{"id":91739,"date":"2026-08-08T11:17:08","date_gmt":"2026-08-08T03:17:08","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/91739.html"},"modified":"2026-08-08T11:17:08","modified_gmt":"2026-08-08T03:17:08","slug":"%e4%bb%8e-pagedattention-%e5%88%b0-prefill-decode-%e5%88%86%e7%a6%bb%e7%9a%84%e6%b7%b1%e5%ba%a6%e8%a7%a3%e6%9e%90","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/91739.html","title":{"rendered":"\u4ece PagedAttention \u5230 Prefill-Decode \u5206\u79bb\u7684\u6df1\u5ea6\u89e3\u6790"},"content":{"rendered":"<h2>\u5927\u6a21\u578b\u63a8\u7406\u67b6\u6784\u6f14\u8fdb&#xff1a;\u4ece PagedAttention \u5230 Prefill-Decode \u5206\u79bb\u7684\u6df1\u5ea6\u89e3\u6790<\/h2>\n<p>\u672c\u6587\u4ece\u5927\u6a21\u578b\u63a8\u7406\u7684\u4e24\u9636\u6bb5\u8ba1\u7b97\u7279\u6027\u51fa\u53d1&#xff0c;\u7cfb\u7edf\u62c6\u89e3 PagedAttention\u3001\u8fde\u7eed\u6279\u5904\u7406\u7684\u6838\u5fc3\u539f\u7406&#xff0c;\u6df1\u5165\u5206\u6790 Prefill-Decode \u5206\u79bb\u67b6\u6784&#xff08;DistServe\u3001Splitwise\u3001Mooncake&#xff09;\u7684\u8bbe\u8ba1\u54f2\u5b66&#xff0c;\u5e76\u7ed3\u5408 vLLM Wide-EP \u4e0e DeepSeek \u751f\u4ea7\u90e8\u7f72\u6848\u4f8b&#xff0c;\u63a2\u8ba8 2025-2026 \u5e74\u63a8\u7406\u57fa\u7840\u8bbe\u65bd\u7684\u524d\u6cbf\u8d8b\u52bf\u3002<\/p>\n<h3>\u4e00\u3001\u63a8\u7406\u6027\u80fd\u7684\u74f6\u9888&#xff1a;Prefill \u4e0e Decode \u7684\u4e0d\u5bf9\u79f0\u6027<\/h3>\n<p>\u7406\u89e3\u5927\u6a21\u578b\u63a8\u7406\u7684\u6240\u4f18\u5316\u6280\u672f&#xff0c;\u9996\u5148\u8981\u56de\u5230\u4e00\u4e2a\u6839\u672c\u4e8b\u5b9e&#xff1a;\u81ea\u56de\u5f52\u751f\u6210\u5305\u542b\u4e24\u4e2a\u8ba1\u7b97\u7279\u6027\u622a\u7136\u4e0d\u540c\u7684\u9636\u6bb5\u3002<\/p>\n<p>Prefill \u9636\u6bb5\u8d1f\u8d23\u5904\u7406\u7528\u6237\u8f93\u5165\u7684\u5b8c\u6574 prompt\u3002\u4ee5\u4e00\u4e2a 4096 token \u7684\u8f93\u5165\u4e3a\u4f8b&#xff0c;\u6a21\u578b\u9700\u8981\u505a\u4e00\u6b21\u5b8c\u6574\u7684\u524d\u5411\u4f20\u64ad&#xff0c;\u6d89\u53ca\u5927\u89c4\u6a21\u77e9\u9635\u4e58\u6cd5\u3002\u6b64\u65f6 GPU \u7684\u7b97\u672f\u5f3a\u5ea6&#xff08;FLOP\/byte&#xff09;\u7ea6\u5728 200-400 \u4e4b\u95f4&#xff0c;\u8fdc\u8d85 H100 \u7684\u5185\u5b58\u5e26\u5bbd\u5c4b\u9876\u7ebf&#xff0c;\u5c5e\u4e8e\u5178\u578b\u7684**\u8ba1\u7b97\u53d7\u9650&#xff08;compute-bound&#xff09;**\u573a\u666f&#xff0c;GPU \u5229\u7528\u7387\u53ef\u4ee5\u63a5\u8fd1\u5cf0\u503c\u3002<\/p>\n<p>Decode \u9636\u6bb5\u5219\u622a\u7136\u4e0d\u540c\u3002\u6bcf\u4e00\u6b65\u53ea\u751f\u6210\u4e00\u4e2a token&#xff0c;\u4f46\u9700\u8981\u8bfb\u53d6\u5168\u90e8\u5386\u53f2 KV Cache \u548c\u6a21\u578b\u6743\u91cd\u624d\u80fd\u8ba1\u7b97\u4e0b\u4e00\u4e2a token\u3002\u7b97\u672f\u5f3a\u5ea6\u9aa4\u964d\u81f3\u7ea6 60-80 FLOP\/byte&#xff0c;GPU \u5b9e\u9645\u5229\u7528\u7387\u4ec5 20%-40%&#xff0c;\u5c5e\u4e8e**\u5185\u5b58\u5e26\u5bbd\u53d7\u9650&#xff08;memory-bandwidth-bound&#xff09;**\u573a\u666f\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260808031706-6a769fb218249.png\" alt=\"\" \/><\/p>\n<p>\u8fd9\u4e24\u9636\u6bb5\u7684\u4e0d\u5bf9\u79f0\u6027&#xff0c;\u662f\u540e\u7eed\u6240\u6709\u63a8\u7406\u4f18\u5316\u6280\u672f\u7684\u51fa\u53d1\u70b9\u3002PagedAttention \u89e3\u51b3\u7684\u662f Decode \u9636\u6bb5\u7684\u663e\u5b58\u7ba1\u7406\u95ee\u9898&#xff0c;\u8fde\u7eed\u6279\u5904\u7406\u89e3\u51b3\u7684\u662f\u4e24\u9636\u6bb5\u6df7\u5408\u8c03\u5ea6\u95ee\u9898&#xff0c;\u800c Prefill-Decode \u5206\u79bb\u67b6\u6784\u5219\u4ece\u6839\u672c\u4e0a\u89e3\u9664\u4e86\u4e24\u9636\u6bb5\u7684\u8d44\u6e90\u8026\u5408\u3002<\/p>\n<h3>\u4e8c\u3001PagedAttention&#xff1a;\u64cd\u4f5c\u7cfb\u7edf\u7ea7\u7684 KV Cache \u7ba1\u7406<\/h3>\n<h4>2.1 \u4f20\u7edf\u65b9\u6848\u7684\u788e\u7247\u5316\u56f0\u5883<\/h4>\n<p>\u5728 PagedAttention \u51fa\u73b0\u4e4b\u524d&#xff0c;\u63a8\u7406\u5f15\u64ce\u4e3a\u6bcf\u4e2a\u8bf7\u6c42\u9884\u5206\u914d\u4e00\u6bb5\u8fde\u7eed\u7684 KV Cache \u663e\u5b58\u7a7a\u95f4\u3002\u8fd9\u79cd\u65b9\u6848\u5b58\u5728\u4e24\u4e2a\u4e25\u91cd\u95ee\u9898&#xff1a;<\/p>\n<ul>\n<li>\u5185\u90e8\u788e\u7247&#xff1a;\u9884\u5206\u914d\u7a7a\u95f4\u6309\u8bf7\u6c42\u53ef\u80fd\u7684\u6700\u5927\u5e8f\u5217\u957f\u5ea6\u9884\u7559&#xff0c;\u4f46\u5b9e\u9645\u751f\u6210\u5f80\u5f80\u8fdc\u77ed\u4e8e\u4e0a\u9650&#xff0c;reserved \u4f46\u672a\u7528\u7684\u663e\u5b58\u767d\u767d\u6d6a\u8d39\u3002\u5b9e\u6d4b\u4e2d&#xff0c;\u5185\u90e8\u788e\u7247\u5bfc\u81f4\u7ea6 60%-80% \u7684\u663e\u5b58\u88ab\u6d6a\u8d39\u3002<\/li>\n<li>\u5916\u90e8\u788e\u7247&#xff1a;\u8bf7\u6c42\u4e0d\u65ad\u521b\u5efa\u548c\u9500\u6bc1&#xff0c;\u663e\u5b58\u4e2d\u51fa\u73b0\u5927\u91cf\u4e0d\u8fde\u7eed\u7a7a\u6d1e&#xff0c;\u65b0\u8bf7\u6c42\u53ef\u80fd\u56e0\u627e\u4e0d\u5230\u8db3\u591f\u5927\u7684\u8fde\u7eed\u5757\u800c\u5931\u8d25&#xff0c;\u5373\u4f7f\u603b\u5269\u4f59\u663e\u5b58\u5145\u8db3\u3002<\/li>\n<\/ul>\n<h4>2.2 \u865a\u62df\u5185\u5b58\u601d\u60f3\u7684\u6838\u5fc3\u6620\u5c04<\/h4>\n<p>vLLM \u7684 PagedAttention \u501f\u9274\u4e86\u64cd\u4f5c\u7cfb\u7edf\u7684\u865a\u62df\u5185\u5b58\u5206\u9875\u673a\u5236&#xff0c;\u6838\u5fc3\u8bbe\u8ba1\u5206\u4e09\u6b65&#xff1a;<\/p>\n<p>\u7b2c\u4e00\u6b65&#xff1a;\u7269\u7406\u5757\u5212\u5206\u3002 \u5c06 KV Cache \u7684\u5b58\u50a8\u7a7a\u95f4\u5212\u5206\u4e3a\u56fa\u5b9a\u5927\u5c0f\u7684\u7269\u7406\u5757&#xff08;Block&#xff09;&#xff0c;\u6bcf\u4e2a Block \u5b58\u50a8\u56fa\u5b9a\u6570\u91cf token \u7684 Key \u548c Value \u5f20\u91cf\u3002<\/p>\n<p>\u7b2c\u4e8c\u6b65&#xff1a;\u9875\u8868\u6620\u5c04\u3002 \u6bcf\u4e2a\u8bf7\u6c42\u7ef4\u62a4\u4e00\u5f20\u9875\u8868&#xff08;Page Table&#xff09;&#xff0c;\u8bb0\u5f55\u903b\u8f91 token \u5e8f\u5217\u5230\u7269\u7406\u5757\u7684\u6620\u5c04\u5173\u7cfb\u3002\u4e00\u4e2a\u8bf7\u6c42\u7684 KV Cache \u53ef\u4ee5\u5206\u6563\u5b58\u50a8\u5728\u591a\u4e2a\u4e0d\u8fde\u7eed\u7684\u7269\u7406\u5757\u4e2d\u3002<\/p>\n<p>\u7b2c\u4e09\u6b65&#xff1a;\u6309\u9700\u5206\u914d\u4e0e\u56de\u6536\u3002 \u7269\u7406\u5757\u6309\u9700\u5206\u914d\u2014\u2014\u8bf7\u6c42\u751f\u6210\u65b0 token \u65f6\u624d\u7533\u8bf7\u65b0\u5757&#xff1b;\u8bf7\u6c42\u5b8c\u6210\u540e&#xff0c;\u5176\u6240\u6709\u7269\u7406\u5757\u7acb\u5373\u56de\u6536\u5230\u5168\u5c40\u6c60&#xff0c;\u4f9b\u5176\u4ed6\u8bf7\u6c42\u590d\u7528\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260808031706-6a769fb24085c.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>\u8fd9\u4e00\u8bbe\u8ba1\u5f7b\u5e95\u6d88\u9664\u4e86\u5185\u90e8\u788e\u7247&#xff08;\u5757\u6309\u9700\u5206\u914d&#xff0c;\u4e0d\u591a\u9884\u7559&#xff09;\u548c\u5916\u90e8\u788e\u7247&#xff08;\u7269\u7406\u5757\u4e0d\u8981\u6c42\u8fde\u7eed&#xff0c;\u4efb\u4f55\u7a7a\u95f2\u5757\u90fd\u53ef\u590d\u7528&#xff09;\u3002\u5728 vLLM \u7684\u5b9e\u9a8c\u4e2d&#xff0c;PagedAttention \u5c06 KV Cache \u7684\u663e\u5b58\u5229\u7528\u7387\u4ece\u7ea6 20%-40% \u63d0\u5347\u81f3\u63a5\u8fd1 100%\u3002<\/p>\n<h4>2.3 \u5171\u4eab\u524d\u7f00\u7684\u989d\u5916\u6536\u76ca<\/h4>\n<p>PagedAttention \u7684\u9875\u8868\u8bbe\u8ba1\u8fd8\u5e26\u6765\u4e86\u4e00\u4e2a\u610f\u5916\u6536\u83b7&#xff1a;\u5171\u4eab\u524d\u7f00\u590d\u7528\u3002\u5f53\u591a\u4e2a\u8bf7\u6c42\u5171\u4eab\u76f8\u540c\u7684 system prompt \u6216 few-shot \u793a\u4f8b\u65f6&#xff0c;\u5b83\u4eec\u53ef\u4ee5\u6620\u5c04\u5230\u540c\u4e00\u7ec4\u7269\u7406\u5757&#xff0c;\u53ea\u8ba1\u7b97\u4e00\u6b21 KV Cache \u5373\u53ef\u8de8\u8bf7\u6c42\u590d\u7528\u3002\u8fd9\u4e00\u601d\u8def\u540e\u6765\u88ab SGLang \u7684 RadixAttention \u8fdb\u4e00\u6b65\u53d1\u626c\u5149\u5927\u3002<\/p>\n<h3>\u4e09\u3001\u8fde\u7eed\u6279\u5904\u7406&#xff1a;\u8ba9 GPU \u6c38\u4e0d\u7a7a\u95f2<\/h3>\n<h4>3.1 \u9759\u6001\u6279\u5904\u7406\u7684\u77ed\u677f<\/h4>\n<p>\u4f20\u7edf\u63a8\u7406\u5f15\u64ce\u91c7\u7528\u9759\u6001\u6279\u5904\u7406&#xff1a;\u6536\u96c6\u4e00\u6279\u8bf7\u6c42&#xff0c;\u51d1\u6ee1 batch size \u540e\u4e00\u8d77\u9001\u5165 GPU \u6267\u884c&#xff0c;\u7b49\u5f85\u6240\u6709\u5e8f\u5217\u751f\u6210\u5b8c\u6bd5\u624d\u91ca\u653e\u8d44\u6e90\u3002\u95ee\u9898\u5728\u4e8e&#xff0c;\u540c\u4e00 batch \u4e2d\u5404\u8bf7\u6c42\u7684\u751f\u6210\u957f\u5ea6\u5dee\u5f02\u5de8\u5927\u2014\u2014\u4e00\u4e2a\u751f\u6210 10 token \u7684\u77ed\u8bf7\u6c42\u548c\u4e00\u4e2a\u751f\u6210 500 token \u7684\u957f\u8bf7\u6c42\u88ab\u7ed1\u5b9a\u5728\u4e00\u8d77&#xff0c;\u77ed\u8bf7\u6c42\u5b8c\u6210\u540e\u5fc5\u987b\u7a7a\u7b49\u957f\u8bf7\u6c42&#xff0c;GPU \u5728\u8fd9\u6bb5\u65f6\u95f4\u5185\u5927\u91cf\u7b97\u529b\u88ab\u6d6a\u8d39\u3002<\/p>\n<h4>3.2 \u52a8\u6001\u63d2\u5165\u4e0e\u79fb\u9664<\/h4>\n<p>\u8fde\u7eed\u6279\u5904\u7406&#xff08;Continuous Batching&#xff0c;\u53c8\u79f0 Inflight Batching&#xff09;\u5728\u6bcf\u4e2a\u89e3\u7801\u8fed\u4ee3\u6b65\u7684\u8fb9\u754c\u4e0a\u505a\u4e24\u4ef6\u4e8b&#xff1a;<\/p>\n<li>\u79fb\u9664\u5df2\u5b8c\u6210\u8bf7\u6c42&#xff1a;\u4efb\u4f55\u5df2\u751f\u6210 EOS token \u6216\u8fbe\u5230\u957f\u5ea6\u4e0a\u9650\u7684\u8bf7\u6c42&#xff0c;\u7acb\u5373\u4ece batch \u4e2d\u79fb\u9664&#xff0c;\u91ca\u653e\u5176\u7269\u7406\u5757\u3002<\/li>\n<li>\u63d2\u5165\u65b0\u5230\u8fbe\u8bf7\u6c42&#xff1a;\u65b0\u5230\u8fbe\u7684\u8bf7\u6c42\u5728\u5f53\u524d\u8fed\u4ee3\u6b65\u5f00\u59cb\u524d\u88ab\u52a0\u5165 batch&#xff0c;\u7acb\u5373\u5f00\u59cb prefill\u3002<\/li>\n<p>\u8fd9\u79cd&#034;\u9010\u6b65\u52a8\u6001\u8c03\u6574&#034;\u4f7f GPU \u5728\u7a81\u53d1\u6d41\u91cf\u4e0b\u65e0\u9700\u624b\u52a8\u6279\u5904\u7406\u903b\u8f91\u5373\u53ef\u7ef4\u6301\u9ad8\u5229\u7528\u7387\u3002\u4e0e PagedAttention \u914d\u5408\u65f6&#xff0c;\u8fde\u7eed\u6279\u5904\u7406\u5c24\u5176\u5f3a\u5927\u2014\u2014\u56e0\u4e3a PagedAttention \u8ba9\u6bcf\u4e2a\u8bf7\u6c42\u7684 KV Cache \u53ef\u4ee5\u72ec\u7acb\u7ba1\u7406&#xff0c;\u65b0\u8bf7\u6c42\u7684\u52a0\u5165\u4e0d\u4f1a\u5e72\u6270\u5df2\u6709\u8bf7\u6c42\u7684\u5185\u5b58\u5e03\u5c40\u3002<\/p>\n<h4>3.3 \u5b9e\u6d4b\u6027\u80fd\u5f71\u54cd<\/h4>\n<p>\u5728 H100 \u4e0a\u7684\u57fa\u51c6\u6d4b\u8bd5\u4e2d&#xff0c;\u8fde\u7eed\u6279\u5904\u7406\u4f7f vLLM \u5728 100 \u5e76\u53d1\u4e0b\u7684\u541e\u5410\u91cf\u8fbe\u5230 2400 tokens\/s&#xff0c;\u76f8\u6bd4\u9759\u6001\u6279\u5904\u7406\u63d0\u5347\u7ea6 3-5 \u500d\u3002TensorRT-LLM \u5728\u76f8\u540c\u573a\u666f\u4e0b\u8fbe\u5230 2780 tokens\/s&#xff0c;\u9886\u5148\u7ea6 13%&#xff0c;\u8fd9\u5f97\u76ca\u4e8e\u5176\u7f16\u8bd1\u578b\u5f15\u64ce\u5bf9 kernel \u7684\u6df1\u5ea6\u4f18\u5316\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260808031706-6a769fb28c69c.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<h3>\u56db\u3001Prefill-Decode \u5206\u79bb\u67b6\u6784&#xff1a;\u7a81\u7834\u8d44\u6e90\u8026\u5408\u74f6\u9888<\/h3>\n<h4>4.1 \u5171\u7f6e\u7cfb\u7edf\u7684\u5e72\u6270\u95ee\u9898<\/h4>\n<p>\u5728\u4f20\u7edf\u7684\u5171\u7f6e&#xff08;colocated&#xff09;\u7cfb\u7edf\u4e2d&#xff0c;Prefill \u548c Decode \u4ea4\u66ff\u8fd0\u884c\u5728\u540c\u4e00\u7ec4 GPU \u4e0a\u3002\u8fde\u7eed\u6279\u5904\u7406\u867d\u7136\u63d0\u9ad8\u4e86\u5229\u7528\u7387&#xff0c;\u4f46\u5e26\u6765\u4e86\u4e00\u4e2a\u9690\u853d\u4f46\u4e25\u91cd\u7684\u95ee\u9898\u2014\u2014ITL&#xff08;Inter-Token Latency&#xff09;\u5c16\u5cf0\u3002<\/p>\n<p>\u5f53\u4e00\u4e2a\u65b0\u7684\u957f prompt \u5230\u8fbe\u5e76\u5f00\u59cb prefill \u65f6&#xff0c;\u5927\u91cf\u8ba1\u7b97\u8d44\u6e90\u88ab prefill \u5360\u636e&#xff0c;\u6b63\u5728\u8fdb\u884c\u7684 decode \u8fed\u4ee3\u88ab\u8feb\u7b49\u5f85\u3002\u7528\u6237\u4f53\u9a8c\u4e0a&#xff0c;\u8fd9\u8868\u73b0\u4e3a token \u6d41\u5f0f\u8f93\u51fa\u7684&#034;\u505c\u987f-\u7a81\u53d1-\u505c\u987f&#034;\u6a21\u5f0f&#xff1a;\u6b63\u5e38\u6bcf 50ms \u8f93\u51fa\u4e00\u4e2a token&#xff0c;\u7a81\u7136\u5361\u987f 500ms&#xff0c;\u7136\u540e\u4e00\u53e3\u6c14\u5410\u51fa\u591a\u4e2a token&#xff0c;\u518d\u7ee7\u7eed\u5361\u987f\u3002<\/p>\n<p>\u6b64\u5916&#xff0c;\u8d44\u6e90\u5206\u914d\u4e5f\u88ab\u8026\u5408&#xff1a;GPU \u663e\u5b58\u3001\u5e76\u884c\u7b56\u7565\u3001batch size \u5fc5\u987b\u540c\u65f6\u670d\u52a1\u4e24\u79cd\u5de5\u4f5c\u8d1f\u8f7d&#xff0c;\u65e0\u6cd5\u540c\u65f6\u4e3a prefill \u548c decode \u5404\u81ea\u627e\u5230\u6700\u4f18\u914d\u7f6e\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260808031706-6a769fb2cbe2a.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<h4>4.2 DistServe&#xff1a;Goodput \u9a71\u52a8\u7684\u8d44\u6e90\u5206\u914d<\/h4>\n<p>DistServe&#xff08;OSDI 2024&#xff0c;\u5317\u5927\/UCSD\/StepFun&#xff09;\u9996\u6b21\u7cfb\u7edf\u5316\u5730\u63d0\u51fa\u4e86\u5206\u79bb\u5f0f\u63a8\u7406\u7684\u6846\u67b6\u3002\u5176\u6838\u5fc3\u8d21\u732e\u6709\u4e09&#xff1a;<\/p>\n<p>\u72ec\u7acb\u8d44\u6e90\u5206\u914d\u3002 Prefill \u548c Decode \u88ab\u5206\u914d\u5230\u72ec\u7acb\u7684 GPU \u96c6\u5408&#xff0c;\u5404\u9636\u6bb5\u72ec\u7acb\u9009\u62e9\u6700\u4f18\u7684\u5e76\u884c\u7b56\u7565\u548c batch size\u3002Prefill \u96c6\u7fa4\u914d\u7f6e\u4e3a\u9ad8 FLOPS \u5229\u7528\u7387&#xff0c;Decode \u96c6\u7fa4\u914d\u7f6e\u4e3a\u9ad8\u663e\u5b58\u5e26\u5bbd\u5229\u7528\u7387\u3002<\/p>\n<p>\u5e26\u5bbd\u611f\u77e5\u653e\u7f6e\u7b97\u6cd5\u3002 Prefill \u5b8c\u6210\u540e\u9700\u8981\u5c06 KV Cache \u4f20\u8f93\u7ed9 Decode worker&#xff0c;\u4f20\u8f93\u5ef6\u8fdf\u662f\u5206\u79bb\u67b6\u6784\u7684\u5173\u952e\u5f00\u9500\u3002DistServe \u7684\u653e\u7f6e\u7b97\u6cd5\u5728\u5e26\u5bbd\u5145\u8db3\u65f6\u8ba9 KV \u4f20\u8f93\u4e0e\u8ba1\u7b97\u91cd\u53e0\u9690\u85cf\u5ef6\u8fdf&#xff0c;\u5728\u5e26\u5bbd\u53d7\u9650\u65f6\u8c03\u6574\u653e\u7f6e\u7b56\u7565\u51cf\u5c11\u4f20\u8f93\u91cf\u3002<\/p>\n<p>Goodput \u6307\u6807\u3002 DistServe \u5f15\u5165\u4e86 Goodput \u6982\u5ff5\u2014\u2014\u7cfb\u7edf\u5728\u6ee1\u8db3\u6307\u5b9a TTFT&#xff08;\u9996\u5b57\u5ef6\u8fdf&#xff09;\u548c TPOT&#xff08;\u6bcf token \u5ef6\u8fdf&#xff09;\u76ee\u6807\u524d\u63d0\u4e0b\u80fd\u627f\u53d7\u7684\u6700\u5927\u8bf7\u6c42\u901f\u7387\u3002\u8fd9\u6bd4\u5355\u7eaf\u8ffd\u6c42 tokens\/s \u66f4\u8d34\u8fd1\u771f\u5b9e\u7528\u6237\u4f53\u9a8c\u3002<\/p>\n<p>\u5b9e\u6d4b\u7ed3\u679c&#xff1a;\u804a\u5929\u673a\u5668\u4eba\u573a\u666f Goodput \u63d0\u5347 2.0-3.41 \u500d&#xff0c;\u4ee3\u7801\u8865\u5168 3.2 \u500d&#xff0c;\u6458\u8981\u573a\u666f 4.48 \u500d\u3002\u7cfb\u7edf\u7ea7\u8fbe\u5230 7.4 \u500d\u8bf7\u6c42\u541e\u5410&#xff0c;SLO \u7ea6\u675f\u6536\u7d27 12.6 \u500d\u3002<\/p>\n<h4>4.3 Splitwise&#xff1a;\u5f02\u6784\u786c\u4ef6\u7684\u9010\u5c42\u6d41\u6c34\u7ebf<\/h4>\n<p>Microsoft Research \u7684 Splitwise&#xff08;ISCA 2024&#xff09;\u4ece\u786c\u4ef6\u6210\u672c\u89c6\u89d2\u5207\u5165&#xff0c;\u63d0\u51fa\u4e86\u5f02\u6784\u786c\u4ef6\u7b56\u7565&#xff1a;<\/p>\n<ul>\n<li>Prefill \u96c6\u7fa4\u4f7f\u7528\u9ad8 FLOPS \u7684\u6700\u65b0 GPU&#xff08;\u5982 H100\/B200&#xff09;&#xff0c;\u5339\u914d\u8ba1\u7b97\u53d7\u9650\u7279\u6027\u3002<\/li>\n<li>Decode \u96c6\u7fa4\u4f7f\u7528\u5927 HBM \u5bb9\u91cf\u3001\u9ad8\u5e26\u5bbd\u4f46 FLOPS \u76f8\u5bf9\u8f83\u4f4e\u7684\u6210\u672c\u6548\u76ca GPU&#xff0c;\u5339\u914d\u5e26\u5bbd\u53d7\u9650\u7279\u6027\u3002<\/li>\n<\/ul>\n<p>Splitwise \u7684\u5173\u952e\u5de5\u7a0b\u8d21\u732e\u662f\u9010\u5c42\u6d41\u6c34\u7ebf\u4f20\u8f93&#xff1a;\u6bcf\u4e00\u5c42 prefill \u8ba1\u7b97\u5b8c\u6210\u540e\u7acb\u5373\u5f02\u6b65\u4f20\u8f93\u8be5\u5c42\u7684 KV Cache&#xff0c;\u4e0d\u9700\u8981\u7b49\u5168\u90e8 prefill \u5b8c\u6210\u624d\u5f00\u59cb\u4f20\u8f93\u3002\u8fd9\u79cd\u6d41\u6c34\u7ebf\u8bbe\u8ba1\u9690\u85cf\u4e86\u5927\u90e8\u5206\u4f20\u8f93\u5ef6\u8fdf&#xff0c;\u5b9e\u73b0\u7ea6 1.4 \u500d\u541e\u5410\u63d0\u5347\u548c 20% \u7684\u6bcf token \u6210\u672c\u4e0b\u964d\u3002\u5728\u7b49\u529f\u7387\/\u6210\u672c\u9884\u7b97\u4e0b&#xff0c;\u541e\u5410\u63d0\u5347\u8fbe 2.35 \u500d\u3002\u8fd9\u4e00\u6280\u672f\u540e\u6765\u88ab\u4e0a\u6e38\u81f3 vLLM \u5f00\u6e90\u9879\u76ee\u3002<\/p>\n<h4>4.4 Mooncake&#xff1a;KVCache-centric \u7684\u751f\u4ea7\u7cfb\u7edf<\/h4>\n<p>Moonshot AI \u7684 Mooncake&#xff08;FAST 2025 \u6700\u4f73\u8bba\u6587&#xff09;\u4ee3\u8868\u4e86\u5206\u79bb\u5f0f\u63a8\u7406\u7684\u751f\u4ea7\u7ea7\u5b9e\u8df5&#xff0c;\u5176\u8bbe\u8ba1\u54f2\u5b66\u662fKVCache-centric\u2014\u2014KV Cache \u7684\u653e\u7f6e\u548c\u79fb\u52a8\u9a71\u52a8\u6240\u6709\u8c03\u5ea6\u51b3\u7b56\u3002<\/p>\n<p>\u5206\u5e03\u5f0f KV Cache \u5b58\u50a8\u3002 Mooncake \u5229\u7528 GPU \u96c6\u7fa4\u8282\u70b9\u4e0a\u672a\u5145\u5206\u4f7f\u7528\u7684 CPU DRAM\u3001\u672c\u5730 SSD \u548c NVMe \u5b58\u50a8&#xff0c;\u6784\u5efa\u4e86\u4e00\u4e2a\u5206\u5c42\u7684\u5206\u5e03\u5f0f KV Cache \u5b58\u50a8\u7cfb\u7edf\u3002\u8bf7\u6c42\u5b8c\u6210\u540e&#xff0c;KV Cache \u4e0d\u7acb\u5373\u4e22\u5f03&#xff0c;\u800c\u662f\u4fdd\u7559\u5728\u5b58\u50a8\u5c42\u4f9b\u672a\u6765\u5177\u6709\u5171\u4eab\u524d\u7f00\u7684\u8bf7\u6c42\u590d\u7528\u3002<\/p>\n<p>Transfer Engine\u3002 \u8fd9\u662f Mooncake \u7684\u6838\u5fc3\u901a\u4fe1\u5c42&#xff0c;\u652f\u6301 RDMA \u548c\u62d3\u6251\u611f\u77e5\u8def\u5f84\u9009\u62e9&#xff08;NVLink\/InfiniBand\/RoCE\/Ethernet&#xff09;&#xff0c;\u80fd\u805a\u5408\u591a\u5361\u5e26\u5bbd\u5b9e\u73b0\u9ad8\u541e\u5410 KV \u4f20\u8f93\u3002\u8be5\u7ec4\u4ef6\u5df2\u4e8e 2024 \u5e74 11 \u6708\u5f00\u6e90\u3002<\/p>\n<p>\u751f\u4ea7\u89c4\u6a21\u3002 Kimi \u6bcf\u65e5\u5904\u7406\u8d85 1000 \u4ebf token&#xff0c;\u8de8\u6570\u5343\u8282\u70b9\u8fd0\u884c\u3002\u6709\u6548\u8bf7\u6c42\u5bb9\u91cf\u63d0\u5347 59%-498%&#xff0c;SLO \u7ea6\u675f\u4e0b\u541e\u5410\u63d0\u5347\u6700\u9ad8 525%&#xff0c;\u771f\u5b9e\u5de5\u4f5c\u8d1f\u8f7d\u4e0b\u591a\u5904\u7406 75% \u7684\u8bf7\u6c42\u3002<\/p>\n<h3>\u4e94\u3001\u751f\u4ea7\u5b9e\u8df5&#xff1a;vLLM Wide-EP \u4e0e DeepSeek \u90e8\u7f72<\/h3>\n<h4>5.1 Wide-EP \u7b56\u7565<\/h4>\n<p>DeepSeek-R1 \u662f\u4e00\u4e2a 671B \u53c2\u6570\u7684 MoE \u6a21\u578b&#xff0c;\u6bcf\u6b21\u63a8\u7406\u4ec5\u6fc0\u6d3b 37B \u53c2\u6570\u3002\u5728\u591a GPU \u90e8\u7f72\u65f6&#xff0c;\u4f20\u7edf\u7684\u5f20\u91cf\u5e76\u884c&#xff08;TP&#xff09;\u7b56\u7565\u5728 MLA&#xff08;Multi-head Latent Attention&#xff09;\u67b6\u6784\u4e0b\u4f1a\u5bfc\u81f4\u6f5c\u53d8\u91cf\u6295\u5f71\u7684\u91cd\u590d\u8ba1\u7b97\u3002vLLM \u7684 Wide-EP&#xff08;Wide Expert Parallelism&#xff09;\u7b56\u7565\u5c06\u4e13\u5bb6\u5e76\u884c&#xff08;EP&#xff09;\u4e0e\u6570\u636e\u5e76\u884c&#xff08;DP&#xff09;\u7ed3\u5408&#xff1a;<\/p>\n<ul>\n<li>\u6ce8\u610f\u529b\u5c42\u91c7\u7528 DP \u90e8\u7f72&#xff0c;\u6bcf\u4e2a rank \u72ec\u7acb\u6301\u6709\u6f5c\u53d8\u91cf\u6295\u5f71&#xff0c;\u589e\u52a0\u6709\u6548 batch size\u3002<\/li>\n<li>MoE \u5c42\u91c7\u7528 EP \u90e8\u7f72&#xff0c;\u4e0d\u540c GPU \u8d1f\u8d23\u4e0d\u540c\u4e13\u5bb6&#xff0c;\u901a\u8fc7 all-to-all \u901a\u4fe1\u4ea4\u6362 token\u3002<\/li>\n<li>TP \u4ec5\u5728\u5355 GPU \u663e\u5b58\u4e0d\u8db3\u65f6\u4f7f\u7528&#xff08;\u6bcf\u5f20 H200 \u5269\u4f59\u7ea6 34GB \u663e\u5b58\u65f6&#xff0c;Wide-EP \u66f4\u4f18&#xff09;\u3002<\/li>\n<\/ul>\n<h4>5.2 \u53cc\u6279\u6b21\u91cd\u53e0\u4e0e\u4e13\u5bb6\u8d1f\u8f7d\u5747\u8861<\/h4>\n<p>vLLM \u96c6\u6210\u4e86\u4e24\u9879\u6765\u81ea DeepSeek \u7684\u5173\u952e\u4f18\u5316&#xff1a;<\/p>\n<p>\u53cc\u6279\u6b21\u91cd\u53e0&#xff08;DBO, &#8211;enable-dbo&#xff09;&#xff1a;\u5c06\u6bcf\u4e2a batch \u5206\u4e3a\u5fae\u6279\u6b21&#xff0c;\u91cd\u53e0\u8ba1\u7b97\u548c\u96c6\u4f53\u901a\u4fe1\u3002\u4e3b\u7ebf\u7a0b\u521b\u5efa\u5fae\u6279\u6b21\u5de5\u4f5c\u7ebf\u7a0b\u5b8c\u6210 CUDA \u56fe\u6355\u83b7&#xff0c;MoE all-to-all \u7b97\u5b50\u5728\u7b49\u5f85 GPU \u5b8c\u6210\u65f6\u51fa\u8ba9\u63a7\u5236\u6743\u3002\u5728\u901a\u4fe1\u5f00\u9500\u9ad8\u7684\u90e8\u7f72&#xff08;\u9ad8 EP \u5ea6&#xff09;\u4e2d\u663e\u8457\u63d0\u5347 GPU \u5229\u7528\u7387\u3002<\/p>\n<p>\u4e13\u5bb6\u5e76\u884c\u8d1f\u8f7d\u5747\u8861&#xff08;EPLB, &#8211;enable-eplb&#xff09;&#xff1a;\u6bcf\u6b21 MoE \u524d\u5411\u8bb0\u5f55 token \u8d1f\u8f7d&#xff0c;\u6ed1\u52a8\u7a97\u53e3\u805a\u5408\u540e&#xff0c;\u8fbe\u5230\u91cd\u5e73\u8861\u95f4\u9694\u65f6\u8ba1\u7b97\u65b0\u7684\u903b\u8f91\u5230\u7269\u7406\u4e13\u5bb6\u6620\u5c04\u5e76\u7f16\u6392\u6743\u91cd\u91cd\u7ec4&#xff08;weight shuffle&#xff09;&#xff0c;\u6574\u4e2a\u8fc7\u7a0b\u65e0\u9700\u91cd\u542f\u6a21\u578b\u3002<\/p>\n<h4>5.3 \u5b9e\u6d4b\u6027\u80fd<\/h4>\n<p>\u5728 CoreWeave H200 \u96c6\u7fa4&#xff08;InfiniBand &#043; ConnectX-7&#xff09;\u7684\u793e\u533a\u57fa\u51c6\u6d4b\u8bd5\u4e2d&#xff0c;vLLM Wide-EP \u5b9e\u73b0\u4e86\u6bcf\u5f20 H200 \u6301\u7eed\u541e\u5410 2.2k tokens\/s&#xff0c;\u76f8\u6bd4\u65e9\u671f\u7ea6 1.5k tokens\/s\/GPU \u6709\u663e\u8457\u63d0\u5347\u3002DeepSeek \u5728\u6570\u5343\u8282\u70b9\u4e0a\u751f\u4ea7\u8fd0\u884c\u5206\u79bb\u5f0f\u670d\u52a1&#xff0c;Decode \u4f7f\u7528 EP144 &#043; DP144 \u8de8 18 \u8282\u70b9&#xff0c;\u6bcf GPU \u7ba1\u7406 2 \u4e2a\u8def\u7531\u4e13\u5bb6\u548c 1 \u4e2a\u5171\u4eab\u4e13\u5bb6&#xff0c;\u6700\u5927\u5316 GroupGEMM \u5229\u7528\u7387\u3002<\/p>\n<h3>\u516d\u3001\u524d\u6cbf\u5c55\u671b<\/h3>\n<h4>6.1 NVIDIA Dynamo&#xff1a;\u5206\u5e03\u5f0f\u63a8\u7406\u670d\u52a1\u6846\u67b6<\/h4>\n<p>NVIDIA \u4e8e GTC 2025 \u53d1\u5e03\u7684 Dynamo \u662f\u9ad8\u541e\u5410\u3001\u4f4e\u5ef6\u8fdf\u7684\u5f00\u6e90\u63a8\u7406\u670d\u52a1\u6846\u67b6&#xff0c;\u5728 Blackwell \u4e0a\u8fd0\u884c DeepSeek-R1 \u65f6\u5c06\u53ef\u670d\u52a1\u8bf7\u6c42\u6570\u63d0\u5347\u81f3 30 \u500d\u3002\u5176\u56db\u5927\u6838\u5fc3\u7ec4\u4ef6\u5305\u62ec&#xff1a;<\/p>\n<ul>\n<li>Dynamo Planner&#xff1a;\u6301\u7eed\u76d1\u63a7 GPU \u5bb9\u91cf\u6307\u6807&#xff0c;\u7ed3\u5408 TTFT\/ITL \u7b49 SLO&#xff0c;\u52a8\u6001\u51b3\u5b9a\u8bf7\u6c42\u91c7\u7528\u5206\u79bb\u5f0f\u6216\u805a\u5408\u5f0f\u670d\u52a1&#xff0c;\u5e76\u5728 prefill GPU \u6210\u4e3a\u74f6\u9888\u65f6\u8ba9 decode GPU \u8f6c\u800c\u6267\u884c prefill\u3002<\/li>\n<li>Smart Router&#xff1a;\u7528 Radix Tree \u54c8\u5e0c\u8bf7\u6c42\u5e76\u5b58\u50a8 KV \u4f4d\u7f6e&#xff0c;\u8ba1\u7b97\u65b0\u8bf7\u6c42\u4e0e\u5df2\u6709 KV cache \u5757\u7684\u91cd\u53e0\u5206\u6570&#xff0c;\u7ed3\u5408\u8d1f\u8f7d\u5747\u8861\u5c06\u8bf7\u6c42\u8def\u7531\u5230\u6700\u5408\u9002 worker\u3002<\/li>\n<li>Distributed KV Cache Manager&#xff1a;\u5c06\u4f4e\u9891\u8bbf\u95ee\u7684 KV cache \u5757\u5378\u8f7d\u5230 CPU \u4e3b\u5b58\u3001SSD \u6216\u5bf9\u8c61\u5b58\u50a8&#xff0c;\u91c7\u7528 GPU \u2192 CPU \u2192 SSD \u2192 \u5bf9\u8c61\u5b58\u50a8\u7684\u5206\u5c42\u7f13\u5b58\u7b56\u7565\u3002<\/li>\n<li>NIXL&#xff1a;\u786c\u4ef6\u4e0e\u7f51\u7edc\u65e0\u5173\u7684\u4f4e\u5ef6\u8fdf\u901a\u4fe1\u5e93&#xff0c;\u652f\u6301 GPUDirect RDMA&#xff0c;\u517c\u5bb9 NVLink\u3001InfiniBand\u3001RoCE\u3001Ethernet\u3002<\/li>\n<\/ul>\n<h4>6.2 NVFP4 KV Cache \u4e0e\u538b\u7f29\u6ce8\u610f\u529b<\/h4>\n<p>NVIDIA \u63a8\u51fa\u7684 NVFP4 KV Cache \u4ee5 4-bit \u5b58\u50a8 KV \u5f20\u91cf&#xff08;attention \u524d\u53cd\u91cf\u5316\u4e3a FP8&#xff09;&#xff0c;\u5b9e\u73b0 HBM \u5360\u7528\u3001\u5185\u5b58\u5e26\u5bbd\u548c\u541e\u5410\u7684\u663e\u8457\u63d0\u5347\u3002NVFP4 \u56e0\u66f4\u7ec6\u7c92\u5ea6\u7684\u5757\u7f29\u653e\u548c E4M3 FP8 \u7f29\u653e\u56e0\u5b50&#xff0c;\u7cbe\u5ea6\u4f18\u4e8e MXFP4&#xff0c;\u5728\u591a\u667a\u80fd\u4f53\u548c MoE \u90e8\u7f72\u4e2d\u8868\u73b0\u7a81\u51fa\u3002<\/p>\n<p>\u5b66\u672f\u754c\u4e5f\u5728\u6301\u7eed\u63a2\u7d22&#xff1a;Zipage&#xff08;2026&#xff09;\u5728 PagedAttention \u57fa\u7840\u4e0a\u5f15\u5165 KV cache \u9a71\u9010\u7b56\u7565&#xff1b;PagedEviction&#xff08;2025&#xff09;\u63d0\u51fa\u4e0e\u5206\u9875\u7ed3\u6784\u534f\u540c\u7684\u7ed3\u6784\u5316\u9010\u5757 token \u9a71\u9010&#xff1b;TPLA&#xff08;Tensor-Parallel Latent Attention&#xff09;\u9488\u5bf9\u5206\u79bb\u5f0f\u63a8\u7406\u8de8\u8bbe\u5907\u5206\u7247\u6f5c\u53d8\u91cf\u8868\u793a&#xff0c;\u4fdd\u7559\u538b\u7f29 KV cache \u4f18\u52bf\u540c\u65f6\u89e3\u9501 TP \u6548\u7387\u3002<\/p>\n<h4>6.3 \u8d85\u5927\u89c4\u6a21\u8bad\u7ec3\u57fa\u7840\u8bbe\u65bd<\/h4>\n<p>\u63a8\u7406\u57fa\u7840\u8bbe\u65bd\u7684\u6f14\u8fdb\u4e0e\u8bad\u7ec3\u57fa\u7840\u8bbe\u65bd\u5bc6\u4e0d\u53ef\u5206\u3002NVIDIA GB200 NVL72 \u5355\u673a\u67dc\u96c6\u6210 72 \u9897 Blackwell GPU \u548c 36 \u9897 Grace CPU&#xff0c;\u6db2\u51b7\u6ee1\u8f7d\u529f\u8017\u7ea6 130kW&#xff0c;\u805a\u5408 NVLink \u5e26\u5bbd 130 TB\/s&#xff0c;FP4 \u7a00\u758f\u7b97\u529b\u8d85 1.4 exaFLOPS\u3002\u5728 MLPerf Training v5.0 \u4e2d&#xff0c;512 \u9897 Blackwell GPU \u5c06 Llama 3.1 405B \u9884\u8bad\u7ec3\u65f6\u95f4\u4ece Hopper \u7684 269 \u5206\u949f\u7f29\u77ed\u81f3 121 \u5206\u949f&#xff0c;\u52a0\u901f 2.2 \u500d&#xff0c;\u5cf0\u503c\u8bad\u7ec3\u541e\u5410\u8fbe 1960 TFLOPS\u3002<\/p>\n<p>\u66f4\u6781\u7aef\u7684\u6848\u4f8b\u662f xAI \u7684 Colossus 2 \u96c6\u7fa4\u2014\u2014\u642d\u8f7d 555,000 \u5f20 GPU&#xff08;GB200 &#043; GB300&#xff09;&#xff0c;\u529f\u8017\u7ea6 1GW&#xff0c;\u4e09\u5c42\u6db2\u51b7\u7cfb\u7edf\u6bcf\u79d2\u5faa\u73af 40 \u5428\u51b7\u5374\u6db2&#xff0c;SemiAnalysis \u8bc4\u4f30\u5176\u4e3a 2025 Q3 \u4e16\u754c\u6700\u5927\u5355\u4f53\u6570\u636e\u4e2d\u5fc3\u3002<\/p>\n<h3>\u4e03\u3001\u603b\u7ed3<\/h3>\n<p>\u5927\u6a21\u578b\u63a8\u7406\u67b6\u6784\u7684\u6f14\u8fdb\u53ef\u4ee5\u5f52\u7eb3\u4e3a\u4e09\u4e2a\u5c42\u6b21&#xff1a;<\/p>\n<p>\u7b2c\u4e00\u5c42&#xff1a;\u663e\u5b58\u7ba1\u7406\u3002 PagedAttention \u7528\u865a\u62df\u5185\u5b58\u5206\u9875\u601d\u60f3\u89e3\u51b3\u4e86 KV Cache \u7684\u788e\u7247\u5316\u95ee\u9898&#xff0c;\u5c06\u663e\u5b58\u5229\u7528\u7387\u4ece 20%-40% \u63d0\u5347\u81f3\u63a5\u8fd1 100%\u3002<\/p>\n<p>\u7b2c\u4e8c\u5c42&#xff1a;\u8c03\u5ea6\u4f18\u5316\u3002 \u8fde\u7eed\u6279\u5904\u7406\u901a\u8fc7\u9010\u6b65\u52a8\u6001\u63d2\u5165\u548c\u79fb\u9664\u8bf7\u6c42&#xff0c;\u6d88\u9664\u4e86\u9759\u6001\u6279\u5904\u7406\u7684\u957f\u5c3e\u7b49\u5f85\u95ee\u9898&#xff0c;\u4f7f GPU \u5728\u7a81\u53d1\u6d41\u91cf\u4e0b\u4fdd\u6301\u9ad8\u5229\u7528\u7387\u3002<\/p>\n<p>\u7b2c\u4e09\u5c42&#xff1a;\u67b6\u6784\u5206\u79bb\u3002 Prefill-Decode \u5206\u79bb\u67b6\u6784\u4ece\u6839\u672c\u4e0a\u89e3\u9664\u4e86\u4e24\u9636\u6bb5\u8ba1\u7b97\u7279\u6027\u7684\u8d44\u6e90\u8026\u5408&#xff0c;\u901a\u8fc7\u72ec\u7acb\u8d44\u6e90\u5206\u914d\u3001KV Cache \u4f20\u8f93\u6d41\u6c34\u7ebf\u548c\u5206\u5e03\u5f0f\u7f13\u5b58\u7ba1\u7406&#xff0c;\u5c06 Goodput \u63d0\u5347\u6570\u500d\u3002<\/p>\n<p>\u622a\u81f3 2025 \u5e74\u5e95&#xff0c;\u5206\u79bb\u5f0f\u63a8\u7406\u5df2\u4ece\u5b66\u672f\u7814\u7a76\u8fdb\u5165\u751f\u4ea7\u4e3b\u6d41\u2014\u2014vLLM\u3001SGLang\u3001TensorRT-LLM\u3001NVIDIA Dynamo \u5747\u5185\u7f6e\u5206\u79bb\u5f0f\u670d\u52a1\u6a21\u5f0f&#xff0c;DeepSeek\u3001Fireworks AI\u3001Perplexity\u3001Meta\u3001Amazon \u5747\u8fd0\u884c\u81ea\u6709\u5206\u79bb\u7cfb\u7edf\u3002DistServe \u4f5c\u8005\u5728 2025 \u5e74 11 \u6708\u7684\u56de\u987e\u4e2d\u5766\u8a00&#xff1a;\u201c\u51e0\u4e4e\u6bcf\u4e2a\u751f\u4ea7\u7ea7 LLM \u670d\u52a1\u6846\u67b6\u90fd\u57fa\u4e8e\u5206\u79bb\u5f0f\u67b6\u6784\u3002\u201d<\/p>\n<p>\u5bf9\u4e8e\u6280\u672f\u9009\u578b\u800c\u8a00&#xff1a;\u5feb\u901f\u4e0a\u7ebf\u548c\u5e7f\u6cdb\u6a21\u578b\u652f\u6301\u9996\u9009 vLLM&#xff1b;\u541e\u5410\u81f3\u4e0a\u7684\u5355\u4e00\u6a21\u578b\u957f\u671f\u751f\u4ea7\u573a\u666f\u9009 TensorRT-LLM&#xff1b;\u5171\u4eab\u524d\u7f00\u5de5\u4f5c\u8d1f\u8f7d&#xff08;\u804a\u5929\u673a\u5668\u4eba\u3001RAG\u3001\u591a\u8f6e\u5bf9\u8bdd&#xff09;\u9009 SGLang\u3002\u800c\u65e0\u8bba\u9009\u62e9\u54ea\u4e2a\u5f15\u64ce&#xff0c;Prefill-Decode \u5206\u79bb\u90fd\u5df2\u6210\u4e3a\u4e0d\u53ef\u5ffd\u89c6\u7684\u67b6\u6784\u9009\u9879\u3002<\/p>\n<hr \/>\n<p>\u672c\u6587\u6570\u636e\u6765\u6e90\u5305\u62ec vLLM \u5b98\u65b9\u535a\u5ba2\u3001NVIDIA Developer Blog\u3001MLPerf \u57fa\u51c6\u6d4b\u8bd5\u62a5\u544a\u3001DistServe\/Splitwise\/Mooncake \u8bba\u6587\u53ca SemiAnalysis \u884c\u4e1a\u5206\u6790&#xff0c;\u65f6\u95f4\u8de8\u5ea6\u8986\u76d6 2024-2026 \u5e74\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u5927\u6a21\u578b\u63a8\u7406\u67b6\u6784\u6f14\u8fdb&#xff1a;\u4ece PagedAttention \u5230 Prefill-Decode \u5206\u79bb\u7684\u6df1\u5ea6\u89e3\u6790\u672c\u6587\u4ece\u5927\u6a21\u578b\u63a8\u7406\u7684\u4e24\u9636\u6bb5\u8ba1\u7b97\u7279\u6027\u51fa\u53d1&#xff0c;\u7cfb\u7edf\u62c6\u89e3 PagedAttention\u3001\u8fde\u7eed\u6279\u5904\u7406\u7684\u6838\u5fc3\u539f\u7406&#xff0c;\u6df1\u5165\u5206\u6790 Prefill-Decode \u5206\u79bb\u67b6\u6784&#xff08;DistServe\u3001Splitwise\u3001Mooncake&#xff09;\u7684\u8bbe\u8ba1\u54f2\u5b66&#xff0c;\u5e76\u7ed3\u5408 vLLM Wide-EP \u4e0e DeepSeek \u751f\u4ea7\u90e8\u7f72\u6848\u4f8b&amp;#x<\/p>\n","protected":false},"author":2,"featured_media":91735,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[10466,10465,10467,132],"topic":[],"class_list":["post-91739","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-pagedattention","tag-pd","tag-prefill-decode","tag-132"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v20.3 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u4ece PagedAttention \u5230 Prefill-Decode \u5206\u79bb\u7684\u6df1\u5ea6\u89e3\u6790 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/91739.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u4ece PagedAttention \u5230 Prefill-Decode \u5206\u79bb\u7684\u6df1\u5ea6\u89e3\u6790 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u5927\u6a21\u578b\u63a8\u7406\u67b6\u6784\u6f14\u8fdb&#xff1a;\u4ece PagedAttention \u5230 Prefill-Decode \u5206\u79bb\u7684\u6df1\u5ea6\u89e3\u6790\u672c\u6587\u4ece\u5927\u6a21\u578b\u63a8\u7406\u7684\u4e24\u9636\u6bb5\u8ba1\u7b97\u7279\u6027\u51fa\u53d1&#xff0c;\u7cfb\u7edf\u62c6\u89e3 PagedAttention\u3001\u8fde\u7eed\u6279\u5904\u7406\u7684\u6838\u5fc3\u539f\u7406&#xff0c;\u6df1\u5165\u5206\u6790 Prefill-Decode \u5206\u79bb\u67b6\u6784&#xff08;DistServe\u3001Splitwise\u3001Mooncake&#xff09;\u7684\u8bbe\u8ba1\u54f2\u5b66&#xff0c;\u5e76\u7ed3\u5408 vLLM Wide-EP \u4e0e DeepSeek \u751f\u4ea7\u90e8\u7f72\u6848\u4f8b&amp;#x\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/91739.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-08-08T03:17:08+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260808031706-6a769fb218249.png\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"4 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/91739.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/91739.html\",\"name\":\"\u4ece PagedAttention \u5230 Prefill-Decode \u5206\u79bb\u7684\u6df1\u5ea6\u89e3\u6790 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-08-08T03:17:08+00:00\",\"dateModified\":\"2026-08-08T03:17:08+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/91739.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/91739.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/91739.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u4ece PagedAttention \u5230 Prefill-Decode \u5206\u79bb\u7684\u6df1\u5ea6\u89e3\u6790\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u4ece PagedAttention \u5230 Prefill-Decode \u5206\u79bb\u7684\u6df1\u5ea6\u89e3\u6790 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/91739.html","og_locale":"zh_CN","og_type":"article","og_title":"\u4ece PagedAttention \u5230 Prefill-Decode \u5206\u79bb\u7684\u6df1\u5ea6\u89e3\u6790 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u5927\u6a21\u578b\u63a8\u7406\u67b6\u6784\u6f14\u8fdb&#xff1a;\u4ece PagedAttention \u5230 Prefill-Decode \u5206\u79bb\u7684\u6df1\u5ea6\u89e3\u6790\u672c\u6587\u4ece\u5927\u6a21\u578b\u63a8\u7406\u7684\u4e24\u9636\u6bb5\u8ba1\u7b97\u7279\u6027\u51fa\u53d1&#xff0c;\u7cfb\u7edf\u62c6\u89e3 PagedAttention\u3001\u8fde\u7eed\u6279\u5904\u7406\u7684\u6838\u5fc3\u539f\u7406&#xff0c;\u6df1\u5165\u5206\u6790 Prefill-Decode \u5206\u79bb\u67b6\u6784&#xff08;DistServe\u3001Splitwise\u3001Mooncake&#xff09;\u7684\u8bbe\u8ba1\u54f2\u5b66&#xff0c;\u5e76\u7ed3\u5408 vLLM Wide-EP \u4e0e DeepSeek \u751f\u4ea7\u90e8\u7f72\u6848\u4f8b&amp;#x","og_url":"https:\/\/www.wsisp.com\/helps\/91739.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-08-08T03:17:08+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260808031706-6a769fb218249.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"4 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/91739.html","url":"https:\/\/www.wsisp.com\/helps\/91739.html","name":"\u4ece PagedAttention \u5230 Prefill-Decode \u5206\u79bb\u7684\u6df1\u5ea6\u89e3\u6790 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-08-08T03:17:08+00:00","dateModified":"2026-08-08T03:17:08+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/91739.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/91739.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/91739.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u4ece PagedAttention \u5230 Prefill-Decode \u5206\u79bb\u7684\u6df1\u5ea6\u89e3\u6790"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/91739","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=91739"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/91739\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/91735"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=91739"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=91739"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=91739"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=91739"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}