

{"id":404,"date":"2025-11-25T09:13:32","date_gmt":"2025-11-25T01:13:32","guid":{"rendered":"https:\/\/high-flyer.in.suopu.cc\/?p=404"},"modified":"2025-11-25T09:13:32","modified_gmt":"2025-11-25T01:13:32","slug":"continuous-batching%ef%bc%9a%e4%b8%80%e7%a7%8d%e6%8f%90%e5%8d%87-llm-%e9%83%a8%e7%bd%b2%e5%90%9e%e5%90%90%e9%87%8f%e7%9a%84%e5%88%a9%e5%99%a8","status":"publish","type":"post","link":"https:\/\/high-flyer.in.suopu.cc\/en\/blog\/404\/","title":{"rendered":"Continuous Batching\uff1a\u4e00\u79cd\u63d0\u5347 LLM \u90e8\u7f72\u541e\u5410\u91cf\u7684\u5229\u5668"},"content":{"rendered":"<p>\u7531\u4e8e LLM \u5de8\u5927\u7684 GPU \u5185\u5b58\u5f00\u9500\u548c\u8ba1\u7b97\u6210\u672c\uff0c\u5728\u5927\u591a\u6570\u5e94\u7528\u4e2d\uff0c\u673a\u5668\u5b66\u4e60\u5de5\u7a0b\u5e08\u901a\u5e38\u901a\u8fc7\u5185\u90e8\u8c03\u6574\uff08\u5982\u91cf\u5316\u548c\u5bf9 CUDA \u6838\u7684\u5b9a\u5236\uff09\u6765\u4f18\u5316\u3002\u7136\u800c\uff0c\u7531\u4e8e LLM \u901a\u8fc7\u8fed\u4ee3\u751f\u6210\u5176\u8f93\u51fa\uff0c\u5e76\u4e14 LLM \u63a8\u7406\u901a\u5e38\u6d89\u53ca\u5185\u5b58\u800c\u4e0d\u662f\u8ba1\u7b97\uff0c\u56e0\u6b64\u5728\u5f88\u591a\u5b9e\u8df5\u4e2d\uff0c\u4f18\u5316\u7cfb\u7edf\u7ea7\u6279\u5904\u7406\u53ef\u4ee5\u4f7f\u6027\u80fd\u5dee\u5f02\u8fbe\u523010\u500d\u751a\u81f3\u66f4\u591a\u3002<\/p>\n<p>\u4e00\u79cd\u6700\u8fd1\u63d0\u51fa\u7684\u4f18\u5316\u65b9\u6cd5\u662f\u8fde\u7eed\u6279\u5904\u7406\uff08Continuous batching\uff09\uff0c\u4e5f\u79f0\u4e3a\u52a8\u6001\u6279\u5904\u7406\u6216\u57fa\u4e8e\u8fed\u4ee3\u7ea7\u7684\u6279\u5904\u7406\u3002\u5176\u5177\u6709\u5982\u4e0b\u60ca\u4eba\u7684\u6548\u679c\uff1a<\/p>\n<ul>\n<li>\u57fa\u4e8evLLM\uff0c\u4f7f\u7528\u8fde\u7eed\u6279\u5904\u7406\u548c\u8fde\u7eed\u6279\u5904\u7406\u7279\u5b9a\u7684\u5185\u5b58\u4f18\u5316\uff0c\u53ef\u4ee5\u5b9e\u73b0\u591a\u8fbe23\u500d\u7684\u541e\u5410\u91cf\u63d0\u5347\uff1b<\/li>\n<li>\u5bf9\u4e8e HuggingFace \u672c\u5730\u751f\u6210\u63a8\u7406\uff0c\u4f7f\u7528\u8fde\u7eed\u6279\u5904\u7406\uff0c\u53ef\u4ee5\u5b9e\u73b08\u500d\u7684\u541e\u5410\u91cf\u63d0\u5347\uff1b<\/li>\n<li>\u57fa\u4e8e NVIDIA \u7684 FasterTransformer\uff0c\u4f7f\u7528\u4f18\u5316\u8fc7\u7684\u6a21\u578b\u5b9e\u73b0\uff0c\u53ef\u4ee5\u5b9e\u73b04\u500d\u7684\u541e\u5410\u91cf\u63d0\u5347\u3002<\/li>\n<\/ul>\n<p>\u672c\u535a\u5ba2\u63a5\u4e0b\u6765\u5c06\u8be6\u7ec6\u4ecb\u7ecd\u76f8\u5173\u6280\u672f\u7684\u7ec6\u8282\u3002<\/p>\n<p><small>\u6587\u7ae0\u7ffb\u8bd1\u81ea:\u00a0<a href=\"https:\/\/www.anyscale.com\/blog\/continuous-batching-llm-inference\">https:\/\/www.anyscale.com\/blog\/continuous-batching-llm-inference<\/a><\/small><\/p>\n<h2 id=\"llm-\u63a8\u7406\">LLM \u63a8\u7406<\/h2>\n<p>LLM \u63a8\u7406\u662f\u4e00\u4e2a\u8fed\u4ee3\u8fc7\u7a0b\uff0c\u5728\u6bcf\u4e2a\u65b0\u524d\u9988\u5faa\u73af\u540e\u83b7\u5f97\u4e00\u4e2a\u989d\u5916\u7684\u5b8c\u6210\u6807\u8bb0\u3002\u4f8b\u5982\uff0c\u5982\u679c\u60a8\u63d0\u793a\u4e00\u4e2a\u53e5\u5b50\u201dWhat is the capital of California:\u201c\uff0c\u5b83\u9700\u8981\u8fdb\u884c\u5341\u6b21\u524d\u9988\u5faa\u73af\u624d\u80fd\u5f97\u5230\u5b8c\u6574\u7684\u56de\u7b54[\u201cS\u201d,\u201ca\u201d,\u201cc\u201d,\u201cr\u201d,\u201ca\u201d,\u201cm\u201d,\u201ce\u201d,\u201cn\u201d,\u201ct\u201d,\u201co\u201d]\u3002<\/p>\n<p>As shown in the figure below:<\/p>\n<p><span class=\"gatsby-resp-image-wrapper\"><a class=\"gatsby-resp-image-link\" href=\"https:\/\/hfai-static.high-flyer.cn\/static\/e12243478cfe843761c4bf905292ad5b\/3075e\/01.png\" target=\"_blank\" rel=\"noopener\"><img fetchpriority=\"high\" decoding=\"async\" class=\"alignnone size-full wp-image-408\" title=\"96a3be3cf272e017046d65\" src=\"http:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/96a3be3cf272e017046d65.png\" alt=\"96a3be3cf272e017046d65\" width=\"621\" height=\"300\" srcset=\"https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/96a3be3cf272e017046d65.png 621w, https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/96a3be3cf272e017046d65-300x145.png 300w, https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/96a3be3cf272e017046d65-150x72.png 150w\" sizes=\"(max-width: 621px) 100vw, 621px\" \/><\/a><\/span><\/p>\n<p>\u4e0a\u56fe\u793a\u4f8b\u663e\u793a\u4e86\u4e00\u4e2a\u652f\u6301\u6700\u5927\u5e8f\u5217\u957f\u5ea6\u4e3a8\u4e2a\u6807\u8bb0\uff08T1\uff0cT2\uff0c\u2026\uff0cT8\uff09\u7684\u5047\u8bbe\u6a21\u578b\u3002\u4ece Prompt\uff08\u9ec4\u8272\uff09\u5f00\u59cb\uff0c\u8fed\u4ee3\u8fc7\u7a0b\u9010\u4e2a\u751f\u6210\u4e00\u4e2a\u6807\u8bb0\uff08\u84dd\u8272\uff09\u3002\u4e00\u65e6\u6a21\u578b\u751f\u6210\u4e86\u4e00\u4e2a\u7ed3\u675f\u5e8f\u5217\u6807\u8bb0\uff08\u7ea2\u8272\uff09\uff0c\u751f\u6210\u5faa\u73af\u505c\u6b62\u3002\u8fd9\u4e2a\u4f8b\u5b50\u663e\u793a\u4e86\u4ec5\u5305\u542b\u4e00\u4e2a\u8f93\u5165\u5e8f\u5217\u7684 Batch\uff0c\u56e0\u6b64 Batch \u5927\u5c0f\u4e3a1\u3002<\/p>\n<p>\u73b0\u5728\uff0c\u6211\u4eec\u5df2\u7ecf\u4e86\u89e3\u4e86 LLM \u63a8\u7406\u7684\u8fed\u4ee3\u8fc7\u7a0b\uff0c\u63a5\u4e0b\u6765\u6211\u4eec\u6765\u63a2\u8ba8\u4e00\u4e9b\u4e0d\u719f\u6089\u7684\u4e8b\u60c5\uff1a<\/p>\n<ol>\n<li>LLM \u63a8\u7406\u662f\u5185\u5b58 IO \u9650\u5236\uff0c\u800c\u4e0d\u662f\u8ba1\u7b97\u9650\u5236\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u76ee\u524d\u52a0\u8f7d 1MB \u7684\u6570\u636e\u5230 GPU \u6240\u9700\u7684\u65f6\u95f4\u6bd4 1MB \u7684\u6570\u636e\u5728GPU\u4e0a\u8ba1\u7b97\u6240\u9700\u7684\u65f6\u95f4\u957f\u3002\u8fd9\u610f\u5473\u7740 LLM \u63a8\u7406\u7684\u541e\u5410\u91cf\u5f88\u5927\u7a0b\u5ea6\u4e0a\u53d6\u51b3\u4e8e\u60a8\u80fd\u5c06\u591a\u5c11\u6279\u6570\u636e\u88c5\u5165\u5230\u9ad8\u901fGPU \u5185\u5b58\u4e2d\uff1b<\/li>\n<li>GPU \u5185\u5b58\u7684\u6d88\u8017\u91cf\u968f\u7740\u57fa\u672c\u6a21\u578b\u5927\u5c0f\u548c\u6807\u8bb0\u957f\u5ea6\u7684\u589e\u52a0\u800c\u589e\u52a0\u3002\u5982\u679c\u6211\u4eec\u5c06\u5e8f\u5217\u957f\u5ea6\u9650\u5236\u4e3a 512\uff0c\u90a3\u4e48\u5728\u4e00\u4e2a\u6279\u5904\u7406\u4e2d\uff0c\u6211\u4eec\u6700\u591a\u53ea\u80fd\u5904\u740628\u4e2a\u5e8f\u5217\uff1b\u4e00\u4e2a\u5e8f\u5217\u957f\u5ea6\u4e3a 2048 \u5219\u6279\u5904\u7406\u5927\u5c0f\u6700\u591a\u53ea\u80fd\u4e3a7\u4e2a\u5e8f\u5217\u3002\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u8fd9\u53ea\u662f\u4e00\u4e2a\u4e0a\u9650\uff0c\u56e0\u4e3a\u4e2d\u95f4\u8ba1\u7b97\u7ed3\u679c\u6ca1\u6709\u7559\u4e0b\u5b58\u50a8\u7684\u7a7a\u95f4\u3002<\/li>\n<\/ol>\n<p>\u8fd9\u610f\u5473\u7740\u4f18\u5316\u5185\u5b58\u4f7f\u7528\u6709\u5f88\u591a\u4f59\u5730\u3002\u8fd9\u5c31\u662f\u4e3a\u4ec0\u4e48\u50cf AutoGPTQ \u8fd9\u6837\u7684\u6a21\u578b\u91cf\u5316\u7b56\u7565\u5177\u6709\u5982\u6b64\u5f3a\u5927\u7684\u529b\u91cf\u7684\u539f\u56e0\uff1b\u5982\u679c\u60a8\u53ef\u4ee5\u5c06\u5185\u5b58\u4f7f\u7528\u91cf\u51cf\u534a\uff0c\u60a8\u5c06\u80fd\u591f\u4e3a\u66f4\u5927\u7684\u6279\u5904\u7406\u63d0\u4f9b\u66f4\u591a\u7684\u7a7a\u95f4\u3002\u7136\u800c\uff0c\u5e76\u4e0d\u662f\u6240\u6709\u7684\u7b56\u7565\u90fd\u9700\u8981\u5bf9\u6a21\u578b\u6743\u91cd\u8fdb\u884c\u4fee\u6539\u3002\u4f8b\u5982\uff0cFlashAttention \u901a\u8fc7\u91cd\u65b0\u7ec4\u7ec7\u6ce8\u610f\u8ba1\u7b97\u6765\u51cf\u5c11\u5185\u5b58IO\uff0c\u4ece\u800c\u53d1\u73b0\u663e\u8457\u7684\u6027\u80fd\u63d0\u5347\u3002<\/p>\n<p>\u8fde\u7eed\u6279\u5904\u7406\u662f\u53e6\u4e00\u79cd\u5185\u5b58\u4f18\u5316\u6280\u672f\uff0c\u5b83\u4e0d\u9700\u8981\u5bf9\u6a21\u578b\u6743\u91cd\u8fdb\u884c\u4fee\u6539\u3002\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u89e3\u91ca\u4e00\u4e0b\u8fde\u7eed\u6279\u5904\u7406\u5982\u4f55\u5de5\u4f5c\uff0c\u4ee5\u53ca\u5982\u4f55\u63d0\u9ad8LLM\u751f\u6210\u8fc7\u7a0b\u7684\u5185\u5b58\u6548\u7387\u3002<\/p>\n<h2 id=\"\u6734\u7d20\u6279\u5904\u7406\u4e0e\u9759\u6001\u6279\u5904\u7406\">\u6734\u7d20\u6279\u5904\u7406\u4e0e\u9759\u6001\u6279\u5904\u7406<\/h2>\n<p>\u6211\u4eec\u79f0\u8fd9\u79cd\u4f20\u7edf\u7684\u6279\u5904\u7406\u65b9\u6cd5\u4e3a\u9759\u6001\u6279\u5904\u7406\uff0c\u56e0\u4e3a\u6279\u5927\u5c0f\u5728\u63a8\u7406\u5b8c\u6210\u4e4b\u524d\u4fdd\u6301\u4e0d\u53d8\u3002\u4ee5\u4e0b\u662fLLM\u63a8\u7406\u4e2d\u9759\u6001\u6279\u5904\u7406\u7684\u793a\u610f\u56fe\uff1a<\/p>\n<p><span class=\"gatsby-resp-image-wrapper\"><a class=\"gatsby-resp-image-link\" href=\"https:\/\/hfai-static.high-flyer.cn\/static\/28d7172c5c27e843bc31bb91aacac16c\/302a4\/02.png\" target=\"_blank\" rel=\"noopener\"><img decoding=\"async\" class=\"alignnone size-full wp-image-405\" title=\"a2ef406e2c2351e0b9e883\" src=\"http:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/a2ef406e2c2351e0b9e883.png\" alt=\"a2ef406e2c2351e0b9e883\" width=\"650\" height=\"151\" srcset=\"https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/a2ef406e2c2351e0b9e883.png 650w, https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/a2ef406e2c2351e0b9e883-300x70.png 300w, https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/a2ef406e2c2351e0b9e883-150x35.png 150w\" sizes=\"(max-width: 650px) 100vw, 650px\" \/><\/a><\/span><\/p>\n<p>\u5982\u4e0a\u56fe\u6240\u793a\uff0c\u5728\u7b2c\u4e00\u904d\u8fed\u4ee3\uff08\u5de6\uff09\u4e2d\uff0c\u6bcf\u4e2a\u5e8f\u5217\u4ece\u63d0\u793a\u8bcd\uff08\u9ec4\uff09\u4e2d\u751f\u6210\u4e00\u4e2a\u6807\u8bb0\uff08\u84dd\u8272\uff09\u3002\u7ecf\u8fc7\u51e0\u8f6e\u8fed\u4ee3\uff08\u53f3\uff09\u540e\uff0c\u5b8c\u6210\u7684\u5e8f\u5217\u5177\u6709\u4e0d\u540c\u7684\u5c3a\u5bf8\uff0c\u56e0\u4e3a\u6bcf\u4e2a\u5e8f\u5217\u5728\u4e0d\u540c\u7684\u8fed\u4ee3\u7ed3\u675f\u65f6\u4ea7\u751f\u4e0d\u540c\u7684\u7ed3\u675f\u5e8f\u5217\u6807\u8bb0\uff08\u7ea2\u8272\uff09\u3002\u5c3d\u7ba1\u5e8f\u52173\u5728\u4e24\u6b21\u8fed\u4ee3\u540e\u5b8c\u6210\uff0c\u4f46\u9759\u6001\u6279\u5904\u7406\u610f\u5473\u7740 GPU \u5c06\u5728\u6279\u5904\u7406\u4e2d\u7684\u6700\u540e\u4e00\u4e2a\u5e8f\u5217\u5b8c\u6210\u3002<\/p>\n<p>\u4e0e\u4f20\u7edf\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u4e0d\u540c\uff0cLLM \u7684\u6279\u5904\u7406\u65b9\u6cd5\u6bd4\u8f83\u590d\u6742\uff0c\u56e0\u4e3a\u5176\u63a8\u7406\u662f\u8fed\u4ee3\u6027\u7684\u3002\u76f4\u89c2\u4e0a\uff0c\u8fd9\u662f\u56e0\u4e3a\u8bf7\u6c42\u53ef\u80fd\u63d0\u524d\u5b8c\u6210\uff0c\u4f46\u662f\u91ca\u653e\u5176\u8d44\u6e90\u5e76\u6dfb\u52a0\u53ef\u80fd\u5904\u4e8e\u4e0d\u540c\u5b8c\u6210\u72b6\u6001\u7684\u65b0\u8bf7\u6c42\u662f\u68d8\u624b\u7684\u3002\u8fd9\u610f\u5473\u7740\uff0c\u968f\u7740 GPU \u5728\u6279\u6b21\u4e2d\u4e0d\u540c\u5e8f\u5217\u7684\u751f\u6210\u957f\u5ea6\u4e0e\u6700\u5927\u751f\u6210\u957f\u5ea6\u4e0d\u540c\u800c\u672a\u5145\u5206\u5229\u7528\u3002\u5728\u4e0a\u9762\u7684\u53f3\u56fe\uff0c\u8fd9\u901a\u8fc7\u5e8f\u52171\u30013\u548c4\u7684\u5e8f\u5217\u672b\u6807\u8bb0\u540e\u7684\u767d\u8272\u65b9\u5757\u6765\u8bf4\u660e\u3002<\/p>\n<p>\u9759\u6001\u6279\u5904\u7406\u4e0d\u5145\u5206\u5229\u7528GPU\u7684\u9891\u7387\u662f\u591a\u5c11\uff1f\u8fd9\u53d6\u51b3\u4e8e\u6279\u6b21\u4e2d\u5e8f\u5217\u7684\u751f\u6210\u957f\u5ea6\u3002\u4f8b\u5982\uff0c\u53ef\u4ee5\u4f7f\u7528LLM\u63a8\u65ad\u6765\u53d1\u5c04\u5355\u4e2a\u6807\u8bb0\u4f5c\u4e3a\u5206\u7c7b\u4efb\u52a1\uff08\u6709\u66f4\u597d\u7684\u65b9\u6cd5\u53ef\u4ee5\u505a\u5230\u8fd9\u4e00\u70b9\uff0c\u4f46\u8ba9\u6211\u4eec\u4ee5\u6b64\u4e3a\u4f8b\uff09\u3002\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u6bcf\u4e2a\u8f93\u51fa\u5e8f\u5217\u5177\u6709\u76f8\u540c\u7684\u5927\u5c0f\uff081\u4e2a\u6807\u8bb0\uff09\u3002\u5982\u679c\u8f93\u5165\u5e8f\u5217\u4e5f\u5177\u6709\u76f8\u540c\u7684\u5927\u5c0f\uff08\u4f8b\u5982\uff0c512\u4e2a\u6807\u8bb0\uff09\uff0c\u5219\u6bcf\u4e2a\u9759\u6001\u6279\u6b21\u5c06\u5b9e\u73b0\u6700\u4f73\u7684GPU\u5229\u7528\u7387\u3002<\/p>\n<p>\u53e6\u4e00\u65b9\u9762\uff0cLLM \u9a71\u52a8\u7684\u804a\u5929\u673a\u5668\u4eba\u670d\u52a1\u4e0d\u80fd\u5047\u5b9a\u56fa\u5b9a\u957f\u5ea6\u7684\u8f93\u5165\u5e8f\u5217\uff0c\u4e5f\u4e0d\u80fd\u5047\u5b9a\u56fa\u5b9a\u957f\u5ea6\u7684\u8f93\u51fa\u5e8f\u5217\u3002\u5728\u64b0\u5199\u672c\u6587\u65f6\uff0c\u4e13\u6709\u6a21\u578b\u63d0\u4f9b\u7684\u6700\u5927\u4e0a\u4e0b\u6587\u957f\u5ea6\u8d85\u8fc78K\u4e2a\u6807\u8bb0\u3002\u4f7f\u7528\u9759\u6001\u6279\u5904\u7406\uff0c\u751f\u6210\u8f93\u51fa\u7684\u53d8\u5316\u53ef\u80fd\u4f1a\u5bfc\u81f4GPU\u4e25\u91cd\u672a\u5145\u5206\u5229\u7528\u3002\u96be\u602aOpenAI\u9996\u5e2d\u6267\u884c\u5b98Sam Altman\u79f0\u8ba1\u7b97\u6210\u672c\u9ad8\u5f97\u60ca\u4eba\u3002<\/p>\n<p>\u5982\u679c\u6ca1\u6709\u5bf9\u7528\u6237\u8f93\u5165\u548c\u6a21\u578b\u8f93\u51fa\u7684\u9650\u5236\u5047\u8bbe\uff0c\u672a\u7ecf\u4f18\u5316\u7684\u751f\u4ea7\u7ea7 LLM \u7cfb\u7edf\u65e0\u6cd5\u5145\u5206\u5229\u7528 GPU \u5e76\u5bfc\u81f4\u4e0d\u5fc5\u8981\u7684\u8f83\u9ad8\u6210\u672c\u3002<\/p>\n<h2 id=\"\u8fde\u7eed\u6279\u5904\u7406\">\u8fde\u7eed\u6279\u5904\u7406<\/h2>\n<p>\u884c\u4e1a\u610f\u8bc6\u5230\u8fd9\u79cd\u6279\u5904\u7406\u7684\u4f4e\u6548\u6027\uff0c\u5e76\u63d0\u51fa\u4e86\u66f4\u6709\u6548\u7684\u89e3\u51b3\u65b9\u6848\u3002OSDI 2022 \u4e0a\u53d1\u8868\u7684 Orca \u662f\u7b2c\u4e00\u7bc7\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\u7684\u8bba\u6587\u3002\u5b83\u91c7\u7528\u4e86\u8fed\u4ee3\u7ea7\u8c03\u5ea6\uff0c\u5176\u4e2d\u6279\u5927\u5c0f\u6839\u636e\u6bcf\u6b21\u8fed\u4ee3\u786e\u5b9a\u3002\u7ed3\u679c\u662f\uff0c\u4e00\u65e6\u6279\u4e2d\u7684\u4e00\u4e2a\u5e8f\u5217\u5b8c\u6210\u751f\u6210\uff0c\u5c31\u53ef\u4ee5\u5728\u5176\u4f4d\u7f6e\u63d2\u5165\u4e00\u4e2a\u65b0\u7684\u5e8f\u5217\uff0c\u4ece\u800c\u5b9e\u73b0\u6bd4\u9759\u6001\u6279\u5904\u7406\u66f4\u9ad8\u7684GPU\u5229\u7528\u7387\u3002<\/p>\n<p><span class=\"gatsby-resp-image-wrapper\"><a class=\"gatsby-resp-image-link\" href=\"https:\/\/hfai-static.high-flyer.cn\/static\/10738c93c0d5b194e103c4bd291b65b8\/302a4\/03.png\" target=\"_blank\" rel=\"noopener\"><img decoding=\"async\" class=\"alignnone size-full wp-image-409\" title=\"e45ee7ce7e88149af8dd0\" src=\"http:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/e45ee7ce7e88149af8dd0.png\" alt=\"e45ee7ce7e88149af8dd0\" width=\"650\" height=\"148\" srcset=\"https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/e45ee7ce7e88149af8dd0.png 650w, https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/e45ee7ce7e88149af8dd0-300x68.png 300w, https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/e45ee7ce7e88149af8dd0-150x34.png 150w\" sizes=\"(max-width: 650px) 100vw, 650px\" \/><\/a><\/span><\/p>\n<p>\u5982\u4e0a\u56fe\u6240\u793a\uff0c\u4f7f\u7528\u8fde\u7eed\u6279\u5904\u7406\u5b8c\u6210\u4e03\u6761\u5e8f\u5217\u3002\u5de6\u56fe\u663e\u793a\u4e86\u5355\u4e2a\u8fed\u4ee3\u540e\u7684\u6279\uff0c\u53f3\u56fe\u663e\u793a\u4e86\u591a\u6b21\u8fed\u4ee3\u540e\u7684\u6279\u3002\u4e00\u65e6\u4e00\u4e2a\u5e8f\u5217\u4ea7\u751f\u7ed3\u675f\u5e8f\u5217\u6807\u8bb0\uff0c\u6211\u4eec\u5728\u5176\u4f4d\u7f6e\u63d2\u5165\u65b0\u7684\u5e8f\u5217\uff08\u5373\u5e8f\u5217S5\u3001S6\u548cS7\uff09\u3002\u8fd9\u5b9e\u73b0\u4e86\u66f4\u9ad8\u7684 GPU \u5229\u7528\u7387\uff0c\u56e0\u4e3a GPU \u4e0d\u9700\u8981\u7b49\u5f85\u6240\u6709\u5e8f\u5217\u5b8c\u6210\u624d\u5f00\u59cb\u65b0\u7684\u4e00\u4e2a\u3002<\/p>\n<p>\u73b0\u5b9e\u60c5\u51b5\u6bd4\u8fd9\u4e2a\u7b80\u5316\u6a21\u578b\u66f4\u590d\u6742\uff1a\u56e0\u4e3a\u9884\u586b\u5145\u9636\u6bb5\u9700\u8981\u8ba1\u7b97\uff0c\u5e76\u4e14\u4e0e\u751f\u6210\u9636\u6bb5\u7684\u8ba1\u7b97\u6a21\u5f0f\u4e0d\u540c\uff0c\u56e0\u6b64\u5b83\u4e0d\u80fd\u5f88\u5bb9\u6613\u5730\u4e0e\u4ee4\u724c\u7684\u751f\u6210\u4e00\u8d77\u8fdb\u884c\u6279\u91cf\u3002\u8fde\u7eed\u6279\u5904\u7406\u6846\u67b6\u76ee\u524d\u901a\u8fc7\u8d85\u53c2\u6570\u6765\u7ba1\u7406\u8fd9\u4e2a\u95ee\u9898\uff1a\u7b49\u5f85\u5df2\u670d\u52a1\u6bd4\u548c\u7b49\u5f85\u7ed3\u675f\u5e8f\u5217\u6807\u8bb0\u7684\u8bf7\u6c42\u6bd4\uff08waiting_served_ratio\uff09\u3002<\/p>\n<p>\u76ee\u524d\uff0cHuggingFace \u5df2\u7ecf\u5c06\u8fde\u7eed\u6279\u5904\u7406\u4f5c\u4e3a\u4ed6\u4eec\u7684\u57fa\u4e8e Rust \u548c Python \u7684\u6587\u672c\u751f\u6210\u63a8\u7406\u670d\u52a1\u5668\u7684\u751f\u4ea7\u7248\u672c\u3002\u6211\u4eec\u4f7f\u7528\u4ed6\u4eec\u7684\u5b9e\u73b0\u6765\u4e86\u89e3\u8fde\u7eed\u6279\u5904\u7406\u5728 benchmarks \u4e2d\u7684\u6027\u80fd\u7279\u5f81\u3002<\/p>\n<h2 id=\"\u5b9e\u9a8c\u9a8c\u8bc1\">\u5b9e\u9a8c\u9a8c\u8bc1<\/h2>\n<p>\u6211\u4eec\u65e8\u5728\u4e86\u89e3\u8fde\u7eed\u6279\u91cf\u5904\u7406\u4e0e\u9759\u6001\u6279\u5904\u7406\u5728\u6a21\u62df\u5b9e\u65f6\u63a8\u7406\u5de5\u4f5c\u8d1f\u8f7d\u4e0a\u7684\u8868\u73b0\u3002\u672c\u8d28\u4e0a\uff0c\u6211\u4eec\u5173\u5fc3\u6210\u672c\u3002\u6211\u4eec\u5c06\u6210\u672c\u5206\u89e3\u4e3a\u541e\u5410\u91cf\uff08Throughput\uff09\u548c\u5ef6\u8fdf\uff08Latency\uff09\u8fd9\u4e24\u4e2a\u6307\u6807\uff0c\u56e0\u4e3a\u6210\u672c\u76f4\u63a5\u53d6\u51b3\u4e8e\u60a8\u5728\u7ed9\u5b9a\u5ef6\u8fdf\u4e0b\u4ee5\u591a\u9ad8\u6548\u7684\u65b9\u5f0f\u63d0\u4f9b\u670d\u52a1\u7684\u7a0b\u5ea6\u3002<\/p>\n<p>\u6211\u4eec\u5728\u4e00\u4e2a NVIDIA A100 GPU \u4e0a\u8fdb\u884c\u541e\u5410\u91cf\uff08Throughput\uff09\u548c\u5ef6\u8fdf\uff08Latency\uff09\u7684\u57fa\u51c6\u6d4b\u8bd5\u3002\u6211\u4eec\u7684 A100 \u6709 40GB \u7684 GPU \u5185\u5b58\u3002\u6211\u4eec\u9009\u62e9\u4e86 Meta\u7684OPT-13B \u6a21\u578b\uff0c\u56e0\u4e3a\u6bcf\u4e2a\u6d4b\u8bd5\u6846\u67b6\u90fd\u4e0e\u6b64\u6a21\u578b\u914d\u5408\u826f\u597d\u3002\u6211\u4eec\u9009\u62e9\u4e86 13B\u53d8\u4f53\uff0c\u56e0\u4e3a\u5b83\u65e2\u4e0d\u9700\u8981\u5f20\u91cf\u5e76\u884c\uff0c\u53c8\u8db3\u591f\u5927\u4ee5\u4ea7\u751f\u5185\u5b58\u6548\u7387\u6311\u6218\u3002\u6211\u4eec\u9009\u62e9\u4e0d\u4f7f\u7528\u5f20\u91cf\u5e76\u884c\uff0c\u5c3d\u7ba1\u9759\u6001\u6279\u5904\u7406\u548c\u8fde\u7eed\u6279\u5904\u7406\u90fd\u53ef\u4ee5\u4f7f\u7528\u5f20\u91cf\u5e76\u884c\u3002<\/p>\n<p>\u6211\u4eec\u6d4b\u8bd5\u4e86\u4e24\u79cd\u9759\u6001\u6279\u5904\u7406\u6846\u67b6\u548c\u4e09\u79cd\u8fde\u7eed\u6279\u5904\u7406\u6846\u67b6\uff0c\u4ed6\u4eec\u5206\u522b\u662f\uff1a<\/p>\n<p><span class=\"gatsby-resp-image-wrapper\"><a class=\"gatsby-resp-image-link\" href=\"https:\/\/hfai-static.high-flyer.cn\/static\/457efa838e8773189ff0babd287c3c06\/5fd3e\/04.png\" target=\"_blank\" rel=\"noopener\"><img loading=\"lazy\" decoding=\"async\" class=\"alignnone size-full wp-image-406\" title=\"7d0665438e81d8eceb9877\" src=\"http:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/7d0665438e81d8eceb9877.png\" alt=\"7d0665438e81d8eceb9877\" width=\"594\" height=\"297\" srcset=\"https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/7d0665438e81d8eceb9877.png 594w, https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/7d0665438e81d8eceb9877-300x150.png 300w, https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/7d0665438e81d8eceb9877-150x75.png 150w\" sizes=\"(max-width: 594px) 100vw, 594px\" \/><\/a><\/span><\/p>\n<h3 id=\"\u541e\u5410\u91cf\u6d4b\u8bd5\u7ed3\u679c\">\u541e\u5410\u91cf\u6d4b\u8bd5\u7ed3\u679c<\/h3>\n<p>\u6839\u636e\u6211\u4eec\u5bf9\u9759\u6001\u6279\u5904\u7406\u7684\u4e86\u89e3\uff0c\u6211\u4eec\u731c\u6d4b\u5728\u5e8f\u5217\u957f\u5ea6\u6ce2\u52a8\u8f83\u5927\u7684\u6bcf\u4e2a\u6279\u6b21\u4e2d\uff0c\u8fde\u7eed\u6279\u5904\u7406\u7684\u8868\u73b0\u4f1a\u663e\u8457\u66f4\u597d\u3002\u4e3a\u4e86\u8bc1\u660e\u8fd9\u4e00\u70b9\uff0c\u6211\u4eec\u9488\u5bf9\u6bcf\u4e2a\u6846\u67b6\u5206\u522b\u8fd0\u884c\u6211\u4eec\u7684\u541e\u5410\u91cf\u57fa\u51c6\u6d4b\u8bd54\u6b21\uff0c\u6bcf\u6b21\u5728\u5177\u6709\u66f4\u9ad8\u5e8f\u5217\u957f\u5ea6\u6ce2\u52a8\u7684\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u3002<\/p>\n<p>\u4e3a\u4e86\u505a\u5230\u8fd9\u4e00\u70b9\uff0c\u6211\u4eec\u521b\u5efa\u4e86\u4e00\u4e2a\u5305\u542b 1000 \u4e2a\u5e8f\u5217\uff0c\u6bcf\u4e2a\u5e8f\u5217\u5177\u6709 512 \u4e2a\u8f93\u5165\u6807\u8bb0\u7684\u6570\u636e\u96c6\u3002\u6211\u4eec\u901a\u8fc7\u5ffd\u7565\u7aef\u5c3e\u6807\u8bb0\u5e76\u8bbe\u7f6e\u6700\u5927\u6807\u8bb0\u914d\u7f6e\u6765\u59cb\u7ec8\u751f\u6210\u6bcf\u4e2a\u8bf7\u6c42\u7684\u5e8f\u5217\u957f\u5ea6\u3002\u7136\u540e\u6211\u4eec\u751f\u6210 1000 \u4e2a\u751f\u6210\u957f\u5ea6\uff0c\u6bcf\u4e2a\u8bf7\u6c42\u90fd\u6765\u81ea\u4e00\u4e2a\u7531\u5747\u503c\u4e3a 128 \u7684\u6307\u6570\u5206\u5e03\u4ea7\u751f\u7684\u6837\u672c\u3002\u6211\u4eec\u4f7f\u7528\u6307\u6570\u5206\u5e03\uff0c\u56e0\u4e3a\u5b83\u662f\u5bf9\u50cf ChatGPT \u8fd9\u6837\u7684\u5e94\u7528\u7a0b\u5e8f\u4e2d\u53ef\u80fd\u9047\u5230\u7684\u751f\u6210\u957f\u5ea6\u7684\u4e00\u4e2a\u597d\u8fd1\u4f3c\u3002\u4e3a\u4e86\u6539\u53d8\u6bcf\u4e2a\u8fd0\u884c\u7684\u65b9\u5dee\uff0c\u6211\u4eec\u4ec5\u9009\u62e9\u5c0f\u4e8e\u6216\u7b49\u4e8e 32\u3001128\u3001512 \u548c 1536 \u7684\u6837\u672c\u3002\u603b\u8f93\u51fa\u5e8f\u5217\u957f\u5ea6\u6700\u591a\u4e3a 512+32=544\u3001512+128=640\u3001512+512=1024 \u548c 512+1536=2048\u3002<\/p>\n<p>\u7136\u540e\u6211\u4eec\u4f7f\u7528\u4e00\u4e2a\u7b80\u5355\u7684\u57fa\u51c6\u6d4b\u8bd5\u811a\u672c\u5411\u6211\u4eec\u7684\u6a21\u578b\u670d\u52a1\u5668\u63d0\u4ea4 HTTP \u8bf7\u6c42\u3002\u57fa\u51c6\u6d4b\u8bd5\u811a\u672c\u4ee5\u7a81\u53d1\u65b9\u5f0f\u63d0\u4ea4\u6240\u6709\u8bf7\u6c42\uff0c\u4f7f\u8ba1\u7b97\u9971\u548c\u3002<\/p>\n<p>\u7ed3\u679c\u5982\u4e0b\uff1a<\/p>\n<p><span class=\"gatsby-resp-image-wrapper\"><a class=\"gatsby-resp-image-link\" href=\"https:\/\/hfai-static.high-flyer.cn\/static\/e751e4793ee9c6b55675cbc11f5fe112\/302a4\/05.png\" target=\"_blank\" rel=\"noopener\"><img loading=\"lazy\" decoding=\"async\" class=\"alignnone size-full wp-image-407\" title=\"751d31dd6b56b26b29da82\" src=\"http:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/751d31dd6b56b26b29da82.png\" alt=\"751d31dd6b56b26b29da82\" width=\"650\" height=\"193\" srcset=\"https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/751d31dd6b56b26b29da82.png 650w, https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/751d31dd6b56b26b29da82-300x89.png 300w, https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/751d31dd6b56b26b29da82-150x45.png 150w\" sizes=\"(max-width: 650px) 100vw, 650px\" \/><\/a><\/span><\/p>\n<p>\u53ef\u4ee5\u770b\u5230\uff0c\u9759\u6001\u6279\u5904\u7406\u548c\u6734\u7d20\u8fde\u7eed\u6279\u5904\u7406\u5728\u8f83\u4f4e\u7684\u5e8f\u5217\u957f\u5ea6\u6ce2\u52a8\u4e0b\u8868\u73b0\u5927\u7ea6\u76f8\u540c\u3002\u7136\u800c\uff0c\u968f\u7740\u65b9\u5dee\u589e\u52a0\uff0c\u9759\u6001\u6279\u5904\u7406\u7684\u6027\u80fd\u663e\u8457\u4e0b\u964d\uff0c\u76f4\u5230\u751f\u6210\u957f\u5ea6\u8fbe\u5230 1536 \u65f6\uff0c\u6027\u80fd\u4e0b\u964d\u81f3 81 \u4e2a\u5b57\u7b26\/\u79d2\u3002FasterTransformers \u5728\u9759\u6001\u6279\u5904\u7406\u65b9\u9762\u7684\u8868\u73b0\u660e\u663e\u4f18\u4e8e\u9759\u6001\u6279\u5904\u7406\uff0c\u51e0\u4e4e\u4e0e\u8fde\u7eed\u6279\u5904\u7406\u4fdd\u6301\u540c\u6b65\uff0c\u76f4\u5230\u751f\u6210\u957f\u5ea6\u8fbe\u5230 1536\u3002<\/p>\n<p>\u4ee4\u4eba\u5370\u8c61\u6df1\u523b\u7684\u5728\u8fd9\u91cc\u7684\u662fvLLM\u3002\u5bf9\u4e8e\u6bcf\u4e2a\u6570\u636e\u96c6\uff0cvLLM\u7684\u6027\u80fd\u662f\u6734\u7d20\u8fde\u7eed\u6279\u5904\u7406\u7684\u4e24\u500d\u4ee5\u4e0a\u3002\u6211\u4eec\u5c1a\u672a\u5206\u6790vLLM \u6027\u80fd\u4e2d\u4f18\u5316\u8d21\u732e\u6700\u5927\u662f\u4ec0\u4e48\uff0c\u4f46\u6211\u5011\u6000\u7591 vLLM \u52a8\u6001\u9884\u7559\u7a7a\u95f4\u7684\u80fd\u529b\uff0c\u800c\u4e0d\u662f\u63d0\u524d\u9884\u7559\u7a7a\u95f4\uff0c\u4f7f\u5f97 vLLM \u53ef\u4ee5\u663e\u8457\u589e\u52a0\u6279\u91cf\u5927\u5c0f\u3002<\/p>\n<p>\u6211\u4eec\u5c06\u8fd9\u4e9b\u6027\u80fd\u7ed3\u679c\u76f8\u5bf9\u9759\u6001\u6279\u5904\u7406\u7684\u5173\u7cfb\u7ed8\u5236\u51fa\u6765\uff1a<\/p>\n<p><span class=\"gatsby-resp-image-wrapper\"><a class=\"gatsby-resp-image-link\" href=\"https:\/\/hfai-static.high-flyer.cn\/static\/b599416aeedb57e037e55a05ef1a931e\/302a4\/06.png\" target=\"_blank\" rel=\"noopener\"><img loading=\"lazy\" decoding=\"async\" class=\"alignnone size-full wp-image-412\" title=\"faeac4e1eef307c2ab7b68\" src=\"http:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/faeac4e1eef307c2ab7b68.png\" alt=\"faeac4e1eef307c2ab7b68\" width=\"650\" height=\"321\" srcset=\"https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/faeac4e1eef307c2ab7b68.png 650w, https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/faeac4e1eef307c2ab7b68-300x148.png 300w, https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/faeac4e1eef307c2ab7b68-150x74.png 150w\" sizes=\"(max-width: 650px) 100vw, 650px\" \/><\/a><\/span><\/p>\n<p>\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0cFasterTransformer \u76844\u500d\u6539\u8fdb\u867d\u7136\u4ee4\u4eba\u5370\u8c61\u6df1\u523b\uff0c\u4f46\u6211\u4eec\u975e\u5e38\u611f\u5174\u8da3\u5728 NVIDIA \u5b9e\u73b0\u5b83\u65f6\u8fdb\u884c\u57fa\u51c6\u6d4b\u8bd5\u3002\u7136\u800c\uff0c\u5373\u4f7f\u4f7f\u7528\u4f18\u5316\u540e\u7684\u6a21\u578b\uff0c\u8fde\u7eed\u6279\u5904\u7406\u5728\u9759\u6001\u6279\u5904\u7406\u65b9\u9762\u4ecd\u7136\u5177\u6709\u663e\u8457\u7684\u6539\u8fdb\u3002\u5982\u679c\u60a8\u50cf vLLM \u4e00\u6837\u91c7\u7528\u8fde\u7eed\u6279\u5904\u7406\u548c\u8fed\u4ee3\u7ea7\u8c03\u5ea6\u6765\u5b9e\u73b0\u8fdb\u4e00\u6b65\u7684\u5185\u5b58\u4f18\u5316\uff0c\u6027\u80fd\u5dee\u8ddd\u5c31\u4f1a\u53d8\u5f97\u975e\u5e38\u5de8\u5927\u3002<\/p>\n<h2 id=\"\u5ef6\u65f6\u6d4b\u8bd5\u7ed3\u679c\">\u5ef6\u65f6\u6d4b\u8bd5\u7ed3\u679c<\/h2>\n<p>\u7c7b\u4f3c\u5730\uff0c\u6211\u4eec\u5728\u4e00\u4e2a\u771f\u5b9e\u7684\u7528\u4f8b\u4e0a\u8fdb\u884c\u5ef6\u8fdf-\u541e\u5410\u91cf\u6743\u8861\u4f18\u5316\u3002\u6211\u4eec\u5bf9\u6bcf\u4e2a\u6846\u67b6\u8fdb\u884c\u57fa\u51c6\u6d4b\u8bd5\uff0c\u5e76\u6d4b\u91cf\u968f\u7740\u6bcf\u4e2a\u6846\u67b6\u7684\u7d2f\u79ef\u5206\u5e03\u51fd\u6570\uff08CDF\uff09\u7684\u53d8\u5316\uff0c\u8be5\u51fd\u6570\u5982\u4f55\u968f\u7740\u6bcf\u4e2a\u6846\u67b6\u7684\u90e8\u7f72\u800c\u53d8\u5316\u3002<\/p>\n<p>\u4e0e\u541e\u5410\u91cf\u57fa\u51c6\u6d4b\u8bd5\u76f8\u540c\uff0c\u6211\u4eec\u914d\u7f6e\u6a21\u578b\u4ee5\u59cb\u7ec8\u751f\u6210\u6307\u5b9a\u6570\u91cf\u7684\u6bcf\u4e2a\u8bf7\u6c42\u7684\u6807\u8bb0\u3002\u6211\u4eec\u901a\u8fc7\u4ece1\u4e2a\u6807\u8bb0\u5230512\u4e2a\u6807\u8bb0\u7684\u5747\u5300\u5206\u5e03\u4e2d\u91c7\u6837\u957f\u5ea6\u6765\u51c6\u5907100\u4e2a\u968f\u673a\u7684\u63d0\u793a\u3002\u6211\u4eec\u901a\u8fc7\u4ece\u5747\u503c\u4e3a128\u4e14\u6700\u5927\u5927\u5c0f\u4e3a1536\u7684\u6307\u6570\u5206\u5e03\u4e2d\u91c7\u6837100\u4e2a\u8f93\u51fa\u957f\u5ea6\u6765\u8fdb\u884c\u9884\u6d4b\u3002\u8fd9\u4e9b\u6570\u5b57\u4e4b\u6240\u4ee5\u9009\u62e9\uff0c\u662f\u56e0\u4e3a\u5b83\u4eec\u5728\u73b0\u5b9e\u4e16\u754c\u4e2d\u662f\u5408\u7406\u7684\uff0c\u5e76\u5141\u8bb8\u751f\u6210\u5668\u5145\u5206\u5229\u7528\u6211\u4eec\u6a21\u578b\u7684\u5168\u90e8\u4e0a\u4e0b\u6587\u957f\u5ea6\uff08512+1536=2048\uff09\u3002<\/p>\n<p>\u6211\u4eec\u6ca1\u6709\u50cf\u5728\u541e\u5410\u91cf\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u90a3\u6837\u540c\u65f6\u63d0\u4ea4\u6240\u6709\u8bf7\u6c42\u3002\u800c\u662f\u901a\u8fc7\u9884\u5148\u5ef6\u8fdf\u6bcf\u4e2a\u8bf7\u6c42\u4e00\u5b9a\u7684\u65f6\u95f4\u6765\u8fdb\u884c\u57fa\u51c6\u6d4b\u8bd5\u3002\u6211\u4eec\u901a\u8fc7\u6cca\u677e\u5206\u5e03\u6765\u786e\u5b9a\u6bcf\u4e2a\u8bf7\u6c42\u63d0\u4ea4\u540e\u7b49\u5f85\u591a\u957f\u65f6\u95f4\u3002\u6cca\u677e\u5206\u5e03\u7531\u03bb\u548c\u9884\u671f\u7387\uff08\u5373\u6bcf\u79d2\u51fb\u4e2d\u6a21\u578b\u7aef\u70b9\u7684\u67e5\u8be2\u6570\u91cf\uff09\u8fdb\u884c\u53c2\u6570\u5316\u3002\u6211\u4eec\u6d4b\u91cf\u5728 QPS \u4e3a1\u548c QPS \u4e3a4\u65f6\u7684\u5ef6\u8fdf\uff0c\u4ee5\u4e86\u89e3\u968f\u7740\u8d1f\u8f7d\u7684\u53d8\u5316\uff0c\u5ef6\u8fdf\u5206\u5e03\u5982\u4f55\u53d8\u5316\u3002<\/p>\n<p>\u7ed3\u679c\u5982\u4e0b\uff1a<\/p>\n<p><span class=\"gatsby-resp-image-wrapper\"><a class=\"gatsby-resp-image-link\" href=\"https:\/\/hfai-static.high-flyer.cn\/static\/98e00f7af77d8872cffafa3a9efb9b7c\/302a4\/07.png\" target=\"_blank\" rel=\"noopener\"><img loading=\"lazy\" decoding=\"async\" class=\"alignnone size-full wp-image-410\" title=\"d72d187df41e10ea7d9f28\" src=\"http:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/d72d187df41e10ea7d9f28.png\" alt=\"d72d187df41e10ea7d9f28\" width=\"650\" height=\"182\" srcset=\"https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/d72d187df41e10ea7d9f28.png 650w, https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/d72d187df41e10ea7d9f28-300x84.png 300w, https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/d72d187df41e10ea7d9f28-150x42.png 150w\" sizes=\"(max-width: 650px) 100vw, 650px\" \/><\/a><\/span><\/p>\n<p>\u53ef\u4ee5\u770b\u5230\uff0c\u5728\u63d0\u9ad8\u541e\u5410\u91cf\u7684\u540c\u65f6\uff0c\u8fde\u7eed\u6279\u5904\u7406\u7cfb\u7edf\u4e5f\u6539\u5584\u4e86\u4e2d\u4f4d\u6570\u5ef6\u8fdf\u3002\u8fd9\u662f\u56e0\u4e3a\u8fde\u7eed\u6279\u5904\u7406\u7cfb\u7edf\u5141\u8bb8\u5728\u6709\u7a7a\u95f4\u7684\u60c5\u51b5\u4e0b\uff0c\u5c06\u65b0\u7684\u8bf7\u6c42\u6dfb\u52a0\u5230\u73b0\u6709\u7684\u6279\u6b21\u4e2d\u3002\u4f46\u662f\uff0c\u5176\u4ed6\u767e\u5206\u4f4d\u6570\u5462\uff1f\u4e8b\u5b9e\u4e0a\uff0c\u6211\u4eec\u53d1\u73b0\u5b83\u4eec\u5728\u6240\u6709\u767e\u5206\u4f4d\u6570\u4e0a\u90fd\u6539\u5584\u4e86\u5ef6\u8fdf\uff1a<\/p>\n<p><span class=\"gatsby-resp-image-wrapper\"><a class=\"gatsby-resp-image-link\" href=\"https:\/\/hfai-static.high-flyer.cn\/static\/c5fee1ba3a27ddf2135cce9fe5c4cd54\/302a4\/08.png\" target=\"_blank\" rel=\"noopener\"><img loading=\"lazy\" decoding=\"async\" class=\"alignnone size-full wp-image-411\" title=\"fad6f4e614a212e80c6769\" src=\"http:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/fad6f4e614a212e80c6769.png\" alt=\"fad6f4e614a212e80c6769\" width=\"650\" height=\"311\" srcset=\"https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/fad6f4e614a212e80c6769.png 650w, https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/fad6f4e614a212e80c6769-300x144.png 300w, https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/fad6f4e614a212e80c6769-150x72.png 150w\" sizes=\"(max-width: 650px) 100vw, 650px\" \/><\/a><\/span><\/p>\n<p>QPS=1\u65f6\uff0c\u6bcf\u4e2a\u6846\u67b6\u7684\u751f\u6210\u8bf7\u6c42\u4e2d\u7d2f\u79ef\u5206\u5e03\u51fd\u6570\uff08CDF\uff09\u7684\u5f62\u72b6\u90fd\u6709\u6240\u4e0d\u540c\uff0c\u8fd9\u662f\u7531\u4e8e\u8fde\u7eed\u6279\u5904\u7406\u7cfb\u7edf\u4e2d\u7684\u8fed\u4ee3\u7ea7\u6279\u8c03\u5ea6\u5bfc\u81f4\u7684\u3002\u6240\u6709\u8fde\u7eed\u6279\u5904\u7406\u5728 QPS=1 \u65f6\u8868\u73b0\u7ea6\u7b49\u4e8e\u9759\u6001\u6279\u5904\u7406\uff1b\u4e0e\u9759\u6001\u6279\u5904\u7406\u76f8\u6bd4\uff0cFasterTransformers \u7684\u8868\u73b0\u660e\u663e\u66f4\u597d\u3002<\/p>\n<p>\u8fde\u7eed\u6279\u5904\u7406\u6539\u5584\u6240\u6709\u767e\u5206\u4f4d\u6570\u7684\u5ef6\u8fdf\u7684\u539f\u56e0\u4e0e\u6539\u5584 p50 \u7684\u5ef6\u8fdf\u539f\u56e0\u76f8\u540c\uff1a\u65b0\u7684\u8bf7\u6c42\u53ef\u4ee5\u6dfb\u52a0\uff0c\u800c\u4e0d\u7ba1\u5176\u4ed6\u5e8f\u5217\u5728\u751f\u6210\u6279\u6b21\u4e2d\u591a\u8fdc\u3002\u7136\u800c\uff0c\u4e0e\u9759\u6001\u6279\u5904\u7406\u4e00\u6837\uff0c\u8fde\u7eed\u6279\u5904\u7406\u4ecd\u7136\u53d7\u5230 GPU \u53ef\u7528\u7a7a\u95f4\u7684\u9650\u5236\u3002\u968f\u7740\u60a8\u7684\u670d\u52a1\u7cfb\u7edf\u63a5\u6536\u5230\u7684\u8bf7\u6c42\u6570\u91cf\u9971\u548c\uff0c\u5373\u5e73\u5747\u6279\u91cf\u5927\u5c0f\u66f4\u9ad8\uff0c\u90a3\u4e48\u5728\u6536\u5230\u65b0\u8bf7\u6c42\u65f6\uff0c\u7acb\u5373\u5728 GPU \u4e0a\u6ce8\u5165\u65b0\u8bf7\u6c42\u7684\u673a\u4f1a\u5c31\u8d8a\u5c11\u3002\u6211\u4eec\u53ef\u4ee5\u770b\u5230\uff0c\u968f\u7740\u5e73\u5747QPS\u589e\u52a0\u52304\uff0c\u8fd9\u4e00\u70b9\u53d8\u5f97\u66f4\u52a0\u660e\u663e\uff1a<\/p>\n<p><span class=\"gatsby-resp-image-wrapper\"><a class=\"gatsby-resp-image-link\" href=\"https:\/\/hfai-static.high-flyer.cn\/static\/ad42db3ac8e54cd2ebfc291571e21573\/302a4\/09.png\" target=\"_blank\" rel=\"noopener\"><img loading=\"lazy\" decoding=\"async\" class=\"alignnone size-full wp-image-413\" title=\"0a8005f5594bd67041f879\" src=\"http:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/0a8005f5594bd67041f879.png\" alt=\"0a8005f5594bd67041f879\" width=\"650\" height=\"311\" srcset=\"https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/0a8005f5594bd67041f879.png 650w, https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/0a8005f5594bd67041f879-300x144.png 300w, https:\/\/high-flyer.in.suopu.cc\/wp-content\/uploads\/2025\/11\/0a8005f5594bd67041f879-150x72.png 150w\" sizes=\"(max-width: 650px) 100vw, 650px\" \/><\/a><\/span><\/p>\n<p>QPS=4 \u65f6\uff0c\u6bcf\u4e2a\u6846\u67b6\u7684\u751f\u6210\u8bf7\u6c42\u4e2d\u7d2f\u79ef\u5206\u5e03\u51fd\u6570\uff08CDF\uff09\u7684\u5f62\u72b6\u4e0e QPS=1 \u65f6\u7684\u60c5\u51b5\u76f8\u4f3c\uff0cFasterTransformer \u7684\u5206\u5e03\u4e0e\u9759\u6001\u6279\u5904\u7406\u7684\u5206\u5e03\u76f8\u4f3c\u3002\u7136\u800c\uff0c\u4e0e QPS=1 \u76f8\u6bd4\uff0cFasterTransformer \u7684\u66f2\u7ebf\u66f4\u63a5\u8fd1\u4e8e\u9759\u6001\u6279\u5904\u7406\uff0c\u800c text-generation-inference \u548c Ray Serve \u7684\u8fde\u7eed\u6279\u5904\u7406\u5b9e\u73b0\u4e0e FasterTransformer \u7684\u66f2\u7ebf\u76f8\u4f3c\uff0c\u4f46\u660e\u663e\u4e0d\u5982 vLLM\u3002<\/p>\n<p>\u6211\u4eec\u89c2\u5bdf\u5230\uff0c\u968f\u7740\u7cfb\u7edf\u7684\u9971\u548c\uff0cFasterTransformer \u53d8\u5f97\u66f4\u52a0\u63a5\u8fd1\u4e8e\u9759\u6001\u6279\u5904\u7406\uff0c\u800c text-generation-inference \u548c Ray Serve \u7684\u8fde\u7eed\u6279\u5904\u7406\u5b9e\u73b0\u6b63\u5728\u671d\u7740\u4e0e FasterTransformer \u7684\u66f2\u7ebf\u76f8\u4f3c\u7684\u65b9\u5411\u53d1\u5c55\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u968f\u7740\u7cfb\u7edf\u7684\u9971\u548c\uff0c\u7acb\u5373\u5728 GPU \u4e0a\u6ce8\u5165\u65b0\u8bf7\u6c42\u7684\u673a\u4f1a\u66f4\u5c11\uff0c\u56e0\u6b64\u8bf7\u6c42\u5ef6\u8fdf\u4e0a\u5347\u3002\u8fd9\u4e0e vLLM \u7684\u66f2\u7ebf\u76f8\u543b\u5408\uff1a\u5728 QPS=1 \u548c QPS=4 \u4e4b\u95f4\uff0c\u5b83\u7684\u6700\u5927\u6279\u91cf\u5927\u7ea6\u589e\u52a0\u4e86 1500 \u4e2atoken\u3002<\/p>\n<h2 id=\"\u603b\u7ed3\">summarize<\/h2>\n<p>LLMs \u5177\u6709\u4ee4\u4eba\u60ca\u53f9\u7684\u529f\u80fd\uff0c\u6211\u4eec\u76f8\u4fe1\u5176\u5f71\u54cd\u4ecd\u6709\u5f88\u5927\u7684\u53d1\u6398\u7a7a\u95f4\u3002\u6211\u4eec\u5df2\u7ecf\u5206\u4eab\u4e86\u5982\u4f55\u4f7f\u7528\u4e00\u79cd\u65b0\u7684\u670d\u52a1\u6280\u672f\uff1a\u8fde\u7eed\u6279\u5904\u7406\uff0c\u4ee5\u53ca\u5b83\u5982\u4f55\u8d85\u8d8a\u9759\u6001\u6279\u5904\u7406\u3002\u5b83\u901a\u8fc7\u51cf\u5c11\u6d6a\u8d39\u5b89\u6392\u65b0\u8bf7\u6c42\u7684\u673a\u4f1a\u6765\u63d0\u9ad8\u541e\u5410\u91cf\uff0c\u5e76\u80fd\u591f\u7acb\u5373\u5c06\u65b0\u8bf7\u6c42\u6ce8\u5165\u8ba1\u7b97\u6d41\u4e2d\u6765\u6539\u5584\u5ef6\u8fdf\u3002\u6211\u4eec\u671f\u5f85\u4eba\u4eec\u5982\u4f55\u5229\u7528\u8fde\u7eed\u6279\u5904\u7406\uff0c\u4ee5\u53ca\u8fd9\u4e2a\u884c\u4e1a\u5c06\u4ece\u6b64\u8d70\u5411\u4f55\u65b9\u3002<\/p>","protected":false},"excerpt":{"rendered":"<p>\u7531\u4e8e LLM \u5de8\u5927\u7684 GPU \u5185\u5b58\u5f00\u9500\u548c\u8ba1\u7b97\u6210\u672c\uff0c\u5728\u5927\u591a\u6570\u5e94\u7528\u4e2d\uff0c\u673a\u5668\u5b66\u4e60\u5de5\u7a0b\u5e08\u901a\u5e38\u901a\u8fc7\u5185\u90e8\u8c03\u6574\uff08\u5982\u91cf\u5316\u548c\u5bf9  [&hellip;]<\/p>","protected":false},"author":1,"featured_media":414,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[1],"tags":[],"class_list":["post-404","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-basic-research"],"acf":[],"_links":{"self":[{"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/posts\/404","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/comments?post=404"}],"version-history":[{"count":1,"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/posts\/404\/revisions"}],"predecessor-version":[{"id":415,"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/posts\/404\/revisions\/415"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/media\/414"}],"wp:attachment":[{"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/media?parent=404"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/categories?post=404"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/high-flyer.in.suopu.cc\/en\/wp-json\/wp\/v2\/tags?post=404"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}