{"id":10021539,"date":"2026-11-04T09:28:41","date_gmt":"2026-11-04T01:28:41","guid":{"rendered":"https:\/\/itmax.cn\/?p=10021539"},"modified":"2026-11-04T09:28:50","modified_gmt":"2026-11-04T01:28:50","slug":"lmcache-20261104","status":"publish","type":"post","link":"https:\/\/itmax.cn\/?p=10021539","title":{"rendered":"vLLM \u4e00\u91cd\u542f\u7f13\u5b58\u5c31\u5931\u5fc6\uff5cLMCache \u628a KV cache \u642c\u51fa\u8fdb\u7a0b\u5916\uff0c\u8fd9\u7b14\u8d26\u6211\u66ff\u4f60\u7b97\u8fc7"},"content":{"rendered":"<h1 id=\"vllm-lmcache-kv-cache\">vLLM \u4e00\u91cd\u542f\u7f13\u5b58\u5c31\u5931\u5fc6\uff5cLMCache \u628a KV cache \u642c\u51fa\u8fdb\u7a0b\u5916\uff0c\u8fd9\u7b14\u8d26\u6211\u66ff\u4f60\u7b97\u8fc7<\/h1>\n<figure style=\"margin:2em 0;text-align:center;\">\n<img decoding=\"async\" src=\"https:\/\/itmax.cn\/wp-content\/uploads\/2026\/11\/lmcache-cover.png\" alt=\"LMCache \u5c01\u9762\u56fe\uff1aGitHub 12,272 \u661f\uff0c\u5b98\u65b9\u57fa\u51c6 TTFT \u964d 75%\uff0c\u8fd1\u56db\u5468 311 \u6b21\u63d0\u4ea4\" style=\"max-width:100%;height:auto;border-radius:8px;\"><figcaption style=\"font-size:0.9em;color:#666;margin-top:0.5em;\">LMCache\uff1a\u7ed9\u5927\u6a21\u578b\u63a8\u7406\u52a0\u4e00\u5c42\u7684 <code>KV cache<\/code> \u7ba1\u7406\u5c42\uff0cApache-2.0 \u5f00\u6e90\uff0cvLLM \u751f\u6001\u4e4b\u5916\u8fd8\u652f\u6301 SGLang \u4e0e\u591a\u5bb6\u786c\u4ef6<\/figcaption><\/figure>\n<h2 id=\"gpu\">\u6211\u7684 GPU \u5728\u7ed9\u540c\u4e00\u6bb5\u8bdd\u53cd\u590d\u4e70\u5355<\/h2>\n<p>\u5148\u4ea4\u4ee3\u4e00\u4e2a\u6211\u81ea\u5df1\u8eab\u4e0a\u771f\u5b9e\u5b58\u5728\u7684\u573a\u666f\u3002\u4e4b\u524d\u642d\u77e5\u8bc6\u5e93\u95ee\u7b54\uff0c\u6211\u628a\u51e0\u4e07\u5b57\u7684\u5185\u90e8\u6587\u6863\u5207\u6210\u80cc\u666f\u6750\u6599\u585e\u8fdb system prompt\uff0c\u518d\u62fc\u4e0a\u7528\u6237\u95ee\u9898\u53d1\u7ed9\u63a8\u7406\u670d\u52a1\u3002\u90a3\u65f6\u5019\u6211\u8fd8\u6ca1\u610f\u8bc6\u5230\u95ee\u9898\u51fa\u5728\u54ea\uff0c\u53ea\u662f\u7eb3\u95f7\uff1a\u540c\u6837\u7684\u80cc\u666f\u6750\u6599\uff0c\u4e00\u5c0f\u65f6\u5185\u88ab\u51e0\u767e\u4e2a\u8bf7\u6c42\u5404\u7b97\u4e86\u4e00\u904d\uff0cGPU \u5229\u7528\u7387\u770b\u7740\u633a\u9971\u6ee1\uff0c\u771f\u6b63&#8221;\u751f\u6210&#8221;\u7684\u65f6\u95f4\u5374\u4e0d\u591a\u3002\u540e\u6765\u5f04\u660e\u767d prefill \u548c decode \u7684\u533a\u522b\u624d\u56de\u8fc7\u795e\u6765\u2014\u2014\u90a3\u4e9b\u91cd\u590d\u7684\u524d\u7f00\uff0c\u6bcf\u4e00\u904d\u90fd\u5728\u70e7\u94b1\u3002<\/p>\n<p>vLLM \u5176\u5b9e\u81ea\u5e26\u4e00\u62db\uff1aautomatic prefix caching\uff08APC\uff09\uff0c\u76f8\u540c\u524d\u7f00\u547d\u4e2d\u5c31\u76f4\u63a5\u590d\u7528\u663e\u5b58\u91cc\u7684 KV cache\u3002\u542c\u8d77\u6765\u591f\u7528\u4e86\uff1f\u6211\u7528\u4e0b\u6765\uff08\u51c6\u786e\u8bf4\uff0c\u662f\u6211\u76ef\u7740\u522b\u4eba\u7684\u8fd0\u7ef4\u8bb0\u5f55\u548c issue \u770b\u4e0b\u6765\uff09\u6709\u4e24\u4e2a\u786c\u4f24\u7ed5\u4e0d\u8fc7\u53bb\u3002\u7b2c\u4e00\uff0c\u8fd9\u4e2a\u7f13\u5b58\u4f4f\u5728 vLLM \u8fdb\u7a0b\u81ea\u5df1\u809a\u5b50\u91cc\uff0c\u8fdb\u7a0b\u4e00\u91cd\u542f\uff0c\u5168\u90e8\u5931\u5fc6\uff0c\u6240\u6709\u8bf7\u6c42\u4ece\u5934 prefill\uff0c\u9ad8\u5cf0\u671f\u91cd\u542f\u4e00\u6b21\u5c31\u662f\u4e00\u573a\u96ea\u5d29\u3002\u7b2c\u4e8c\uff0c\u5b83\u53ea\u8ba4&#8221;\u524d\u7f00&#8221;\uff0c\u4ece\u7b2c\u4e00\u4e2a token \u5f00\u59cb\u4e25\u683c\u5339\u914d\uff1b\u591a\u8f6e\u5bf9\u8bdd\u91cc\u6bcf\u8f6e\u8ffd\u52a0\u7684\u5185\u5bb9\u3001RAG \u91cc\u68c0\u7d22\u5757\u63d2\u5728\u4e2d\u95f4\u7684\u573a\u666f\uff0c\u5168\u90fd\u4e0d\u547d\u4e2d\u3002\u6211\u7ffb\u5230\u6709\u540c\u884c\u7b97\u8fc7\u4e00\u7b14\u8d26\uff1a\u4e00\u4e2a\u4e24\u4e07 token \u957f\u6587\u6863\u7684\u95ee\u7b54\u670d\u52a1\uff0c\u6587\u6863\u90e8\u5206\u7684 prefill \u8ba1\u7b97\u88ab\u6210\u5343\u4e0a\u4e07\u4e2a\u8bf7\u6c42\u91cd\u590d\u652f\u4ed8\u2014\u2014\u8fd9\u4e9b GPU \u65f6\u95f4\uff0c\u672c\u6765\u4e00\u5206\u90fd\u4e0d\u7528\u82b1\u3002<\/p>\n<p>LMCache \u5c31\u662f\u51b2\u7740\u8fd9\u4e24\u5200\u6765\u7684\u3002\u5b83\u628a KV cache \u4ece\u63a8\u7406\u8fdb\u7a0b\u91cc\u642c\u51fa\u6765\uff0c\u843d\u5230 CPU \u5185\u5b58\u3001\u672c\u5730 SSD\u3001Redis\u3001Mooncake\u3001S3 \u8fd9\u4e9b\u5c42\u7ea7\u91cc\uff0c\u505a\u6210\u53ef\u4ee5\u8de8\u8bf7\u6c42\u3001\u8de8\u5b9e\u4f8b\u3001\u751a\u81f3\u8de8\u5f15\u64ce\u590d\u7528\u7684\u8d44\u4ea7\u3002\u5f00\u6e90\u5728 GitHub \u4e0a\uff0cApache-2.0 \u534f\u8bae\uff0c2024 \u5e74 5 \u6708\u5efa\u4ed3\uff0c\u73b0\u5728\u5df2\u7ecf\u8fdb\u4e86 PyTorch Foundation \u751f\u6001\uff0cNVIDIA \u7684 Dynamo \u63a8\u7406\u6846\u67b6\u4e5f\u628a\u5b83\u5217\u4e3a\u96c6\u6210\u7684\u7f13\u5b58\u5c42\u3002\u6211\u7b2c\u4e00\u773c\u770b\u5230&#8221;KV cache \u7ba1\u7406\u5c42&#8221;\u8fd9\u4e2a\u5b9a\u4f4d\u65f6\u7684\u53cd\u5e94\u662f\uff1a\u8fd9\u4e0d\u5c31\u662f\u7ed9\u63a8\u7406\u670d\u52a1\u52a0\u4e86\u4e2a Redis \u5417\uff1f\u628a\u6587\u6863\u548c issue \u8bfb\u5b8c\u4e4b\u540e\u6211\u627f\u8ba4\uff0c\u4e8b\u60c5\u6ca1\u90a3\u4e48\u7b80\u5355\uff0c\u4f46\u5751\u4e5f\u6ca1\u90a3\u4e48\u5c11\u3002<\/p>\n<h2 id=\"_1\">\u5b83\u5e72\u7684\u4e8b\uff0c\u662f\u628a\u6d88\u8017\u54c1\u53d8\u6210\u8d44\u4ea7<\/h2>\n<p>\u6211\u628a\u5b83\u7684\u80fd\u529b\u62c6\u6210\u56db\u7ec4\uff0c\u987a\u4fbf\u8bf4\u8bf4\u54ea\u51e0\u7ec4\u662f\u771f\u672c\u4e8b\u3002<\/p>\n<p><strong>\u7b2c\u4e00\u7ec4\uff1a\u5206\u5c42\u5b58\u50a8\u52a0\u8de8\u8bf7\u6c42\u590d\u7528\uff0c\u8fd9\u662f\u547d\u6839\u5b50\u3002<\/strong> KV cache \u7b97\u5b8c\u4e4b\u540e\u4e0d\u518d\u53ea\u4f4f\u5728\u663e\u5b58\u91cc\uff0c\u800c\u662f\u6309 GPU \u663e\u5b58 \u2192 CPU \u5185\u5b58 \u2192 \u672c\u5730 SSD \u2192 \u8fdc\u7aef\u5b58\u50a8\uff08Redis\/Valkey\u3001Mooncake\u3001InfiniStore\u3001S3 \u517c\u5bb9\u5bf9\u8c61\u5b58\u50a8\uff09\u5206\u7ea7\u4e0b\u6c89\uff0c\u547d\u4e2d\u65f6\u518d\u88c5\u56de\u6765\u3002\u5b98\u65b9\u6587\u6863\u91cc\u6709\u4e2a\u5bf9\u4f53\u611f\u5f88\u6709\u7528\u7684\u6570\u5b57\uff1aQwen3-8B \u8fd9\u4e2a\u6a21\u578b\uff0c1GB \u7684 KV cache \u5927\u7ea6\u5bf9\u5e94 46,020 \u4e2a token\uff08\u8fd9\u662f\u5b98\u65b9 benchmark \u914d\u7f6e\u91cc\u7ed9\u51fa\u7684\u6362\u7b97\u503c\uff09\u3002\u4e5f\u5c31\u662f\u8bf4\u4e00\u4efd 2 \u4e07 token \u7684\u957f\u6587\u6863\u524d\u7f00\uff0cKV \u5927\u6982\u534a GB \u4e0a\u4e0b\uff0cCPU \u5185\u5b58\u8f7b\u677e\u88c5\u4e0b\u51e0\u5341\u4efd\u3002\u5b98\u65b9 benchmark \u9875\u7528 Qwen3-8B \u8dd1\u4e86\u4e00\u7ec4\u957f\u6587\u6863\u95ee\u7b54\u5bf9\u6bd4\uff1a46 \u4e2a\u8bf7\u6c42\uff0c\u88f8 vLLM \u5e73\u5747\u9996 token \u5ef6\u8fdf 757 \u6beb\u79d2\uff0c\u6302\u4e0a LMCache \u4e4b\u540e 185 \u6beb\u79d2\uff0c\u964d\u4e86 75%\uff1b\u6574\u4e2a\u57fa\u51c6\u7684\u8017\u65f6\u4ece 23.47 \u79d2\u7f29\u5230 13.79 \u79d2\u3002\u539f\u56e0\u6587\u6863\u4e5f\u8bb2\u5f97\u76f4\u767d\uff1a\u88f8 vLLM \u7684\u5de5\u4f5c\u96c6\u4e00\u65e6\u8d85\u51fa\u663e\u5b58\u91cc\u7684 KV \u7a7a\u95f4\uff0c\u7b2c\u4e8c\u8f6e\u5c31\u5f97\u5168\u90e8\u91cd\u7b97\u524d\u7f00\uff0c\u800c LMCache \u628a\u88ab\u6324\u51fa\u7684\u5757\u7559\u5728 CPU \u5185\u5b58\u91cc\u6309\u9700\u6062\u590d\u3002<\/p>\n<p><strong>\u7b2c\u4e8c\u7ec4\uff1a\u72ec\u7acb\u5b88\u62a4\u8fdb\u7a0b\uff0c\u5f15\u64ce\u5d29\u4e86\u7f13\u5b58\u4e0d\u966a\u846c\u3002<\/strong> \u8fd9\u5c31\u662f\u5b83\u4e3b\u63a8\u7684 MP\uff08multiprocess\uff09\u6a21\u5f0f\uff1aLMCache \u81ea\u5df1\u8dd1\u6210\u4e00\u4e2a\u72ec\u7acb server\uff0cvLLM \u901a\u8fc7 connector \u8ddf\u5b83\u63e1\u624b\u3002vLLM \u5d29\u6e83\u3001\u91cd\u542f\u3001\u6eda\u52a8\u66f4\u65b0\uff0c\u7f13\u5b58\u90fd\u5728 LMCache \u90a3\u8fb9\u6d3b\u7740\uff0c\u65b0\u5b9e\u4f8b\u8d77\u6765\u76f4\u63a5\u63a5\u7740\u7528\u3002\u6211\u5224\u65ad\u8fd9\u4e00\u7ec4\u662f\u771f\u6709\u4e1c\u897f\u7684\uff1a\u91cd\u542f\u5931\u5fc6\u662f\u6240\u6709\u8fdb\u7a0b\u5185\u7f13\u5b58\u65b9\u6848\u7684\u901a\u75c5\uff0c\u628a\u7f13\u5b58\u751f\u547d\u5468\u671f\u8ddf\u5f15\u64ce\u89e3\u8026\uff0c\u4ef7\u503c\u5728\u6eda\u52a8\u66f4\u65b0\u9891\u7e41\u7684\u751f\u4ea7\u73af\u5883\u91cc\u4f1a\u88ab\u53cd\u590d\u5151\u73b0\u3002\u4ee3\u4ef7\u5f53\u7136\u4e5f\u6709\uff0c\u540e\u9762\u5355\u72ec\u7b97\u3002<\/p>\n<p><strong>\u7b2c\u4e09\u7ec4\uff1a\u975e\u524d\u7f00\u590d\u7528 CacheBlend\uff0c\u6700\u4eae\u773c\u4f46\u8981\u51b7\u9759\u770b\u3002<\/strong> \u4f20\u7edf\u524d\u7f00\u7f13\u5b58\u8981\u6c42 token \u5e8f\u5217\u4ece\u5934\u4e25\u683c\u5339\u914d\uff0cCacheBlend \u5141\u8bb8\u590d\u7528 prompt \u4e2d\u95f4\u4efb\u610f\u4f4d\u7f6e\u7684\u7f13\u5b58\u5757\uff0c\u53ea\u5bf9\u62fc\u63a5\u5904\u5c11\u91cf\u91cd\u7b97\u6765\u6062\u590d\u8d28\u91cf\u3002\u8fd9\u5bf9 RAG\uff08\u68c0\u7d22\u5757\u6bcf\u6b21\u63d2\u7684\u4f4d\u7f6e\u90fd\u4e0d\u540c\uff09\u548c\u591a\u8f6e\u5bf9\u8bdd\uff08\u4e2d\u95f4\u5939\u7740\u52a8\u6001\u5de5\u5177\u7ed3\u679c\uff09\u662f\u7406\u8bba\u4e0a\u7684\u5927\u6740\u5668\u3002\u4f46\u6211\u8981\u6cfc\u7684\u51b7\u6c34\u662f\uff1a\u8fd9\u662f\u8bba\u6587\u8f6c\u5316\u6765\u7684\u7279\u6027\uff0c\u751f\u4ea7\u73af\u5883\u91cc\u5927\u89c4\u6a21\u9a8c\u8bc1\u7684\u516c\u5f00\u8bb0\u5f55\u8fd8\u4e0d\u591a\uff0c\u5b98\u65b9\u6587\u6863\u81ea\u5df1\u4e5f\u6ca1\u628a\u5b83\u5217\u4e3a\u9ed8\u8ba4\u8def\u5f84\u3002\u6211\u4f1a\u628a\u5b83\u5f53&#8221;\u503c\u5f97\u5173\u6ce8&#8221;\u800c\u4e0d\u662f&#8221;\u53ef\u4ee5\u4f9d\u8d56&#8221;\u3002<\/p>\n<p><strong>\u7b2c\u56db\u7ec4\uff1a\u53ef\u89c2\u6d4b\u6027\uff0c\u610f\u5916\u5730\u624e\u5b9e\u3002<\/strong> \u7f13\u5b58\u547d\u4e2d\u7387\uff08\u8bf7\u6c42\u7ea7\u548c token \u7ea7\uff09\u3001\u5404\u5c42\u8bfb\u5199\u8017\u65f6\u3001\u7f13\u5b58\u5bb9\u91cf\u3001\u6dd8\u6c70\u60c5\u51b5\uff0cPrometheus \u6307\u6807\u9f50\u5168\u3002\u7f13\u5b58\u7cfb\u7edf\u6ca1\u6709\u6307\u6807\u5c31\u662f\u76f2\u5f00\uff0c\u8fd9\u4e00\u70b9\u5b83\u6bd4\u5f88\u591a\u540c\u7c7b\u9879\u76ee\u505a\u5f97\u8bda\u5b9e\u3002<\/p>\n<p>\u8fd8\u6709\u4e00\u5757\u8981\u63d0\u4e00\u4e0b\u6ca1\u90a3\u4e48\u5149\u9c9c\uff1aSGLang \u7684\u96c6\u6210\u76ee\u524d\u53ea\u6709 MP \u6a21\u5f0f\uff0c\u5f97\u5355\u72ec\u8d77\u4e00\u4e2a <code>lmcache\/standalone<\/code> \u5bb9\u5668\u518d\u5f00 <code>--enable-lmcache<\/code>\uff1bprefill \u548c decode \u5206\u79bb\uff08PD disaggregation\uff09\u90a3\u6761\u8def\u8981 NIXL \u52a0 RDMA \u6216 NVLink \u7684\u4f20\u8f93\u73af\u5883\uff0c\u666e\u901a\u5355\u673a\u73a9\u5bb6\u78b0\u4e0d\u5230\u4e5f\u4e0d\u7528\u78b0\u3002\u6211\u7684\u5efa\u8bae\u662f\uff1a\u5148\u628a vLLM \u52a0 MP \u6a21\u5f0f\u8fd9\u4e00\u6761\u4e3b\u8def\u8d70\u901a\uff0c\u522b\u7684\u5f53\u5730\u56fe\u6807\u8bb0\u5c31\u597d\u3002<\/p>\n<h2 id=\"_2\">\u72ec\u7acb\u8fdb\u7a0b\u8fd9\u7b14\u8d26\uff0c\u6536\u76ca\u548c\u5229\u606f\u90fd\u8981\u8ba4<\/h2>\n<p>LMCache \u7684\u67b6\u6784\u9009\u62e9\uff0c\u6211\u8ba4\u4e3a\u4e00\u53e5\u8bdd\u5c31\u80fd\u6982\u62ec\uff1a\u628a KV cache \u4ece&#8221;\u8fdb\u7a0b\u5185\u7684\u4e34\u65f6\u72b6\u6001&#8221;\u5347\u7ea7\u6210&#8221;\u8fdb\u7a0b\u5916\u7684\u6301\u4e45\u8d44\u4ea7&#8221;\u3002\u806a\u660e\u5728\u54ea\u5df2\u7ecf\u8bf4\u8fc7\u4e86\uff0c\u73b0\u5728\u7b97\u5229\u606f\u3002<\/p>\n<p>\u7b2c\u4e00\u7b14\u5229\u606f\u662f\u4f20\u8f93\u548c\u5e8f\u5217\u5316\u672c\u8eab\u8981\u94b1\u3002KV cache \u4ece\u5f15\u64ce\u642c\u5230 LMCache \u518d\u642c\u56de\u6765\uff0c\u6bcf\u4e00\u8d9f\u90fd\u662f\u5b9e\u6253\u5b9e\u7684\u5e8f\u5217\u5316\u3001\u7f51\u7edc\u4f20\u8f93\u3001\u53cd\u5e8f\u5217\u5316\u3002\u4ed3\u5e93\u91cc\u6709\u4e2a\u6700\u8fd1\u5f00\u7684\u9ad8\u8d28\u91cf issue\uff08#5440\uff09\u628a\u8fd9\u4e2a\u4ee3\u4ef7\u91cf\u5316\u4e86\uff1a\u5728 PCIe \u4e3b\u673a\u4e0a\uff0cMP \u6a21\u5f0f\u4e0b\u6bcf\u4e2a\u4f20\u8f93\u5bf9\u8c61\u6709\u5927\u7ea6 0.55 \u79d2\u7684\u56fa\u5b9a\u5f00\u9500\uff0c\u7ebf\u4e0a\u7684\u4f20\u8f93\u65f6\u95f4\u53cd\u800c\u88ab\u8fd9\u4e2a\u56fa\u5b9a\u6210\u672c\u6df9\u6ca1\u2014\u2014\u6ce8\u610f\u8fd9\u662f\u4ed6\u4eec\u5728\u7528 pickle \u4f20\u8f93\u7684\u573a\u666f\u4e0b\u6d4b\u7684\uff0c\u5b98\u65b9\u5df2\u7ecf\u8ba4\u4e86\u8fd9\u4e2a\u95ee\u9898\u5728\u8ddf\u8fdb\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u7f13\u5b58\u547d\u4e2d\u7701\u4e0b\u7684 prefill \u65f6\u95f4\uff0c\u53ef\u80fd\u88ab\u642c\u8fd0\u7684\u56fa\u5b9a\u5f00\u9500\u5403\u6389\u4e00\u5927\u5757\uff0c\u77ed prompt \u573a\u666f\u751a\u81f3\u5012\u8d34\u3002<\/p>\n<p>\u7b2c\u4e8c\u7b14\u5229\u606f\u662f\u517c\u5bb9\u6027\u7684\u590d\u6742\u5ea6\u3002\u5b98\u65b9\u4e13\u95e8\u5199\u4e86\u4e00\u6574\u9875\u517c\u5bb9\u6027\u6587\u6863\uff0c\u628a&#8221;\u80fd\u4e0d\u80fd\u4e00\u8d77\u8dd1&#8221;\u62c6\u6210\u4e09\u5c42\uff1aPython\/PyTorch\/CUDA \u7684 ABI \u8981\u5bf9\u4e0a\uff1bvLLM \u7248\u672c\u8981\u80fd\u52a0\u8f7d\u4f60\u8981\u7684\u90a3\u4e2a connector\uff08vLLM 0.20.0 \u4e4b\u524d\u53ea\u80fd\u7528\u5b83\u5185\u7f6e\u7684\u5b9e\u73b0\uff0c\u4e4b\u540e\u624d\u80fd\u663e\u5f0f\u6307\u5b9a LMCache \u81ea\u5e26\u7684\u6a21\u5757\u8def\u5f84\uff09\uff1b\u5177\u4f53\u6a21\u578b\u8fd8\u8981\u770b\u6709\u6ca1\u6709\u5bf9\u5e94\u7684\u9a8c\u8bc1\u8bb0\u5f55\u2014\u2014\u6587\u6863\u91cc\u5199\u660e GLM-5.2 \u662f\u5728 vLLM 0.23.0 \u52a0 LMCache 0.4.7 \u8fd9\u4e2a\u7cbe\u786e\u7ec4\u5408\u4e0b\u9a8c\u8bc1\u7684\uff0cKimi K3 \u8981 0.5.3 \u4ee5\u4e0a\u3002\u8fd9\u5957\u73b0\u5b9e\u610f\u5473\u7740&#8221;\u88c5\u4e2a\u6700\u65b0\u7248\u5c31\u5b8c\u4e8b\u4e86&#8221;\u5728\u8fd9\u91cc\u884c\u4e0d\u901a\uff0c\u5347\u7ea7\u524d\u8981\u5bf9\u7740\u77e9\u9635\u6838\u4e00\u904d\u3002\u6211\u4e0d\u592a\u559c\u6b22\u8fd9\u79cd\u590d\u6742\u5ea6\uff0c\u4f46\u8f6c\u5ff5\u4e00\u60f3\uff0c\u4e00\u4e2a\u8981\u78b0\u5e95\u5c42\u663e\u5b58\u5e03\u5c40\u548c\u539f\u751f\u6269\u5c55\u7684\u7f13\u5b58\u5c42\uff0c\u517c\u5bb9\u6027\u95ee\u9898\u4e0d\u662f\u5b83\u80fd\u9009\u62e9\u7684\uff0c\u662f\u5b83\u6ce8\u5b9a\u8981\u80cc\u7684\u3002<\/p>\n<p>\u6211\u7684\u53cd\u5bf9\u610f\u89c1\u4e5f\u7ed9\u5355\u5b9e\u4f8b\u73a9\u5bb6\uff1a\u5982\u679c\u4f60\u7684\u670d\u52a1\u53ea\u6709\u4e00\u4e2a vLLM \u5b9e\u4f8b\u3001\u51e0\u4e4e\u4e0d\u91cd\u542f\u3001prompt \u53c8\u5168\u662f\u77ed\u6587\u672c\uff0cvLLM \u81ea\u5e26\u7684 APC \u5c31\u591f\u4e86\uff0c\u5f15\u5165 LMCache \u7eaf\u5c5e\u7ed9\u81ea\u5df1\u52a0\u8fd0\u7ef4\u9762\u3002\u8fd9\u4e1c\u897f\u7684\u4ef7\u503c\u5bc6\u5ea6\u8ddf&#8221;\u91cd\u542f\u9891\u7387 \u00d7 \u591a\u5b9e\u4f8b\u6570 \u00d7 prompt \u91cd\u590d\u5ea6&#8221;\u6210\u6b63\u6bd4\uff0c\u4e09\u4e2a\u90fd\u4f4e\u5c31\u522b\u6298\u817e\u3002<\/p>\n<h2 id=\"_3\">\u5b89\u88c5\u8fd9\u4e00\u5173\uff0c\u5b98\u65b9\u6587\u6863\u5199\u5f97\u5f88\u8001\u5b9e<\/h2>\n<p>\u5148\u628a\u4e11\u8bdd\u8bf4\u5728\u524d\u9762\uff1a\u5b98\u65b9\u6ca1\u6709\u7ed9\u51fa\u6574\u673a\u7ea7\u522b\u7684\u5185\u5b58\u6216\u78c1\u76d8\u95e8\u69db\u8868\uff0c\u4e0b\u9762\u4e24\u6863\u91cc\u6807&#8221;\u5b98\u65b9&#8221;\u7684\u6211\u90fd\u6ce8\u660e\u4e86\u51fa\u5904\uff0c\u6807&#8221;\u6211\u6309\u6587\u6863\u548c issue \u63a8\u7684&#8221;\u5c31\u662f\u63a8\u7b97\uff0c\u522b\u5f53\u6210\u5b98\u65b9\u53e3\u5f84\u3002<\/p>\n<p>\u5b98\u65b9\u660e\u786e\u5199\u7684\u8981\u6c42\uff08\u6765\u81ea\u5b89\u88c5\u6587\u6863\u539f\u6587\uff09\uff1aLinux\uff0cPython 3.10 \u5230 3.14\uff0cNVIDIA GPU \u7b97\u529b 7.0 \u8d77\uff08\u4e5f\u5c31\u662f Volta \u5f80\u540e\uff09\uff0cCUDA 12.1 \u4ee5\u4e0a\uff1bAMD ROCm\uff08MI300X\/MI350X \u7cfb\u5217\uff09\u3001Intel XPU\u3001\u6469\u5c14\u7ebf\u7a0b MUSA \u8fd9\u4e9b\u975e N \u5361\u901a\u9053\u4e5f\u5404\u6709\u53d1\u5e03\u7248\uff0c\u6628\u5929\uff0811 \u6708 3 \u65e5\uff09\u8fd8\u5728\u53d1 v0.5.7rc4 \u7684 xpu\u3001rocm\u3001musa\u3001rbln \u591a\u8def\u6784\u5efa\u3002<\/p>\n<p>\u6700\u4f4e\u53ef\u7528\u6863\uff08\u6211\u7684\u63a8\u7b97\uff09\uff1a\u5355\u5f20 24GB \u663e\u5b58\u7684\u5361\uff084090\u30013090 \u8fd9\u4e00\u7ea7\uff09\u52a0 64GB \u7cfb\u7edf\u5185\u5b58\uff0c\u591f\u8dd1 Qwen3-8B \u8fd9\u7c7b 8B \u6a21\u578b\u52a0\u4e00\u4e2a\u8fdb\u7a0b\u5185\uff08in-process\uff09\u6a21\u5f0f\u7684 LMCache\uff0cCPU \u4fa7\u7559 20GB \u7ed9\u7f13\u5b58\u5c31\u6709\u6548\u679c\u3002\u5efa\u8bae\u6863\uff08\u6211\u7684\u63a8\u7b97\uff0c\u53c2\u8003\u5b98\u65b9 MP \u793a\u4f8b\uff09\uff1a\u751f\u4ea7\u4e0a\u8d70 MP \u6a21\u5f0f\uff0c\u5b98\u65b9 benchmark \u793a\u4f8b\u91cc L1 \u7f13\u5b58\u76f4\u63a5\u7ed9\u4e86 66GB\uff0c\u518d\u52a0\u4e0a\u6a21\u578b\u672c\u4f53\u548c vLLM \u7684\u5f00\u9500\uff0c128GB \u5185\u5b58\u3001\u4e00\u5f20 A100\/H100 \u6216\u53cc 4090 \u662f\u6bd4\u8f83\u8212\u670d\u7684\u4f4d\u7f6e\uff1b\u78c1\u76d8\u5c42\u8981\u5f00\u7684\u8bdd\uff0c\u6309\u4e0a\u9762 46,020 token\/GB \u7684\u6362\u7b97\u81ea\u5df1\u4f30\u5bb9\u91cf\u3002<\/p>\n<p>\u5b89\u88c5\u672c\u8eab\u5012\u662f\u4e1d\u6ed1\uff0c\u5b98\u65b9\u63a8\u8350\u7528 uv\uff1a<\/p>\n<pre><code class=\"language-bash\">uv venv --python 3.12\nsource .venv\/bin\/activate\nuv pip install lmcache vllm\n<\/code><\/pre>\n<p>\u9a8c\u8bc1\u88c5\u6ca1\u88c5\u5bf9\uff0c\u5b98\u65b9\u7ed9\u4e86\u4e00\u884c\uff1a<\/p>\n<pre><code class=\"language-bash\">python -c &quot;import lmcache.cuda_ops&quot;\n<\/code><\/pre>\n<p>\u4e0d\u60f3\u78b0 CUDA \u7684\u8fd8\u6709\u4e2a\u53d6\u5de7\u8def\u7ebf\uff1a<code>pip install lmcache-cli<\/code>\uff0c\u53ea\u88c5\u547d\u4ee4\u884c\u5de5\u5177\uff08\u8ddf\u4e3b\u5305\u662f\u540c\u4e00\u4e2a <code>lmcache<\/code> \u547d\u4ee4\uff0c\u5b98\u65b9\u63d0\u9192\u522b\u4e24\u4e2a\u4e00\u8d77\u88c5\uff09\uff0c\u53ef\u4ee5\u5148\u628a server \u548c benchmark \u5de5\u5177\u6478\u719f\u3002\u8981\u6211\u63a8\u8350\u7684\u8bdd\uff0c\u6700\u7701\u5fc3\u7684\u662f\u76f4\u63a5\u4e0a\u5b98\u65b9 Docker \u955c\u50cf\uff0cvLLM \u548c LMCache \u7684\u7248\u672c\u5728\u91cc\u9762\u662f\u914d\u597d\u5bf9\u7684\uff1a<\/p>\n<pre><code class=\"language-bash\">docker pull lmcache\/vllm-openai\n<\/code><\/pre>\n<p>\u955c\u50cf\u9ed8\u8ba4\u7528\u975e root \u7684 <code>lmcache<\/code> \u7528\u6237\u8dd1\uff0cHF \u6a21\u578b\u7f13\u5b58\u8981\u6302\u5230 <code>\/home\/lmcache\/.cache\/huggingface<\/code>\uff0c\u8fd9\u4e2a\u7ec6\u8282\u5b98\u65b9\u6587\u6863\u5199\u4e86\u4f46\u5f88\u5bb9\u6613\u6f0f\uff0c\u6f0f\u4e86\u7684\u75c7\u72b6\u5c31\u662f\u5bb9\u5668\u91cc\u4e0b\u4e0d\u52a8\u6a21\u578b\u3002<\/p>\n<h2 id=\"_4\">\u4e0a\u624b\u8dd1\u4e00\u904d\uff1a\u5b98\u65b9\u8fd9\u6761\u6700\u5c0f\u94fe\u8def\uff0c\u6211\u9010\u884c\u5bf9\u8fc7<\/h2>\n<p>\u5148\u8bda\u5b9e\u4ea4\u4ee3\uff1a\u6211\u624b\u4e0a\u6ca1\u6709\u80fd\u8dd1 CUDA \u7684\u673a\u5668\uff0c\u8fd9\u4e00\u8282\u6211\u6ca1\u6cd5\u5199\u771f\u8dd1\u7684\u8f93\u51fa\uff0c\u4e0b\u9762\u7684\u547d\u4ee4\u548c\u65e5\u5fd7\u5168\u90e8\u9010\u6761\u6765\u81ea\u5b98\u65b9 quickstart \u6587\u6863\uff0c\u6211\u6838\u5bf9\u8fc7\u7684\u662f\u547d\u4ee4\u4e4b\u95f4\u7684\u4f9d\u8d56\u5173\u7cfb\u548c\u53c2\u6570\u542b\u4e49\u3002\u60f3\u6284\u4f5c\u4e1a\u7684\u53ef\u4ee5\u76f4\u63a5\u7167\u8d70\u3002<\/p>\n<p>\u6700\u5feb\u89c1\u6548\u7684\u662f\u8fdb\u7a0b\u5185\u6a21\u5f0f\uff0c\u4e00\u6761\u547d\u4ee4\u8d77\u670d\u52a1\uff1a<\/p>\n<pre><code class=\"language-bash\">LMCACHE_CHUNK_SIZE=8 \\\nvllm serve Qwen\/Qwen3-8B \\\n    --port 8000 --kv-transfer-config \\\n    '{&quot;kv_connector&quot;:&quot;LMCacheConnectorV1&quot;, &quot;kv_role&quot;:&quot;kv_both&quot;}'\n<\/code><\/pre>\n<p>\uff08chunk size \u8bbe\u6210 8 \u53ea\u662f\u5b98\u65b9\u4e3a\u4e86\u7528\u77ed prompt \u4e5f\u80fd\u770b\u5230\u7f13\u5b58\u52a8\u9759\uff0c\u751f\u4ea7\u4fdd\u6301\u9ed8\u8ba4 256\u3002\uff09<\/p>\n<p>\u7136\u540e\u7528\u540c\u4e00\u4e2a\u524d\u7f00\u8fde\u6253\u4e24\u6b21\uff1a<\/p>\n<pre><code class=\"language-bash\">curl http:\/\/localhost:8000\/v1\/completions \\\n  -H &quot;Content-Type: application\/json&quot; \\\n  -d '{\n    &quot;model&quot;: &quot;Qwen\/Qwen3-8B&quot;,\n    &quot;prompt&quot;: &quot;Qwen3 is the latest generation of large language models in Qwen series, offering a comprehensive suite of dense and mixture-of-experts&quot;,\n    &quot;max_tokens&quot;: 100,\n    &quot;temperature&quot;: 0.7\n  }'\n<\/code><\/pre>\n<p>\u5b98\u65b9\u6587\u6863\u7ed9\u51fa\u7684\u9884\u671f\u65e5\u5fd7\u662f\u8fd9\u6837\u7684\uff1a\u7b2c\u4e00\u6b21\u8bf7\u6c42\u7f13\u5b58\u662f\u7a7a\u7684\uff0c\u9010\u5757\u5199\u5165\uff08<code>Stored 16 tokens in 0.023 seconds<\/code> \u8fd9\u4e00\u7c7b\u7684\u8f93\u51fa\uff09\uff1b\u7b2c\u4e8c\u6b21\u8bf7\u6c42\u5c31\u80fd\u770b\u5230\u547d\u4e2d\u2014\u201432 \u4e2a token \u7684 prompt \u91cc\u547d\u4e2d 24 \u4e2a\uff08\u5176\u4e2d 16 \u4e2a\u6765\u81ea vLLM \u81ea\u5df1\u7684 GPU \u524d\u7f00\u7f13\u5b58\uff0c8 \u4e2a\u7531 LMCache \u4ece CPU \u5185\u5b58\u8865\u4e0a\uff09\uff0c\u65e5\u5fd7\u91cc\u4f1a\u6709 <code>Retrieved 8 out of 24 required tokens ... cost 0.55 ms, throughput: 1.98 GB\/s<\/code> \u8fd9\u6837\u7684\u884c\u3002\u770b\u5230\u8fd9\u4e2a\u8f93\u51fa\uff0c\u94fe\u8def\u5c31\u7b97\u901a\u4e86\u3002<\/p>\n<p>\u8981\u8d70 MP \u6a21\u5f0f\u7684\u8bdd\u591a\u4e00\u6b65\uff0c\u5148\u8d77\u72ec\u7acb server\uff1a<\/p>\n<pre><code class=\"language-bash\">lmcache server \\\n    --host localhost --port 5555 \\\n    --l1-size-gb 20 --eviction-policy LRU --chunk-size 16\n<\/code><\/pre>\n<p>\u518d\u628a vLLM \u7684 connector \u6362\u6210 <code>LMCacheMPConnector<\/code> \u5e76\u6307\u5230 5555 \u7aef\u53e3\u3002\u4e00\u4e2a\u5bb9\u6613\u7ffb\u8f66\u7684\u7ec6\u8282\uff1avLLM 0.20.0 \u4ee5\u4e0b\u52a0\u8f7d\u4e0d\u5230 LMCache \u81ea\u5e26\u7684 connector \u5b9e\u73b0\uff0c\u7248\u672c\u5bf9\u4e0d\u4e0a\u65f6\u8981\u4e48\u5347\u7ea7 vLLM\uff0c\u8981\u4e48\u663e\u5f0f\u52a0 <code>kv_connector_module_path<\/code> \u6307\u5230 <code>lmcache.integration.vllm.lmcache_mp_connector<\/code>\u2014\u2014\u5b98\u65b9\u517c\u5bb9\u6027\u9875\u7684\u539f\u8bdd\u662f\uff0c\u4e0d\u52a0\u8fd9\u4e2a\u8def\u5f84\uff0cvLLM \u53ef\u80fd\u9759\u9ed8\u9009\u4e2d\u4e00\u4e2a\u534f\u8bae\u66f4\u65e7\u7684\u5185\u7f6e connector\uff0c\u7136\u540e\u4f60\u5c31\u5f00\u59cb\u6392\u67e5\u4e3a\u4ec0\u4e48\u529f\u80fd\u5bf9\u4e0d\u4e0a\u3002<\/p>\n<h2 id=\"_5\">\u522b\u4eba\u8e29\u8fc7\u7684\u5751\uff0c\u6211\u4e00\u6761\u6761\u5bf9\u8fc7<\/h2>\n<p>issue \u6c60\u5b50\u4e0d\u5c0f\uff08\u5f00\u653e\u4e2d\u7684\u6709 909 \u4e2a\uff09\uff0c\u6211\u628a\u9ad8\u9891\u7684\u548c\u8fd1\u671f\u7684\u7528\u6237\u62a5\u9519\u7c7b\u90fd\u8fc7\u4e86\u4e00\u904d\uff0c\u518d\u4ea4\u53c9\u641c\u4e86\u7b2c\u4e09\u65b9\u5b9e\u6d4b\uff0c\u6311\u4e94\u6761\u503c\u5f97\u8bb0\u7684\u3002<\/p>\n<p><strong>\u7b2c\u4e00\uff0c\u5f02\u5e38\u9000\u51fa\u540e\u91cd\u542f\u5361\u6b7b\u3002<\/strong> #5265 \u8bb0\u5f55\u4e86\u4e00\u4e2a\u5f88\u5201\u7684\u573a\u666f\uff1avLLM \u7684 worker \u5f02\u5e38\u9000\u51fa\u540e\uff0c\u6b8b\u7559\u7684 CUDA IPC \u6620\u5c04\u4f1a\u6321\u4f4f\u4e0b\u4e00\u6b21 vLLM \u91cd\u542f\u3002\u8fd9\u7c7b\u95ee\u9898\u5728 MP \u6a21\u5f0f\u5f15\u5165\u5171\u4eab\u5185\u5b58\u4f20\u8f93\u540e\u6982\u7387\u53d8\u9ad8\uff0c\u75c7\u72b6\u662f\u8fdb\u7a0b\u9000\u4e86\u4f46\u8d44\u6e90\u6ca1\u653e\u5e72\u51c0\uff0c\u53ea\u80fd\u624b\u52a8\u6e05\u7406\u3002\u751f\u4ea7\u4e0a\u5efa\u8bae\u7ed9\u91cd\u542f\u6d41\u7a0b\u7559\u4e00\u4e2a&#8221;\u786e\u8ba4\u65e0\u6b8b\u7559&#8221;\u7684\u68c0\u67e5\u6b65\u9aa4\u3002<\/p>\n<p><strong>\u7b2c\u4e8c\uff0cGPU \u4e0a\u4e0b\u6587\u4e22\u5931\u540e\u7684\u5185\u5b58\u6cc4\u6f0f\u3002<\/strong> #5270\uff1aMP \u5ba2\u6237\u7aef\u53d1\u51fa\u53bb\u4e86\u4f46\u6ca1\u4eba\u5e94\u7b54\u7684 STORE\/RETRIEVE \u8bf7\u6c42\uff0c\u5728 GPU context \u4e22\u5931\u4e4b\u540e\u4f1a\u5728\u5ba2\u6237\u7aef\u4fa7\u65e0\u9650\u5806\u79ef\u3002\u8fd9\u4e2a issue \u76ee\u524d\u8fd8\u5f00\u7740\uff0c\u4e5f\u5c31\u662f\u8bf4\u6ca1\u6709\u5f7b\u5e95\u4fee\u590d\uff0cGPU \u51fa\u8fc7\u6545\u969c\u7684\u673a\u5668\u4e0a\u8981\u7559\u610f\u5185\u5b58\u66f2\u7ebf\u3002<\/p>\n<p><strong>\u7b2c\u4e09\uff0c\u9759\u9ed8\u8fd4\u56de\u574f\u7684\u7f13\u5b58\u3002<\/strong> #5344 \u662f\u6211\u6700\u8b66\u60d5\u7684\u4e00\u4e2a\uff1a\u6279\u91cf\u5206\u914d L1 \u5931\u8d25\u65f6\u4f1a\u63d0\u4ea4\u4e00\u4e2a&#8221;\u6495\u88c2&#8221;\u7684 TP \u524d\u7f00\uff0c\u4e4b\u540e retrieve \u4e0d\u62a5\u9519\uff0c\u76f4\u63a5\u628a\u635f\u574f\u7684 KV \u8fd4\u56de\u7ed9\u4f60\u3002\u4e0d\u62a5\u9519\u7684\u574f\u6570\u636e\u6bd4\u62a5\u9519\u53ef\u6015\u5341\u500d\uff0c\u5982\u679c\u4f60\u7684\u670d\u52a1\u51fa\u73b0&#8221;\u83ab\u540d\u5176\u5999\u7684\u8f93\u51fa\u8d28\u91cf\u4e0b\u964d&#8221;\uff0c\u503c\u5f97\u628a\u7f13\u5b58\u547d\u4e2d\u7387\u65e5\u5fd7\u548c\u5f15\u64ce\u65e5\u5fd7\u5bf9\u4e00\u4e0b\u65f6\u95f4\u7ebf\u3002<\/p>\n<p><strong>\u7b2c\u56db\uff0cPCIe \u4e3b\u673a\u4e0a\u7684\u56fa\u5b9a\u642c\u8fd0\u5f00\u9500\u3002<\/strong> \u524d\u9762\u63d0\u8fc7\u7684 #5440\uff0c\u6bcf\u4e2a\u4f20\u8f93\u5bf9\u8c61\u7ea6 0.55 \u79d2\u56fa\u5b9a\u6210\u672c\uff0cpickle \u4f20\u8f93\u573a\u666f\u3002NVLink \u673a\u5668\u6ca1\u8fd9\u4e2a\u95ee\u9898\uff0cPCIe \u5355\u673a\u7528\u6237\u5728 MP \u6a21\u5f0f\u4e0b\u8981\u5bf9&#8221;\u547d\u4e2d\u6536\u76ca\u51cf\u56fa\u5b9a\u5f00\u9500&#8221;\u6709\u5fc3\u7406\u51c6\u5907\u3002<\/p>\n<p><strong>\u7b2c\u4e94\uff0c\u78c1\u76d8\u5c42\u7684\u6301\u4e45\u5316\u6ca1\u90a3\u4e48\u7a33\u3002<\/strong> #1175 \u65e9\u5728\u8bb2 SSD offloading \u7684\u7f13\u5b58\u6301\u4e45\u5316\u95ee\u9898\uff1b\u7b2c\u4e09\u65b9\u4e5f\u6709\u5b9e\u9524\u2014\u2014kubesimplify \u90a3\u7bc7\u5b9e\u6d4b\uff08\u4f5c\u8005\u5728\u4e00\u53f0\u7edf\u4e00\u5185\u5b58\u7684 ARM \u673a\u5668\u4e0a\u8fde\u6d4b\u4e09\u4e2a\u7248\u672c\u914d\u5bf9\uff09\u91cc\uff0c\u78c1\u76d8\u6301\u4e45\u5316\u8def\u5f84&#8221;\u5199\u5165\u6210\u529f\uff0c\u4f46\u91cd\u542f\u540e\u627e\u4e0d\u5230&#8221;\uff0c\u4f5c\u8005\u81ea\u5df1\u90fd\u628a\u8fd9\u6807\u8bb0\u6210\u5b9e\u9a8c\u7ed3\u8bba\u800c\u4e0d\u662f\u63a8\u8350\u914d\u7f6e\u3002<\/p>\n<p>\u7b2c\u4e09\u65b9\u4f53\u9a8c\u91cc\u8fd8\u6709\u4e24\u6761\u6211\u8ba4\u53ef\u5ea6\u9ad8\u7684\uff1alearncode.college \u90a3\u7bc7\u628a&#8221;\u4ec0\u4e48\u4eba\u522b\u7528&#8221;\u8bb2\u5f97\u5f88\u900f\uff0c\u4f5c\u8005\u8bd5\u8fc7\u7ed9\u4e00\u4e2a\u6587\u6863\u6458\u8981\u5de5\u5177\u6302\u7f13\u5b58\uff0c\u547d\u4e2d\u7387\u53ea\u6709 3%\uff0c\u7f51\u7edc\u548c\u5e8f\u5217\u5316\u5f00\u9500\u53cd\u800c\u8ba9\u5ef6\u8fdf\u6da8\u4e86 15%\uff1bMindStudio \u7684\u5bf9\u6bd4\u6587\u7ae0\u5219\u63d0\u9192 MP \u6a21\u5f0f\u4e0b vLLM \u91cd\u542f\u65f6\u7f13\u5b58\u8fdb\u7a0b\u7684\u5185\u5b58\u91ca\u653e\u6709\u5ef6\u8fdf\u3002\u5171\u540c\u6307\u5411\u4e00\u4e2a\u7ed3\u8bba\uff1a\u8fd9\u5de5\u5177\u7684\u6536\u76ca\u5b8c\u5168\u7531\u8d1f\u8f7d\u7279\u5f81\u51b3\u5b9a\uff0c\u4e0a\u4e4b\u524d\u5148\u62ff\u771f\u5b9e prompt \u8bc1\u660e\u91cd\u590d\u5ea6\uff0c\u522b\u770b star \u6570\u5934\u8111\u53d1\u70ed\u3002<\/p>\n<h2 id=\"_6\">\u6211\u6700\u540e\u7684\u5224\u65ad<\/h2>\n<p>\u503c\u4e0d\u503c\u5f97\u4e0a\u624b\uff0c\u6211\u7684\u5224\u65ad\u6807\u51c6\u5c31\u4e09\u95ee\uff1a\u4f60\u7684\u670d\u52a1\u91cd\u542f\u9891\u7e41\u5417\uff1f\u5b9e\u4f8b\u4e0d\u6b62\u4e00\u4e2a\u5417\uff1fprompt \u91cc\u6709\u5927\u6bb5\u91cd\u590d\u5185\u5bb9\uff08\u957f system prompt\u3001\u56fa\u5b9a\u6a21\u677f\u3001\u5171\u4eab\u6587\u6863\uff09\u5417\uff1f\u4e09\u4e2a\u91cc\u4e2d\u4e86\u4e24\u4e2a\uff0cLMCache \u5c31\u662f\u6b63\u6536\u76ca\uff0c\u5b98\u65b9\u57fa\u51c6\u91cc 75% \u7684 TTFT \u964d\u5e45\u548c\u8de8\u91cd\u542f\u590d\u7528\u90fd\u662f\u771f\u91d1\u767d\u94f6\uff1b\u4e09\u4e2a\u5168\u4e0d\u6cbe\uff0c\u56de\u53bb\u628a vLLM \u7684 APC \u6253\u5f00\u5c31\u597d\uff0c\u522b\u7ed9\u81ea\u5df1\u52a0\u4e00\u5c42\u8981\u7ef4\u62a4\u7684\u5206\u5e03\u5f0f\u7cfb\u7edf\u3002<\/p>\n<p>\u66ff\u4ee3\u65b9\u6848\u6211\u8bf4\u4e24\u4e2a\u3002\u7b2c\u4e00\u4e2a\u662f vLLM \u5185\u7f6e APC\uff1a\u96f6\u8fd0\u7ef4\u3001\u9ed8\u8ba4\u5c31\u8be5\u5f00\uff0c\u8f93\u5728\u8fdb\u7a0b\u5185\uff08\u91cd\u542f\u5931\u5fc6\uff09\u3001\u4ec5\u524d\u7f00\u5339\u914d\u3001\u5b9e\u4f8b\u4e4b\u95f4\u4e0d\u5171\u4eab\u2014\u2014\u5b83\u8ddf LMCache \u4e0d\u662f\u4e8c\u9009\u4e00\uff0c\u5b98\u65b9\u94fe\u8def\u91cc\u4e24\u8005\u662f\u53e0\u7740\u7528\u7684\uff08GPU \u5185\u547d\u4e2d\u5f52 APC\uff0cCPU \u5c42\u8865\u4f4d\u5f52 LMCache\uff09\u3002\u7b2c\u4e8c\u4e2a\u662f SGLang \u52a0 RadixAttention\uff1a\u5f15\u64ce\u5185\u7684\u57fa\u6570\u6811\u7f13\u5b58\uff0c\u591a\u8f6e\u5bf9\u8bdd\u7684\u5206\u652f\u590d\u7528\u505a\u5f97\u5f88\u4f18\u96c5\uff0c\u5982\u679c\u4f60\u7684\u6808\u672c\u6765\u5c31\u613f\u610f\u6574\u4f53\u8fc1\u5230 SGLang\uff0c\u5b83\u81ea\u5e26\u7684\u7f13\u5b58\u4f53\u7cfb\u80fd\u8986\u76d6\u5927\u90e8\u5206\u5355\u673a\u9700\u6c42\uff1b\u4f46\u5b83\u540c\u6837\u6d3b\u5728\u8fdb\u7a0b\u91cc\uff0c\u8de8\u5f15\u64ce\u590d\u7528\u548c\u6301\u4e45\u5316\u5c42\u7ea7\u5c31\u662f LMCache \u7684\u4e3b\u573a\u4e86\u3002\u6211\u7684\u6392\u5e8f\u5f88\u7b80\u5355\uff1a\u5355\u5f15\u64ce\u5355\u673a\u5148\u69a8\u5e72\u5f15\u64ce\u81ea\u5e26\u80fd\u529b\uff0c\u591a\u5b9e\u4f8b\u3001\u8981\u6301\u4e45\u3001\u8de8\u5f15\u64ce\uff0c\u518d\u8bf7 LMCache \u51fa\u5c71\u3002<\/p>\n<p>\u6700\u540e\u8bf4\u53e5\u5b9e\u5728\u7684\uff0c\u8fd9\u4e2a\u9879\u76ee\u4e0d\u662f\u6211\u65e5\u5e38\u4f1a\u88c5\u5728\u81ea\u5df1\u7b14\u8bb0\u672c\u4e0a\u7684\u90a3\u79cd\u73a9\u5177\uff0c\u800c\u662f\u4e00\u5c42\u660e\u786e\u5c5e\u4e8e\u751f\u4ea7\u63a8\u7406\u67b6\u6784\u7684\u57fa\u7840\u8bbe\u65bd\u3002\u987a\u624b\u62c9\u4e86\u4e0b\uff0c\u73b0\u5728\u662f 12,272 \u4e2a star\uff0c\u6700\u8fd1\u56db\u5468 311 \u6b21\u63d0\u4ea4\uff0c\u7a33\u5b9a\u7248 0.5.6\uff0c\u6628\u5929\u8fd8\u5728\u4e00\u53e3\u6c14\u53d1 0.5.7rc4 \u7684\u4e94\u79cd\u786c\u4ef6\u6784\u5efa\u2014\u2014\u7ef4\u62a4\u5f3a\u5ea6\u548c\u751f\u6001\u7ad9\u4f4d\u90fd\u5728\u90a3\uff0c\u552f\u4e00\u7684\u95ee\u9898\u662f\u4f60\u7684\u8d1f\u8f7d\u914d\u4e0d\u914d\u5f97\u4e0a\u5b83\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>LMCache \u628a KV cache \u4ece\u63a8\u7406\u8fdb\u7a0b\u91cc\u642c\u51fa\u6765\uff0c\u843d\u5230 CPU \u5185\u5b58\u548c SSD \u91cc\u8de8\u8bf7\u6c42\u3001\u8de8\u91cd\u542f\u590d\u7528\u3002\u6211\u7ffb\u5b8c\u5b98\u65b9\u6587\u6863\u3001\u517c\u5bb9\u6027\u77e9\u9635\u548c 909 \u4e2a\u5f00\u653e issue\uff1a\u5b98\u65b9\u57fa\u51c6 TTFT \u964d 75%\uff0c\u4f46\u77ed prompt\u3001\u5355\u5b9e\u4f8b\u573a\u666f\u6536\u76ca\u53ef\u80fd\u5012\u8d34\uff0c\u503c\u4e0d\u503c\u5f97\u4e0a\u8981\u770b\u8d1f\u8f7d\u7279\u5f81\u3002<\/p>\n","protected":false},"author":1,"featured_media":10021538,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[4],"tags":[],"class_list":["post-10021539","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-share"],"jetpack_sharing_enabled":true,"jetpack_featured_media_url":"https:\/\/itmax.cn\/wp-content\/uploads\/2026\/11\/lmcache-cover.png","_links":{"self":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10021539","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=10021539"}],"version-history":[{"count":1,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10021539\/revisions"}],"predecessor-version":[{"id":10021540,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10021539\/revisions\/10021540"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/media\/10021538"}],"wp:attachment":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=10021539"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=10021539"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=10021539"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}