{"id":7961,"date":"2026-05-01T02:30:02","date_gmt":"2026-04-30T18:30:02","guid":{"rendered":"https:\/\/infernews.com\/blog\/?p=7961"},"modified":"2026-05-01T02:31:47","modified_gmt":"2026-04-30T18:31:47","slug":"7961-2","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/7961-2\/","title":{"rendered":"tLLM \u63a8\u8ad6\u5f15\u64ce\u64f4\u5c55"},"content":{"rendered":"\n<p><a href=\"https:\/\/github.com\/LinesHogan\/tLLM\">tLLM<\/a> \u662f vLLM \u63a8\u8ad6\u5f15\u64ce\u7684\u904b\u884c\u6642\u64f4\u5c55\u5c64\uff0c\u63d0\u4f9b\u751f\u7522\u8005\/\u6d88\u8cbb\u8005\uff08Producer\/Consumer\uff09\u67b6\u69cb\uff0c\u80fd\u5728\u751f\u7522\u74b0\u5883\u4e2d<strong>\u8a13\u7df4\u548c\u7ba1\u7406\u84b8\u993e\u5668<\/strong>\u3002<a href=\"https:\/\/github.com\/LinesHogan\/tLLM\" target=\"_blank\" rel=\"noreferrer noopener\"><\/a><\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"tllm\">tLLM \u7684\u89d2\u8272<\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u751f\u7522\u8005\u7ba1\u9053<\/strong>\uff1a\u5f9e vLLM \u63a8\u8ad6\u4e2d\u5373\u6642\u6355\u6349 LLM \u7684\u6df1\u5c64\u96b1\u85cf\u72c0\u614b\uff08latent representations\uff09<\/li>\n\n\n\n<li><strong>\u6d88\u8cbb\u8005\u7ba1\u9053<\/strong>\uff1a\u975e\u540c\u6b65\u8a13\u7df4\u8f15\u91cf MLP\uff08~1M \u53c3\u6578\uff09\uff0c<strong>\u9019\u5c31\u662f Latent Distiller\uff08\u6f5b\u5728\u84b8\u993e\u5668\uff09<\/strong><\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image size-full\"><img data-dominant-color=\"dfdfe0\" data-has-transparency=\"false\" style=\"--dominant-color: #dfdfe0;\" loading=\"lazy\" decoding=\"async\" width=\"712\" height=\"558\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/05\/tLLM.png\" alt=\"\" class=\"wp-image-7962 not-transparent\" srcset=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/05\/tLLM.png 712w, https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/05\/tLLM-300x235.png 300w\" sizes=\"auto, (max-width: 712px) 100vw, 712px\" \/><\/figure>\n\n\n\n<p><a href=\"https:\/\/github.com\/LinesHogan\/tLLM\">tLLM<\/a> \u53ef\u61c9\u7528\u65bc\u91ab\u7642\u554f\u7b54\u7cfb\u7d71\u4e2d\uff0c\u63d0\u5347 RAG \u751f\u6210\u7684\u591a\u6a23\u6027\u8207\u6e96\u78ba\u6027\uff0c\u7279\u5225\u9069\u5408\u914d\u5408 MedGemma \u5c08\u6848\u3002 \u900f\u904e ESamp (Exploratory Sampling\u00a0)\u65b9\u6cd5\uff0c\u5728\u9ad8\u541e\u5410 vLLM \u670d\u52d9\u4e0b\u52d5\u614b\u8a13\u7df4\u8f15\u91cf\u84b8\u993e\u5668\uff0c\u5f15\u5c0e\u6a21\u578b\u63a2\u7d22\u672a\u898b\u8a9e\u7fa9\u5340\u57df\uff0c\u907f\u514d\u91cd\u8907\u7b54\u6848\u3002<a href=\"https:\/\/arxiv.org\/html\/2604.24927v1\" target=\"_blank\" rel=\"noreferrer noopener\"><\/a><\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"rag\">\u6848\u4f8b\uff1a\u91ab\u7642 RAG \u7cfb\u7d71<\/h2>\n\n\n\n<p>\u5047\u8a2d\u60a8\u5efa\u7f6e\u4e00\u500b\u57fa\u65bc MedGemma \u7684\u7e41\u9ad4\u4e2d\u6587\u91ab\u7642 RAG \u7cfb\u7d71\uff0c\u8655\u7406\u60a3\u8005\u67e5\u8a62\u5982\u300c\u7cd6\u5c3f\u75c5\u4f75\u767c\u75c7\u9810\u9632\u300d\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6a19\u6e96 vLLM<\/strong>\uff1a\u4f9d\u8cf4\u6aa2\u7d22\u6587\u4ef6\u751f\u6210\u55ae\u4e00\u7b54\u6848\uff0c\u6613\u9677\u5165\u5e38\u898b\u6a21\u5f0f\uff0cPass@1 \u4f4e\u65bc 60%\u3002<\/li>\n\n\n\n<li><strong>tLLM + ESamp<\/strong>\uff1a\u555f\u7528\u751f\u7522\u8005\u7ba1\u9053\u6355\u6349\u96b1\u85cf\u8868\u793a\uff0c\u6d88\u8cbb\u8005\u8a13\u7df4 Latent Distiller\uff08MLP\uff0c~1M \u53c3\u6578\uff09\uff0c\u4f7f\u7528\u9810\u6e2c\u8aa4\u5dee\u4f5c\u70ba\u65b0\u7a4e\u5ea6\u4fe1\u865f\u3002<\/li>\n\n\n\n<li><strong>\u904b\u884c\u6d41\u7a0b<\/strong>\uff1a\u6279\u6b21 32 \u67e5\u8a62\u4e26\u884c\u751f\u6210\uff0c\u84b8\u993e\u5668\u7dda\u4e0a\u66f4\u65b0\uff08&lt;5% \u958b\u92b7\uff09\uff0c\u91cd\u65b0\u53d6\u6a23\u7522\u751f\u591a\u6a23\u5019\u9078\uff08\u5982\u751f\u6d3b\u8abf\u6574\u3001\u85e5\u7269\u3001\u98f2\u98df\u591a\u8996\u89d2\uff09\uff0cPass@k \u63d0\u5347 20-30%\u3002<a href=\"https:\/\/github.com\/LinesHogan\/tLLM\" target=\"_blank\" rel=\"noreferrer noopener\"><\/a><\/li>\n<\/ul>\n\n\n\n<p>\u6b64\u6848\u4f8b\u5728 RTX 4090 \u4e0a\u541e\u5410 4000+ tokens\/sec\uff0c\u9069\u5408\u90e8\u7f72\u65bc WhatsApp \u67e5\u8a62\u9598\u9053\uff0c\u63d0\u5347 Cantonese\/\u7e41\u4e2d\u91ab\u7642 NLP \u6548\u80fd\u3002 \u7a0b\u5f0f\u78bc\u7bc4\u4f8b\uff1a\u6574\u5408 tLLM \u81f3 vLLM Engine\uff0c<code>engine = TLlmEngine.from_engine_args(engine_args)<\/code>&nbsp;\u555f\u52d5\u751f\u7522\u8005\/\u6d88\u8cbb\u8005\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>tLLM \u662f vLLM \u63a8\u8ad6\u5f15\u64ce\u7684\u904b\u884c\u6642\u64f4\u5c55\u5c64\uff0c\u63d0\u4f9b\u751f\u7522\u8005\/\u6d88\u8cbb\u8005\uff08Producer\/Consumer\uff09\u67b6\u69cb\uff0c\u80fd\u5728\u751f\u7522\u74b0\u5883\u4e2d\u8a13\u7df4\u548c\u7ba1\u7406\u84b8\u993e\u5668\u3002 tLLM \u7684\u89d2\u8272 tLLM \u53ef\u61c9\u7528\u65bc\u91ab\u7642\u554f\u7b54\u7cfb\u7d71\u4e2d\uff0c\u63d0\u5347 RAG \u751f\u6210\u7684\u591a\u6a23\u6027\u8207\u6e96\u78ba\u6027\uff0c\u7279\u5225\u9069\u5408\u914d\u5408 MedGemma \u5c08\u6848\u3002 \u900f\u904e ESamp (Exploratory Sampling\u00a0)\u65b9\u6cd5\uff0c\u5728\u9ad8\u541e\u5410 vLLM \u670d\u52d9\u4e0b\u52d5\u614b\u8a13\u7df4\u8f15\u91cf\u84b8\u993e\u5668\uff0c\u5f15\u5c0e\u6a21\u578b\u63a2\u7d22\u672a\u898b\u8a9e\u7fa9\u5340\u57df\uff0c\u907f\u514d\u91cd\u8907\u7b54\u6848\u3002 \u6848\u4f8b\uff1a\u91ab\u7642 RAG \u7cfb\u7d71 \u5047\u8a2d\u60a8\u5efa\u7f6e\u4e00\u500b\u57fa\u65bc MedGemma \u7684\u7e41\u9ad4\u4e2d\u6587\u91ab\u7642 RAG \u7cfb\u7d71\uff0c\u8655\u7406\u60a3\u8005\u67e5\u8a62\u5982\u300c\u7cd6\u5c3f\u75c5\u4f75\u767c\u75c7\u9810\u9632\u300d\u3002 \u6b64\u6848\u4f8b\u5728 RTX 4090 \u4e0a\u541e\u5410 4000+ tokens\/sec\uff0c\u9069\u5408\u90e8\u7f72\u65bc WhatsApp \u67e5\u8a62\u9598\u9053\uff0c\u63d0\u5347 Cantonese\/\u7e41\u4e2d\u91ab\u7642 NLP \u6548\u80fd\u3002 \u7a0b\u5f0f\u78bc\u7bc4\u4f8b\uff1a\u6574\u5408 tLLM \u81f3 vLLM Engine\uff0cengine = TLlmEngine.from_engine_args(engine_args)&nbsp;\u555f\u52d5\u751f\u7522\u8005\/\u6d88\u8cbb\u8005\u3002<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"googlesitekit_rrm_CAowvqSiDA:productID":"","footnotes":""},"categories":[127,146,133],"tags":[],"class_list":["post-7961","post","type-post","status-publish","format-standard","hentry","category-127","category-146","category-133"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/7961","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\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=7961"}],"version-history":[{"count":2,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/7961\/revisions"}],"predecessor-version":[{"id":7964,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/7961\/revisions\/7964"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=7961"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=7961"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=7961"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}