{"id":12554,"date":"2026-09-23T16:41:12","date_gmt":"2026-09-23T08:41:12","guid":{"rendered":"https:\/\/infernews.com\/blog\/flash-dllm-io-aware-kv-caching-and-parallel-decoding\/"},"modified":"2026-09-23T16:41:12","modified_gmt":"2026-09-23T08:41:12","slug":"flash-dllm-io-aware-kv-caching-and-parallel-decoding","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/flash-dllm-io-aware-kv-caching-and-parallel-decoding\/","title":{"rendered":"Flash-dLLM \u8b93\u64f4\u6563\u5f0f\u8a9e\u8a00\u6a21\u578b\u89e3\u78bc\u66f4\u5feb\u3001\u66f4\u7701\u8a18\u61b6\u9ad4"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/09\/pasted-d17cf39bb2a3.jpg\" alt=\"Flash-dLLM overview: Flash-Cache selectively refreshes tracked tokens, while Flash-Verify checks draft predictions using\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u64f4\u6563\u5f0f\u8a9e\u8a00\u6a21\u578b\u6bcf\u8f2a\u53bb\u566a\u90fd\u53ef\u80fd\u540c\u6642\u7522\u751f\u591a\u500b token\uff0c\u4f46\u53cd\u8986\u66f4\u65b0 KV-cache \u6703\u4ee4 GPU \u82b1\u5927\u91cf\u6642\u9593\u8b80\u5beb\u8cc7\u6599\u3002Flash-dLLM \u5c6c\u65bc\u514d\u8a13\u7df4\u63a8\u7406\u52a0\u901f\u6846\u67b6\uff0c\u91dd\u5c0d\u9019\u500b\u8a18\u61b6\u9ad4\u74f6\u9838\u63d0\u5347\u89e3\u78bc\u901f\u5ea6\uff0c\u4ea6\u4fdd\u7559\u53ef\u8abf\u6821\u7684\u6e96\u78ba\u5ea6\u8207\u541e\u5410\u91cf\u53d6\u6368\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Flash-Cache \u4ee5\u878d\u5408\u5f0f Triton kernel \u4e00\u6b21\u8655\u7406 QKV projection\u3001rotary positional embedding\uff08RoPE\uff09\u53ca cache writes\uff0c\u4e26\u900f\u904e block-scheduled attention \u652f\u63f4\u6279\u6b21\u5167\u4e0d\u540c\u67e5\u8a62\u9577\u5ea6\u3002\u5b83\u53ea\u5237\u65b0\u65b0\u89e3\u78bc token \u53ca\u6309 attention importance \u9078\u51fa\u7684\u5c11\u91cf\u95dc\u9375\u4f4d\u7f6e\uff0c\u907f\u514d\u6bcf\u8f2a\u91cd\u7b97\u5168\u90e8\u5167\u5bb9\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Flash-Verify \u518d\u5229\u7528\u540c\u4e00\u500b diffusion model \u540c\u6642\u64d4\u4efb drafter \u548c verifier\uff0c\u6bcb\u9808\u984d\u5916\u6a21\u578b\u6216\u8a13\u7df4\uff1bdraft token \u8207 [MASK] token \u6703\u900f\u904e\u5169\u7a2e\u8996\u5716\u9a57\u8b49\uff0c\u7b26\u5408\u4fe1\u5fc3\u9580\u6abb\u7684\u9810\u6e2c\u53ef\u4e00\u4f75\u63a5\u53d7\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u2022 LLaDA-1.5 \u6e2c\u5f97 148\u2013211 tokens\/s<br \/>\n\u2022 \u6bd4\u7121 caching \u7684 greedy decoding \u5feb 22.3 \u81f3 148.2 \u500d<br \/>\n\u2022 GSM8K \u548c HumanEval \u6bd4 Elastic-Cache \u5feb 5.1 \u500d\u53ca 11.0 \u500d<br \/>\n\u2022 GPU \u8a18\u61b6\u9ad4\u7d04\u6bd4 Fast-dLLM \u5c11 48%\uff0c\u53ef\u64f4\u5c55\u81f3 batch size 32<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9019\u500b\u9805\u76ee\u9069\u5408\u7814\u7a76 diffusion LLM \u63a8\u7406\u3001GPU kernel \u6216\u9ad8\u541e\u5410\u91cf\u670d\u52d9\u7684\u5718\u968a\u3002\u4f7f\u7528\u524d\u9700\u6309\u5132\u5b58\u5eab\u63d0\u4f9b\u7684\u74b0\u5883\u53ca\u6a21\u578b\u8a2d\u5b9a\u9032\u884c\u5b89\u88dd\u548c evaluation\uff1b\u7d50\u679c\u6703\u53d7 GPU\u3001\u6279\u6b21\u5927\u5c0f\u3001\u4fe1\u5fc3\u9580\u6abb\u53ca\u6a21\u578b\u652f\u63f4\u7a0b\u5ea6\u5f71\u97ff\uff0c\u6578\u5b57\u672a\u5fc5\u80fd\u76f4\u63a5\u5957\u7528\u5230\u5176\u4ed6 dLLM\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/huggingface.co\/papers\/2609.26796\" rel=\"noopener noreferrer\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/VILA-Lab\/Flash-dLLM\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>Flash-dLLM \u628a KV-cache \u7ba1\u7406\u548c\u4e26\u884c\u89e3\u78bc\u653e\u5728\u540c\u4e00\u5957\u63a8\u7406\u6846\u67b6\u8655\u7406\uff0c\u6e1b\u5c11\u64f4\u6563\u5f0f\u8a9e\u8a00\u6a21\u578b\u53cd\u8986\u642c\u79fb\u8cc7\u6599\u9020\u6210\u7684\u901f\u5ea6\u640d\u5931\u3002<\/p>\n","protected":false},"author":8,"featured_media":12553,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"wpai_generated_summary":"","footnotes":""},"categories":[133,131,127,197,150],"tags":[],"class_list":["post-12554","post","type-post","status-publish","format-standard","hentry","category-133","category-embedding","category-127","category-framework","category-150"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/12554","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=12554"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/12554\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/12553"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=12554"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=12554"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=12554"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}