{"id":13647,"date":"2026-10-20T22:58:39","date_gmt":"2026-10-20T14:58:39","guid":{"rendered":"https:\/\/infernews.com\/blog\/deepinfer\/"},"modified":"2026-10-20T23:01:02","modified_gmt":"2026-10-20T15:01:02","slug":"deepinfer","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/deepinfer\/","title":{"rendered":"DeepInfer \u958b\u653e DeepSeek \u96e2\u7dda\u63a8\u7406\u6838\u5fc3"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/10\/DeepInfer-DeepSeek.jpg\" alt=\"Prefill: non-cached input throughput vs. latency\"\/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u5c0d\u9700\u8981\u81ea\u884c\u5efa\u7acb\u9ad8\u6548\u6a21\u578b\u670d\u52d9\u7684\u5718\u968a\uff0cDeepInfer \u63d0\u4f9b\u4e86\u4e00\u689d\u66f4\u63a5\u8fd1 DeepSeek \u751f\u7522\u74b0\u5883\u7684\u6280\u8853\u8def\u5f91\u3002\u9019\u500b\u6846\u67b6\u8ca0\u8cac\u96e2\u7dda\u63a8\u7406\uff0c\u8655\u7406\u6a21\u578b\u57f7\u884c\u3001\u6279\u6b21\u6392\u7a0b\u8207\u8f38\u51fa\u751f\u6210\uff1bKV cache \u7cfb\u7d71\u3001Prefill\uff0fDecode \u5206\u79bb\u53ca\u8de8\u8cc7\u6599\u5e73\u884c\u526f\u672c\u8ca0\u8f09\u5e73\u8861\u5247\u9808\u53e6\u884c\u6574\u5408\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u652f\u63f4 DeepSeek-V4.1-Flash\uff0c\u6db5\u84cb\u6a21\u578b\u8207 sampler runtime\u3001FlashMLA\u3001DeepGEMM\u3001DeepEP \u7b49\u6700\u4f73\u5316 kernels\u3002<\/li>\n\n\n\n<li>\u4ee5 DSpark draft tokens \u914d\u5408\u4fe1\u5fc3\u5ea6\u6392\u7a0b\u9a57\u8b49\uff0c\u6e1b\u5c11\u89e3\u78bc\u968e\u6bb5\u7b49\u5f85\uff1b\u9810\u8a2d\u6700\u5927 draft \u9577\u5ea6\u70ba 5\u3002<\/li>\n\n\n\n<li>\u652f\u63f4 MoE EP\u3001EPLB\u3001Attention DP\u3001balanced CP\uff0c\u4ee5\u53ca CPU \u8207\u88dd\u7f6e\u57f7\u884c\u91cd\u758a\u7684\u975e\u540c\u6b65\u6279\u6b21\u6392\u7a0b\u3002<\/li>\n\n\n\n<li>\u63d0\u4f9b SWA bounded replay\u3001structured outputs\u3001batch-invariant inference \u8207 deterministic sampling\uff0c\u65b9\u4fbf\u63a7\u5236\u6210\u672c\u548c\u91cd\u73fe\u7d50\u679c\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b83\u8207\u4e00\u822c\u6a21\u578b\u6846\u67b6\u7684\u5206\u5225\uff0c\u4e0d\u5728\u65b0\u589e\u6a21\u578b\u80fd\u529b\uff0c\u800c\u5728\u65bc\u628a DeepSeek \u7dda\u4e0a\u670d\u52d9\u7d2f\u7a4d\u7684\u4f4e\u5ef6\u9072\u6a5f\u5236\u6574\u7406\u6210\u53ef\u7de8\u8b6f\u7684\u63a8\u7406\u6846\u67b6\uff0c\u4ee3\u50f9\u662f\u786c\u4ef6\u8207\u6574\u5408\u9580\u6abb\u8f03\u9ad8\u3002\u74b0\u5883\u8981\u6c42\u5305\u62ec Linux\u3001Python 3.12\u3001GCC 13 \u6216\u4ee5\u4e0a\uff1bNVIDIA \u8def\u7dda\u4f7f\u7528 CUDA 13 \u53ca SM100 GPU\uff0c\u53e6\u6709 Huawei Ascend 950 NPU \u7248\u672c\uff0c\u90e8\u5206\u529f\u80fd\u53ea\u652f\u63f4 NVIDIA\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5c0d\u541e\u5410\u91cf\u3001\u63a8\u7406\u6210\u672c\u548c\u7d50\u679c\u4e00\u81f4\u6027\u654f\u611f\u7684 AI \u5e73\u53f0\u5718\u968a\uff0cDeepInfer \u7684\u50f9\u503c\u8f03\u660e\u986f\u3002\u5b83\u63d0\u4f9b batch-invariant inference\u3001regex\u3001JSON schema \u8207 EBNF grammar \u7d04\u675f\uff0c\u4f46\u5b98\u65b9\u672a\u63d0\u4f9b\u541e\u5410\u91cf\u6216\u5ef6\u9072\u6578\u64da\uff1b\u516c\u958b\u5167\u5bb9\u4ea6\u672a\u6db5\u84cb\u5b8c\u6574\u90e8\u7f72\u6307\u5357\u8207\u7dda\u4e0a\u670d\u52d9\u5143\u4ef6\uff0c\u8f03\u9069\u5408\u5177\u5099 GPU\u3001NPU \u53ca\u5206\u6563\u5f0f\u7cfb\u7d71\u7d93\u9a57\u7684\u5718\u968a\u5f9e\u6e90\u78bc\u6574\u5408\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/github.com\/deepseek-ai\/DeepInfer\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>DeepInfer \u628a DeepSeek \u5927\u898f\u6a21\u670d\u52d9\u7684\u96e2\u7dda\u63a8\u7406\u6838\u5fc3\u958b\u6e90\uff0c\u6db5\u84cb\u9ad8\u6548 kernels\u3001\u4e26\u884c\u6392\u7a0b\u8207\u89e3\u78bc\u52a0\u901f\uff0c\u4f46\u672a\u5305\u62ec\u5b8c\u6574\u7dda\u4e0a\u670d\u52d9\u7d44\u4ef6\u3002<\/p>\n","protected":false},"author":8,"featured_media":13646,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"wpai_generated_summary":"","wpai_meta_description":"","footnotes":""},"categories":[133,182,76,179,147,197,28,168,152],"tags":[],"class_list":["post-13647","post","type-post","status-publish","format-standard","hentry","category-133","category-182","category-76","category-nvidia","category-deepseek","category-framework","category-services","category-linux","category-python"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/13647","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=13647"}],"version-history":[{"count":1,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/13647\/revisions"}],"predecessor-version":[{"id":13650,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/13647\/revisions\/13650"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/13646"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=13647"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=13647"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=13647"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}