{"id":10076,"date":"2026-07-10T02:00:47","date_gmt":"2026-07-09T18:00:47","guid":{"rendered":"https:\/\/infernews.com\/blog\/vllm\/"},"modified":"2026-07-10T02:00:47","modified_gmt":"2026-07-09T18:00:47","slug":"vllm","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/vllm\/","title":{"rendered":"vLLM \u65b0\u5f8c\u7aef\u8dd1\u51fa\u539f\u751f\u7d1a\u901f\u5ea6"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/07\/thumbnail.jpg\" alt=\"Og image\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u5361\u4f4d\u4e00\u76f4\u5728\u65bc\uff1a\u60f3\u7528 vLLM \u7684\u9ad8\u541e\u5410\u63a8\u7406\u80fd\u529b\uff0c\u904e\u53bb\u5f80\u5f80\u8981\u70ba\u500b\u5225\u6a21\u578b\u5beb\u6216\u7b49\u5c08\u7528\u5be6\u4f5c\u3002\u5462\u7bc7\u5167\u5bb9\u8b1b\u7684\u662f Hugging Face \u628a transformers \u76f4\u63a5\u4f5c\u70ba vLLM \u7684 modeling backend\uff0c\u800c\u4e14\u9801\u9762\u6c92\u6709\u63d0\u4f9b base model \u8cc7\u8a0a\uff0c\u56e0\u70ba\u5b83\u4e0d\u662f\u55ae\u4e00\u6a21\u578b\u9801\uff0c\u800c\u662f\u91dd\u5c0d\u63a8\u7406\u5f8c\u7aef\u6574\u5408\u7684\u6280\u8853\u66f4\u65b0\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u91cd\u9ede\u50f9\u503c\u5f88\u76f4\u63a5\uff1a\u6a21\u578b\u4f5c\u8005\u53ea\u8981\u5df2\u6709 transformers \u5be6\u4f5c\uff0c\u5c31\u6709\u6a5f\u6703\u4e0d\u7528\u518d\u984d\u5916\u79fb\u690d\u5230 vLLM\uff0c\u4e5f\u80fd\u62ff\u5230\u63a5\u8fd1\u539f\u751f\uff0c\u751a\u81f3\u66f4\u5feb\u7684\u63a8\u7406\u8868\u73fe\u3002\u5c0d LLM \u8207 VLM \u90fd\u6709\u610f\u7fa9\uff0c\u56e0\u70ba serving \u8a2d\u5b9a\u57fa\u672c\u4e0d\u8b8a\uff0c\u53ea\u662f\u52a0\u5165 <code>--model-impl transformers<\/code> \u65d7\u6a19\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6587\u4e2d\u5c55\u793a\u4e86\u4e09\u7d44 Qwen3 \u6e2c\u8a66\uff1aQwen3-4B \u55ae GPU\u3001Qwen3-32B \u4ee5 tensor parallelism \u8dd1 2 GPU\uff0c\u4ee5\u53ca Qwen3-235B-A22B-FP8 Mixture-of-Experts \u5728\u540c\u4e00\u500b 8\u00d7H100 \u7bc0\u9ede\u4e0a\u4ee5 data parallelism \u52a0 expert parallelism \u57f7\u884c\u3002\u7d50\u679c\u6307\u5411\u540c\u4e00\u4ef6\u4e8b\uff1atransformers backend \u7684 throughput \u5df2\u7d93\u8ffd\u5e73\u6216\u8d85\u904e vLLM \u624b\u5beb native implementation\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>transformers<\/code> \u5df2\u652f\u63f4 450+ architectures\uff0c\u89d2\u8272\u50cf\u53c3\u8003\u7d1a modeling library<\/li>\n<li><code>vLLM<\/code> \u7e7c\u7e8c\u8ca0\u8cac continuous batching\u3001custom attention kernels \u7b49\u9ad8\u6548\u63a8\u7406\u512a\u5316<\/li>\n<li>\u555f\u7528\u65b9\u5f0f\u5f88\u7c21\u55ae\uff1a\u5347\u7d1a <code>vllm<\/code>\uff0c\u4e26\u5728 serve \u6642\u52a0\u5165 <code>--model-impl transformers<\/code><\/li>\n<li>\u53ef\u8207 <code>--tensor-parallel-size<\/code>\u3001<code>--data-parallel-size<\/code>\u3001<code>--enable-expert-parallel<\/code> \u4e00\u8d77\u4f7f\u7528<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u53d6\u6368\u4ea6\u8981\u8b1b\u6e05\u695a\uff1a\u9801\u9762\u91cd\u9ede\u5728 backend \u6574\u5408\u8207\u6548\u80fd\u5c55\u793a\uff0c\u4e0d\u662f GGUF \u767c\u5e03\u9801\uff0c\u6240\u4ee5\u6c92\u6709\u63d0\u4f9b GGUF \u683c\u5f0f\u3001\u91cf\u5316\u7b49\u7d1a\u3001mmproj\u3001chat template\u3001MTP draft speculation \u6216 LM Studio\uff0fOllama\uff0fllama.cpp \u6a94\u6848\u8cc7\u8a0a\u3002\u786c\u9ad4\u9700\u6c42\u65b9\u9762\uff0c\u793a\u4f8b\u81f3\u5c11\u6db5\u84cb\u55ae GPU\u30012 GPU\uff0c\u540c\u57cb 8\u00d7H100 \u7bc0\u9ede\uff1b\u4e0d\u540c\u6a21\u578b\u662f\u5426\u90fd\u80fd\u8907\u88fd\u540c\u6a23\u589e\u76ca\uff0c\u4ecd\u8981\u8996\u67b6\u69cb\u8207\u90e8\u7f72\u74b0\u5883\u800c\u5b9a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/huggingface.co\/blog\/native-speed-vllm-transformers-backend\" rel=\"noopener noreferrer\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/huggingface\/blog\/blob\/main\/native-speed-vllm-transformers-backend.md\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u540c\u4e00\u4efd transformers \u6a21\u578b\u5be6\u4f5c\uff0c\u4f9d\u5bb6\u53ef\u4ee5\u76f4\u63a5\u98df\u5230 vLLM \u7684\u9ad8\u901f\u63a8\u7406\u3002\u5c0d\u90e8\u7f72\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u7684\u4eba\u800c\u8a00\uff0c\u5c11\u5497\u79fb\u690d\u5de5\u5e8f\uff0c\u541e\u5410\u91cf\u4ea6\u672a\u5fc5\u8f38\u539f\u751f\u5be6\u4f5c\u3002<\/p>\n","protected":false},"author":8,"featured_media":10075,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,185,122,152,150,197],"tags":[],"class_list":["post-10076","post","type-post","status-publish","format-standard","hentry","category-133","category-qwen","category-ollama","category-python","category-150","category-framework"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10076","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=10076"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10076\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/10075"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=10076"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=10076"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=10076"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}