u5185\u5b58\u5bf9\u4e8e\u9ad8\u6548\u6269\u5c55 AI \u5e94\u7528\u7a0b\u5e8f\u81f3\u5173\u91cd\u8981\u3002\u4f7f\u7528\u4e0a\u9762\u7684\u7b80\u5355\u516c\u5f0f\uff0c\u60a8\u53ef\u4ee5\u4f30\u7b97\u4e0d\u540c\u7cbe\u5ea6\u7ea7\u522b\u6240\u9700\u7684 VRAM\uff0c\u5e76\u76f8\u5e94\u5730\u4f18\u5316\u90e8\u7f72\u3002<\/p>\n\n\n\n

\u5982\u679c\u60a8\u6b63\u5728\u5904\u7406\u50cf Llama 70B \u8fd9\u6837\u7684\u5927\u89c4\u6a21\u6a21\u578b\uff0c\u91cf\u5316\u548c\u5e76\u884c\u6027\u662f\u63a7\u5236 GPU \u6210\u672c\u7684\u6700\u4f73\u65b9\u6cd5<\/strong>\u3002\u901a\u8fc7\u5e94\u7528\u8fd9\u4e9b\u4f18\u5316\uff0c\u60a8\u53ef\u4ee5\u8fd0\u884c\u5f3a\u5927\u7684 AI \u6a21\u578b\uff0c\u800c\u65e0\u9700\u5728\u9ad8\u7aef\u786c\u4ef6\u4e0a\u6295\u5165\u5de8\u8d44\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

\u968f\u7740 DeepSeek\u3001GPT\u3001Llama\u3001Mistral \u7b49\u5927\u89c4\u6a21 LLM \u7684\u5174\u8d77\uff0cAI \u4ece\u4e1a\u8005\u9762\u4e34\u7684\u6700\u5927\u6311 […]<\/p>\n","protected":false},"author":1,"featured_media":174,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[7],"tags":[],"class_list":["post-169","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-7"],"_links":{"self":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/169","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/comments?post=169"}],"version-history":[{"count":1,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/169\/revisions"}],"predecessor-version":[{"id":173,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/169\/revisions\/173"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/media\/174"}],"wp:attachment":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/media?parent=169"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/categories?post=169"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/tags?post=169"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}