{"id":10932,"date":"2026-08-11T18:01:42","date_gmt":"2026-08-11T10:01:42","guid":{"rendered":"https:\/\/infernews.com\/blog\/oasiskv-llm-kv-cache-hbm\/"},"modified":"2026-08-11T18:06:50","modified_gmt":"2026-08-11T10:06:50","slug":"oasiskv-llm-kv-cache-hbm","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/oasiskv-llm-kv-cache-hbm\/","title":{"rendered":"OasisKV \u628a LLM KV Cache \u5e36\u51fa HBM \u74f6\u9838"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/08\/pasted-84ada6959205.jpg\" alt=\"Hugging Face\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">OasisKV \u4f86\u81ea Microsoft Research\u3001Imperial College London\u3001KAIST \u548c University of Edinburgh\u3002\u662f\u5efa\u57fa\u65bc vLLM \u7684 LLM \u63a8\u7406\u7cfb\u7d71\u8a2d\u8a08\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9577\u4e0a\u4e0b\u6587\u53ca\u9577\u7bc7\u63a8\u7406\u6703\u4ee4 Large Language Model (LLM) \u7684 key-value (KV) cache \u4f54\u7528\u5927\u91cf\u8a18\u61b6\u9ad4\u548c\u983b\u5bec\uff0cHigh Bandwidth Memory (HBM) \u5bb9\u91cf\u9042\u6210\u70ba\u6279\u6b21\u5927\u5c0f\u8207\u541e\u5410\u91cf\u7684\u9650\u5236\u3002OasisKV \u4e0d\u518d\u628a\u5b8c\u6574 KV cache \u5168\u90e8\u653e\u5728 HBM\uff0c\u800c\u662f\u5728\u89e3\u78bc\u968e\u6bb5\u53ea\u4fdd\u7559\u6ce8\u610f\u529b\u8f03\u9ad8\u7684 KV \u9805\u76ee\uff0c\u5176\u4ed6\u8cc7\u6599\u5b58\u653e\u65bc\u4e3b\u6a5f\u6216\u9060\u7aef\u8a18\u61b6\u9ad4\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7cfb\u7d71\u5229\u7528 speculative decoding (SD) \u7522\u751f\u7684 lookahead tokens\uff0c\u9810\u6e2c\u4e0b\u4e00\u6b65\u53ef\u80fd\u91cd\u8981\u7684 KV blocks\uff0c\u518d\u900f\u904e\u80cc\u666f attention pipeline \u9810\u5148\u8f09\u5165 HBM\u3002\u8ad6\u6587\u5831\u544a\u6307\uff0c\u5728 2,048-token KV \u9810\u7b97\u4e0b\uff0c\u6e96\u78ba\u5ea6\u53ea\u6bd4 full attention \u4f4e\u6700\u591a 0.7 \u5206\uff0c\u63a8\u7406\u5de5\u4f5c\u8ca0\u8f09\u541e\u5410\u91cf\u53ef\u9054 dense vLLM \u7684 1.69 \u500d\uff0c\u591a GPU \u9577\u4e0a\u4e0b\u6587\u670d\u52d9\u6700\u9ad8\u9054 2.1 \u500d\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u9810\u53d6\u7a00\u758f KV\uff0c\u964d\u4f4e HBM \u5bb9\u91cf\u58d3\u529b<\/li>\n<li>\u652f\u63f4 host \u6216 remote memory \u4f5c\u8f03\u5927\u5bb9\u91cf\u5c64\u7d1a<\/li>\n<li>prefill\u2013decode disaggregation \u4e0b\u541e\u5410\u91cf\u7d04\u70ba dense \u65b9\u6cd5 2 \u500d<\/li>\n<li>\u6bcf\u500b\u8acb\u6c42\u6240\u9700 KV \u53ef\u6e1b\u5c11 6.5 \u81f3 9.7 \u500d<\/li>\n<li>decode node host memory \u53ef\u964d\u4f4e 2.2 \u81f3 2.6 \u500d<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u53d6\u6368\u5728\u65bc\u7cfb\u7d71\u4f9d\u8cf4 lookahead \u9810\u6e2c\u53ca\u7a00\u758f attention\uff1b\u9810\u6e2c\u5931\u6e96\u53ef\u80fd\u5f71\u97ff\u6e96\u78ba\u5ea6\uff0c\u4e14\u8de8\u8a18\u61b6\u9ad4\u5c64\u7d1a\u9810\u53d6\u6703\u589e\u52a0\u7cfb\u7d71\u8907\u96dc\u5ea6\u3002\u66ab\u6642\u53ea\u78ba\u8a8d\u4ee5 vLLM \u5be6\u4f5c\uff0c\u6c92\u6709\u63d0\u4f9b GGUF \u6a94\u6848\u3001\u91cf\u5316\u7248\u672c\u3001mmproj \u9644\u52a0\u6a94\u6848\u3001Ollama \u6216 LM Studio \u652f\u63f4\u8cc7\u6599\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/huggingface.co\/papers\/2608.08097\" rel=\"noopener noreferrer\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/arxiv.org\/pdf\/2608.08097\" rel=\"noopener noreferrer\"><strong>Paper<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>HBM \u5bb9\u91cf\u5df2\u6210\u70ba\u4e00\u7a2e\u7a00\u7f3a\u4e14\u6602\u8cb4\u7684\u8cc7\u6e90\uff0c\u56b4\u91cd\u9650\u5236\u4e86\u63a8\u7406\u6279\u91cf\u5927\u5c0f\u548c\u7cfb\u7d71\u541e\u5410\u91cf\u3002OasisKV \u4ee5\u8a18\u61b6\u9ad4\u70ba\u4e2d\u5fc3\u9032\u884c\u63a8\u7406\uff0c\u900f\u904e\u89e3\u78bc\u671f\u9593\u5c07\u5b8c\u6574\u7684 KV \u5feb\u53d6\u5132\u5b58\u8207 HBM \u89e3\u8026\u4f86\u6e1b\u8f15 HBM \u5bb9\u91cf\u58d3\u529b\u3002<\/p>\n","protected":false},"author":8,"featured_media":10931,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[173,138],"tags":[],"class_list":["post-10932","post","type-post","status-publish","format-standard","hentry","category-173","category-138"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10932","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=10932"}],"version-history":[{"count":2,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10932\/revisions"}],"predecessor-version":[{"id":10936,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10932\/revisions\/10936"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/10931"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=10932"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=10932"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=10932"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}