{"id":8514,"date":"2026-09-27T09:00:31","date_gmt":"2026-09-27T01:00:31","guid":{"rendered":"https:\/\/1ow.top\/index.php\/2026\/09\/27\/%e8%b0%b7%e6%ad%8ctpu%e8%b7%91kimi%e6%af%94%e8%8b%b1%e4%bc%9f%e8%be%begpu%e5%bf%ab57%ef%bc%81%e7%94%a8%e7%9a%84%e8%bf%98%e6%98%afdeepseek%e6%8e%a8%e7%90%86%e6%a1%86%e6%9e%b6\/"},"modified":"2026-09-27T09:00:31","modified_gmt":"2026-09-27T01:00:31","slug":"%e8%b0%b7%e6%ad%8ctpu%e8%b7%91kimi%e6%af%94%e8%8b%b1%e4%bc%9f%e8%be%begpu%e5%bf%ab57%ef%bc%81%e7%94%a8%e7%9a%84%e8%bf%98%e6%98%afdeepseek%e6%8e%a8%e7%90%86%e6%a1%86%e6%9e%b6","status":"publish","type":"post","link":"https:\/\/1ow.top\/index.php\/2026\/09\/27\/%e8%b0%b7%e6%ad%8ctpu%e8%b7%91kimi%e6%af%94%e8%8b%b1%e4%bc%9f%e8%be%begpu%e5%bf%ab57%ef%bc%81%e7%94%a8%e7%9a%84%e8%bf%98%e6%98%afdeepseek%e6%8e%a8%e7%90%86%e6%a1%86%e6%9e%b6\/","title":{"rendered":"\u8c37\u6b4cTPU\u8dd1Kimi\u6bd4\u82f1\u4f1f\u8fbeGPU\u5feb57%\uff01\u7528\u7684\u8fd8\u662fDeepSeek\u63a8\u7406\u6846\u67b6"},"content":{"rendered":"
\"\u8c37\u6b4cTPU\u8dd1Kimi\u6bd4\u82f1\u4f1f\u8fbeGPU\u5feb57%\uff01\u7528\u7684\u8fd8\u662fDeepSeek\u63a8\u7406\u6846\u67b6\"\/<\/div><\/figure>\n
\"\u8c37\u6b4cTPU\u8dd1Kimi\u6bd4\u82f1\u4f1f\u8fbeGPU\u5feb57%\uff01\u7528\u7684\u8fd8\u662fDeepSeek\u63a8\u7406\u6846\u67b6\"\/<\/div><\/figure>\n
\"\u8c37\u6b4cTPU\u8dd1Kimi\u6bd4\u82f1\u4f1f\u8fbeGPU\u5feb57%\uff01\u7528\u7684\u8fd8\u662fDeepSeek\u63a8\u7406\u6846\u67b6\"\/<\/div><\/figure>\n
\"\u8c37\u6b4cTPU\u8dd1Kimi\u6bd4\u82f1\u4f1f\u8fbeGPU\u5feb57%\uff01\u7528\u7684\u8fd8\u662fDeepSeek\u63a8\u7406\u6846\u67b6\"\/<\/div><\/figure>\n

vLLM\u4eba\u9a6c\u521b\u4e1a\u516c\u53f8\u56e2\u961f\u51fa\u54c1<\/p>\n

16\u5757\u8c37\u6b4cTPU v7\u8dd1Kimi K3\uff0c\u6bcf\u79d2\u8dd1\u51fa\u4e86709\u4e2aToken\uff0c\u6bd4\u8001\u9ec4\u7684GB200\u5feb\u4e8657%\u3002<\/p>\n

\u505a\u51fa\u8fd9\u4e2a\u6210\u7ee9\u7684\uff0c\u65e2\u4e0d\u662f\u9020Kimi\u7684\u6708\u4e4b\u6697\u9762\uff0c\u4e5f\u4e0d\u662f\u9020TPU\u7684\u8c37\u6b4c\uff0c\u662f\u4e00\u5bb6\u53ebInferact\u7684\u63a8\u7406\u521b\u4e1a\u516c\u53f8\u3002<\/p>\n

Inferact\u521b\u59cb\u56e2\u961f\u5c31\u662fvLLM\u7684\u539f\u73ed\u4eba\u9a6c\uff0c\u4eca\u5e74\u62ff\u4e86a16z\u9886\u6295\u76841.5\u4ebf\u7f8e\u5143\u79cd\u5b50\u8f6e\uff0c\u4f30\u503c8\u4ebf\u7f8e\u5143\u3002<\/p>\n

\u4ed6\u4eec\u7ed9TPU\u5199\u4e86\u4e00\u79cd\u53ebmegakernel\u7684\u63a8\u7406\u5185\u6838\uff0c\u628a\u6a21\u578b\u63a8\u7406\u65f6\u539f\u672c\u8981\u8c03\u5ea6\u7684\u51e0\u767e\u4e2a\u5c0f\u7a0b\u5e8f\u710a\u6210\u4e00\u4e2a\u5927\u7a0b\u5e8f\u3002<\/p>\n

\u914d\u5408TPU\u72ec\u7279\u7684\u7247\u4e0a\u5185\u5b58\u67b6\u6784\uff0cmegakernel\u628a\u5185\u5b58\u5e26\u5bbd\u5229\u7528\u7387\u903c\u5230\u4e86\u786c\u4ef6\u7406\u8bba\u5cf0\u503c\u9644\u8fd1\u3002<\/p>\n

\u914d\u4e0aDeepSeek\u63d0\u51fa\u7684DSpark\u52a0\u901f\u6846\u67b6\uff0cK3\u5728TPU\u4e0a\u8dd1\u51fa\u4e86\u524d\u9762\u63d0\u5230\u7684\u6bcf\u79d2709 token\u7684\u6210\u7ee9\u3002<\/p>\n

\u8fd9\u5957\u4ee3\u7801\uff0c\u73b0\u5728\u5df2\u7ecf\u5f00\u6e90\u3002<\/p>\n

\u540c\u683716\u5757\u82af\u7247\uff0cTPU\u5feb57%<\/p>\n

Inferact\u628aTPU\u548c\u82f1\u4f1f\u8fbeGPU\u653e\u5728\u5b8c\u5168\u76f8\u540c\u7684\u6761\u4ef6\u4e0b\u8dd1\u4e86\u4e00\u8f6ebenchmark\uff0cTPU\u8d62\u4e86\u3002<\/p>\n

\u6d4b\u8bd5\u662f\u752816\u5757TPU v7 Ironwood\u5bf9\u963516\u5757\u82f1\u4f1f\u8fbeGB200\uff0c\u4e24\u8fb9\u90fd\u8dd1Kimi K3\uff0c\u90fd\u7528vLLM\u63a8\u7406\u5f15\u64ce\uff0c\u552f\u4e00\u7684\u53d8\u91cf\u662f\u82af\u7247\u548c\u5e95\u5c42\u7684kernel\u5b9e\u73b0\u3002<\/p>\n

\u6700\u7ec8\uff0cTPU\u8dd1\u51fa\u7684\u6210\u7ee9\u662f\u6bcf\u79d2709\u4e2atoken\uff0cGB200\u8dd1\u51fa\u6bcf\u79d2452\u4e2a\uff0cTPU\u7684\u901f\u5ea6\u5feb\u4e8657%\u3002<\/p>\n

\u8fd9\u4e2a\u901f\u5ea6\u91cc\u6709DSpark\u63a8\u6d4b\u89e3\u7801\u7684\u529f\u52b3\u3002<\/p>\n

DSpark\u662f\u7531DeepSeek\u63d0\u51fa\u7684\u63a8\u7406\u52a0\u901f\u6846\u67b6\uff0c\u5176\u539f\u7406\u662f\u8ba9\u4e00\u4e2a\u5c0f\u6a21\u578b\u5148\u5feb\u901f\u731c\u51fa\u4e00\u4e32\u5019\u9009token\uff0c\u7136\u540e\u5927\u6a21\u578b\u518d\u5bf9\u8fd9\u4e32\u5019\u9009token\u505a\u6279\u91cf\u9a8c\u8bc1\uff0c\u731c\u5bf9\u7684\u76f4\u63a5\u8df3\u8fc7\uff0c\u731c\u9519\u7684\u56de\u9000\u91cd\u6765\u3002<\/p>\n

Inferact\u5728TPU\u4e0a\u8dd1\u901a\u4e86\u8fd9\u5957\u6d41\u7a0b\uff0cacceptance length\u8fbe\u5230\u4e866\uff0c\u4e5f\u5c31\u662f\u6bcf\u8f6e\u731c\u51fa\u7684token\u91cc\uff0c\u5e73\u5747\u67096\u4e2a\u80fd\u76f4\u63a5\u901a\u8fc7\u5927\u6a21\u578b\u7684\u9a8c\u8bc1\uff0c\u5355\u6b65decode\u7684\u8017\u65f6\u5927\u7ea6\u57288.5\u6beb\u79d2\u3002<\/p>\n

\u5173\u6389\u63a8\u6d4b\u89e3\u7801\u770b\u88f8\u901f\uff0c\u5dee\u8ddd\u540c\u6837\u62c9\u5f97\u5f00\u3002<\/p>\n

\u5728batch size\u4e3a1\u7684\u60c5\u51b5\u4e0b\uff0cTPU\u6bcf\u79d2\u80fd\u8dd1249\u4e2atoken\uff0cGB200\u53ea\u6709127\u4e2a\uff0c\u5dee\u8ddd\u63a5\u8fd1\u4e00\u500d\u3002<\/p>\n

\u628abatch size\u653e\u5927\u52308\uff0cTPU\u8fbe\u5230865\u4e2atoken\u6bcf\u79d2\uff0cGB200\u53ea\u6709636\u4e2a\u3002<\/p>\n

\u53e6\u5916\u5728Qwen\u4e0a\uff0c\u4e24\u79cd\u82af\u7247\u7684\u5dee\u8ddd\u66f4\u5927\u3002<\/p>\n

Inferact\u75284\u5757TPU v7\u8dd1Qwen 3.8 27B\uff0c\u6bcf\u79d2\u8dd1\u51fa1515\u4e2atoken\uff0c\u540c\u6837\u914d\u7f6e\u7684GB200\u53ea\u8dd1\u51fa\u4e86695\u4e2a\u3002<\/p>\n

\u800c\u4e14\u8fd9\u79cd\u63d0\u901f\u5e76\u6ca1\u6709\u9020\u6210\u7cbe\u5ea6\u635f\u5931\uff0cInferact\u7528greedy decoding\u505a\u4e86\u9a8c\u8bc1\uff0cKimi K3\u5728TPU\u4e0a\u7684GPQA-Diamond\u5f97\u5206\u662f94.4%\uff0cGSM8K\u662f97.2%\uff0c\u548cGPU\u4e0a\u7684\u7ed3\u679c\u5b8c\u5168\u4e00\u81f4\u3002<\/p>\n

\u63a8\u7406\u5f15\u64ce\u4e00\u6837\uff0c\u6a21\u578b\u4e00\u6837\uff0c\u82af\u7247\u6362\u4e00\u4e0b\uff0c\u901f\u5ea6\u5c31\u5dee\u51fa\u4e8657%\u3002<\/p>\n

\u8fd9\u79cd\u7a0b\u5ea6\u7684\u5dee\u8ddd\uff0c\u6309\u7406\u8bf4\u5e94\u8be5\u80fd\u4ece\u786c\u4ef6\u89c4\u683c\u8868\u4e0a\u627e\u5230\u89e3\u91ca\uff0c\u4f46\u5b9e\u9645\u60c5\u51b5\u6070\u597d\u76f8\u53cd\u3002<\/p>\n

\u4e24\u6b3e\u82af\u7247\u7684\u7b97\u529b\u4e5f\u5728\u540c\u4e00\u4e2a\u91cf\u7ea7\uff0cTPU v7\u7684HBM\u5e26\u5bbd\u662f7380 GB\/s\uff0c\u800cGB200\u7684HBM\u5e26\u5bbd\u53cd\u800c\u66f4\u9ad8\uff0c\u8fbe\u52308000 GB\/s\uff0c\u5e26\u5bbd\u66f4\u5927\u7684\u90a3\u5757\u82af\u7247\uff0c\u8dd1\u51fa\u6765\u7684\u901f\u5ea6\u53cd\u800c\u66f4\u4f4e\u3002<\/p>\n

\u8fd9\u79cd\u901f\u5ea6\u7684\u5dee\u8ddd\uff0c\u5b9e\u9645\u4e0a\u6765\u6e90\u4e8e\u82af\u7247\u4e0a\u9762\u8dd1\u7684\u90a3\u5c42\u8f6f\u4ef6\u3002<\/p>\n

\u628a\u51e0\u767e\u4e2a\u5c0fkernel\uff0c\u710a\u6210\u4e00\u4e2a\u5927\u7a0b\u5e8f<\/p>\n

Inferact\u7528\u4e00\u4e2a\u53ebmegakernel\u7684\u65b9\u6848\u89e3\u51b3\u4e86TPU\u4e0a\u7684\u63a8\u7406\u6548\u7387\u95ee\u9898\uff0c\u6838\u5fc3\u601d\u8def\uff0c\u662f\u628a\u6a21\u578b\u63a8\u7406\u65f6\u539f\u672c\u8981\u8c03\u5ea6\u7684\u51e0\u767e\u4e2a\u72ec\u7acb\u5c0f\u7a0b\u5e8f\uff0c\u5408\u5e76\u6210\u4e00\u4e2a\u5927\u7a0b\u5e8f\uff0c\u4ece\u800c\u6d88\u9664\u7a0b\u5e8f\u4e4b\u95f4\u5207\u6362\u65f6\u7684\u5e26\u5bbd\u6d6a\u8d39\u3002<\/p>\n

\u5927\u6a21\u578b\u63a8\u7406\u7684\u901f\u5ea6\u74f6\u9888\u4e0d\u5728\u8ba1\u7b97\uff0c\u5728\u6570\u636e\u642c\u8fd0\u3002<\/p>\n

\u6bcf\u751f\u6210\u4e00\u4e2atoken\uff0c\u6a21\u578b\u7684\u5168\u90e8\u6743\u91cd\u90fd\u8981\u4eceHBM\u4e3b\u5b58\u642c\u8fdb\u82af\u7247\u5185\u90e8\u7684\u9ad8\u901f\u7f13\u5b58\u91cc\u7b97\u4e00\u904d\uff0c\u7b97\u5b8c\u8fd9\u4e00\u8f6e\uff0c\u4e0b\u4e00\u4e2atoken\u518d\u91cd\u65b0\u642c\u4e00\u904d\u3002<\/p>\n

\u4f20\u7edf\u7684\u505a\u6cd5\u662f\u628a\u4e00\u6b21\u63a8\u7406\u62c6\u6210\u51e0\u767e\u4e2a\u72ec\u7acb\u7684kernel\uff0c\u6bcf\u4e2akernel\u8d1f\u8d23\u4e00\u5c0f\u5757\u8ba1\u7b97\u3002<\/p>\n

\u8fd9\u4e9bkernel\u6392\u7740\u961f\u4e00\u4e2a\u63a5\u4e00\u4e2a\u5730\u6267\u884c\uff0c\u4f46\u95ee\u9898\u51fa\u4f9d\u7136\u4f1a\u5728\u4ea4\u63a5\u7684\u95f4\u9699\u3002<\/p>\n

\u4e0a\u4e00\u4e2akernel\u5e72\u5b8c\u4e86\uff0c\u4e0b\u4e00\u4e2a\u8fd8\u6ca1\u6709\u542f\u52a8\u8d77\u6765\uff0c\u8fd9\u6bb5\u65f6\u95f4\u91cc\u5185\u5b58\u5e26\u5bbd\u5c31\u8fd9\u4e48\u7a7a\u8f6c\u7740\u3002<\/p>\n

CUDA Graphs\u548c\u82f1\u4f1f\u8fbe\u6700\u65b0\u7684PDL\u6280\u672f\u80fd\u7f29\u77ed\u8fd9\u4e2a\u95f4\u9699\uff0c\u4f46kernel\u4e4b\u95f4\u7684\u8fb9\u754c\u8fd8\u5728\u3002<\/p>\n

Inferact\u7684megakernel\u76f4\u63a5\u6d88\u9664\u4e86\u8fd9\u4e2a\u8fb9\u754c\u3002<\/p>\n

Kimi K3\u4e00\u5171\u670992\u5c42MoE\u8ba1\u7b97\uff0cInferact\u628a\u8fd992\u5c42\u7684\u8ba1\u7b97\u903b\u8f91\u4ece\u5934\u5230\u5c3e\u585e\u8fdb\u4e86\u4e00\u4e2aPallas\u7a0b\u5e8f\u91cc\uff0c\u4e00\u6b21\u8c03\u7528\u5c31\u8d70\u5b8c\u6574\u4e2a\u6a21\u578b\u7684\u524d\u5411\u4f20\u64ad\u3002<\/p>\n

\u8fb9\u754c\u6d88\u5931\u4e4b\u540e\uff0c\u8de8\u5c42\u6743\u91cd\u9884\u53d6\u5c31\u53d8\u5f97\u987a\u7406\u6210\u7ae0\u4e86\u3002<\/p>\n

\u7b2cN\u5c42\u8fd8\u5728\u7b97MoE\u7684\u65f6\u5019\uff0c\u7b2cN+1\u5c42\u7684attention\u6743\u91cd\u5df2\u7ecf\u5f00\u59cb\u4eceHBM\u5f80\u7247\u4e0a\u7f13\u5b58\u642c\u4e86\u3002<\/p>\n

\u8ba1\u7b97\u548c\u6570\u636e\u642c\u8fd0\u540c\u65f6\u8fdb\u884c\uff0c\u5185\u5b58\u5e26\u5bbd\u51e0\u4e4e\u6ca1\u6709\u7a7a\u8f6c\u7684\u65f6\u523b\u3002<\/p>\n

TPU\u7684\u786c\u4ef6\u7279\u6027\u8ba9\u8fd9\u79cd\u7f16\u6392\u505a\u8d77\u6765\u683c\u5916\u987a\u7545\u3002<\/p>\n

TPU v7\u7684\u6bcf\u4e2aTensorCore\u5e26\u670964 MiB\u7684VMEM\u7247\u4e0a\u5185\u5b58\uff0c\u8fd9\u5757\u5185\u5b58\u7684\u751f\u547d\u5468\u671f\u5b8c\u5168\u4ea4\u7ed9\u8f6f\u4ef6\u6765\u7ba1\u7406\uff0c\u5de5\u7a0b\u5e08\u53ef\u4ee5\u7cbe\u786e\u63a7\u5236\u4ec0\u4e48\u65f6\u5019\u5f80\u91cc\u9762\u88c5\u4ec0\u4e48\u6570\u636e\u3001\u4ec0\u4e48\u65f6\u5019\u628a\u7a7a\u95f4\u817e\u51fa\u6765\u3002<\/p>\n

64 MiB\u7684\u5bb9\u91cf\u8db3\u591f\u540c\u65f6\u88c5\u4e0b\u5f53\u524d\u5c42\u7684\u8ba1\u7b97\u6570\u636e\u548c\u4e0b\u4e00\u5c42\u9884\u53d6\u7684\u6743\u91cd\u3002<\/p>\n

GPU\u90a3\u8fb9\u60c5\u51b5\u4e0d\u540c\uff0c\u82f1\u4f1f\u8fbeBlackwell\u67b6\u6784\u7684\u7247\u4e0a\u5185\u5b58\u5206\u6563\u5728152\u4e2aSM\u91cc\u9762\uff0c\u603b\u91cf\u5927\u7ea638 MiB\uff0c\u7531\u786c\u4ef6\u81ea\u52a8\u8c03\u5ea6\uff0c\u8981\u505a\u7c7b\u4f3c\u7684\u8de8\u5c42\u7f16\u6392\u9650\u5236\u66f4\u591a\u3002<\/p>\n

Inferact\u7528Pallas\u8bed\u8a00\u624b\u5199\u4e86\u6574\u4e2amegakernel\u7684\u4ee3\u7801\uff0cPallas\u662f\u8c37\u6b4c\u7ed9TPU\u505a\u7684\u5e95\u5c42\u7f16\u7a0b\u8bed\u8a00\uff0c\u89d2\u8272\u7c7b\u4f3c\u82f1\u4f1f\u8fbeGPU\u4e0a\u7684CUDA\u3002<\/p>\n

TPU\u9ed8\u8ba4\u7684\u7f16\u8bd1\u5de5\u5177\u94feXLA\u64c5\u957f\u4f18\u5316\u5355\u5c42\u8ba1\u7b97\uff0c\u4f46\u8de892\u5c42\u534f\u8c03\u6570\u636e\u642c\u8fd0\u7684\u51b3\u7b56\u5206\u652f\u592a\u591a\uff0cXLA\u7684\u81ea\u52a8\u8c03\u5ea6\u627e\u4e0d\u5230\u6700\u4f18\u89e3\u3002<\/p>\n

Inferact\u7684\u505a\u6cd5\u662f\u7ed5\u8fc7XLA\uff0c\u7528Pallas\u8bed\u8a00\u624b\u5199\u6574\u4e2amegakernel\u7684\u4ee3\u7801\uff0c\u7531\u5de5\u7a0b\u5e08\u81ea\u5df1\u6765\u51b3\u5b9a\u6bcf\u4e00\u6b65\u642c\u4ec0\u4e48\u6570\u636e\u3001\u5b58\u5728\u54ea\u5757\u7f13\u5b58\u3001\u4ec0\u4e48\u65f6\u5019\u91ca\u653e\u3002<\/p>\n

\u8fd9\u6837\u505a\u8fd8\u5e26\u6765\u4e86\u4e00\u4e2a\u989d\u5916\u597d\u5904\uff0c\u90a3\u5c31\u662f\u7f16\u8bd1\u901f\u5ea6\u5927\u5e45\u63d0\u5347\uff0c\u8017\u65f6\u4eceXLA\u768430\u5206\u949f\u4ee5\u4e0a\u964d\u5230\u4e86\u4e0d\u523090\u79d2\uff0c\u5de5\u7a0b\u5e08\u6539\u5b8c\u4e00\u7248kernel\u4e00\u5206\u534a\u949f\u5c31\u80fd\u4e0a\u673a\u9a8c\u8bc1\u3002<\/p>\n

\u76ee\u524d\u8fd9\u5957megakernel\u662f\u9488\u5bf9Kimi K3\u7684\u6a21\u578b\u7ed3\u6784\u505a\u7684\u5b9a\u5236\u4f18\u5316\uff0c\u5982\u679c\u6362\u4e00\u4e2a\u6a21\u578b\u67b6\u6784\u8fd8\u9700\u8981\u91cd\u65b0\u9002\u914d\u3002<\/p>\n

Inferact\u5728\u535a\u5ba2\u91cc\u63d0\u5230\uff0c\u56e2\u961f\u63a5\u4e0b\u6765\u4f1a\u628amegakernel\u7684\u652f\u6301\u6269\u5c55\u5230\u66f4\u591a\u6a21\u578b\u67b6\u6784\u4e0a\u3002<\/p>\n

vLLM\u539f\u73ed\u4eba\u9a6c\u7684\u521b\u4e1a\u516c\u53f8<\/p>\n

Inferact\u53bb\u5e7411\u6708\u6210\u7acb\uff0c\u56e2\u961f\u51e0\u4e4e\u5c31\u662fvLLM\u7684\u539f\u59cb\u5f00\u53d1\u8005\uff0c\u4e00\u7fa4\u9760\u7ed9\u82f1\u4f1f\u8fbeGPU\u5199\u63a8\u7406\u5f15\u64ce\u6210\u540d\u7684\u4eba\u3002<\/p>\n

\u4ed6\u4eec\u4eca\u5e74\u4e00\u6708\u51fa\u6765\u521b\u4e1a\uff0c\u8f6c\u5934\u628aTPU\u7684\u63a8\u7406\u901f\u5ea6\u505a\u5230\u4e86GPU\u524d\u9762\u3002<\/p>\n

vLLM\u662f\u76ee\u524d\u5f00\u6e90\u63a8\u7406\u5f15\u64ce\u9886\u57df\u7684\u91cd\u91cf\u7ea7\u6846\u67b6\uff0c\u652f\u6301\u8d85\u8fc7500\u79cd\u6a21\u578b\u67b6\u6784\uff0c\u793e\u533a\u8d21\u732e\u8005\u8d85\u8fc72000\u4eba\uff0cMeta\u3001Google\u3001Character.ai\u90fd\u62ff\u5b83\u6765\u505a\u7ebf\u4e0a\u7684\u63a8\u7406\u670d\u52a1\u3002<\/p>\n

CEO Simon Mo\u662fvLLM\u9879\u76ee\u7684\u539f\u59cb\u7ef4\u62a4\u8005\uff0c\u4f2f\u514b\u5229EECS\u6bd5\u4e1a\uff0c\u4e4b\u524d\u5728Anyscale\u5de5\u4f5c\u3002<\/p>\n

\u8054\u5408\u521b\u59cb\u4ebaWoosuk Kwon\u662f\u6574\u4e2avLLM\u9879\u76ee\u7684\u53d1\u8d77\u4eba\uff0c\u4f2f\u514b\u5229\u8ba1\u7b97\u673a\u79d1\u5b66\u535a\u58eb\uff0c\u5bfc\u5e08\u662fIon Stoica\u3002<\/p>\n

Kwon\u63d0\u51fa\u7684PagedAttention\u7b97\u6cd5\u89e3\u51b3\u4e86GPU\u663e\u5b58\u7ba1\u7406\u4e2d\u7684\u788e\u7247\u5316\u95ee\u9898\uff0c\u8fd9\u4e2a\u7b97\u6cd5\u81f3\u4eca\u4ecd\u7136\u662fvLLM\u7684\u6838\u5fc3\u6280\u672f\u3002<\/p>\n

\u9996\u5e2d\u79d1\u5b66\u5bb6\u6e38\u51ef\u8d85\u66fe\u83b7\u6e05\u534e\u5927\u5b66\u7279\u7b49\u5956\u5b66\u91d1\uff0c\u4ed6\u5728vLLM\u4e2d\u4e3b\u5bfc\u4e86\u5206\u5e03\u5f0f\u63a8\u7406\u529f\u80fd\u7684\u5f00\u53d1\u3002<\/p>\n

Inferact\u4eca\u5e74\u5b8c\u6210\u4e861.5\u4ebf\u7f8e\u5143\u7684\u79cd\u5b50\u8f6e\u878d\u8d44\uff0c\u516c\u53f8\u4f30\u503c8\u4ebf\u7f8e\u5143\u3002a16z\u548cLightspeed\u9886\u6295\uff0c\u771f\u683c\u3001\u7ea2\u6749\u3001Altimeter\u8ddf\u6295\u3002<\/p>\n

\u8fd9\u6b21\u7684TPU megakernel\u6765\u81eaInferact\u4e0eGoogle Cloud\u7684\u4e00\u9879\u8054\u5408\u5de5\u7a0b\u5408\u4f5c\u3002<\/p>\n

\u53cc\u65b9\u7684\u76ee\u6807\u662f\u8ba9TPU\u6210\u4e3avLLM\u7684\u4e00\u6d41\u652f\u6301\u5bf9\u8c61\uff0c\u6240\u6709\u7684\u4f18\u5316\u6210\u679c\u90fd\u4f1a\u56de\u9988\u7ed9\u4e0a\u6e38\u7684\u5f00\u6e90\u793e\u533a\uff0ctpu-megakernels\u4ee3\u7801\u4ed3\u5e93\u662f\u8fd9\u6b21\u5408\u4f5c\u7684\u7b2c\u4e00\u4e2a\u516c\u5f00\u6210\u679c\u3002<\/p>\n

\u53c2\u8003\u94fe\u63a5\uff1a700 TPS on Kimi K3: A Case for TPU Megakernels<\/p>\n

\n

\u7248\u6743\u58f0\u660e\uff1a<\/strong><\/p>\n

\u672c\u535a\u5ba2\u90e8\u5206\u5185\u5bb9\u4e3a\u8f6c\u8f7d\u6587\u7ae0\uff0c\u65e8\u5728\u5206\u4eab\u6709\u4ef7\u503c\u7684\u4fe1\u606f\uff0c\u7248\u6743\u5f52\u539f\u4f5c\u8005\u6240\u6709\u3002<\/p>\n

\u8f6c\u8f7d\u4ec5\u4e3a\u4e2a\u4eba\u5b66\u4e60\u4e0e\u4ea4\u6d41\u76ee\u7684\uff0c\u4e0d\u5bf9\u6587\u7ae0\u89c2\u70b9\u8d1f\u8d23\uff0c\u4ea6\u4e0d\u7528\u4e8e\u4efb\u4f55\u5546\u4e1a\u7528\u9014\u3002<\/p>\n

\u5982\u6d89\u53ca\u7248\u6743\u95ee\u9898\uff0c\u8bf7\u8054\u7cfb\u672c\u4eba\u5220\u9664\u3002<\/p>\n<\/blockquote>\n","protected":false},"excerpt":{"rendered":"

vLLM\u4eba\u9a6c\u521b\u4e1a\u516c\u53f8\u56e2\u961f\u51fa\u54c1<\/p>\n

16\u5757\u8c37\u6b4cTPU v7\u8dd1Kimi K3\uff0c\u6bcf\u79d2\u8dd1\u51fa\u4e86709\u4e2aToken\uff0c\u6bd4\u8001\u9ec4\u7684GB200\u5feb\u4e8657%\u3002<\/p>\n

\u505a\u51fa\u8fd9\u4e2a\u6210\u7ee9\u7684\uff0c\u65e2\u4e0d\u662f\u9020Kimi\u7684\u6708\u4e4b\u6697\u9762\uff0c\u4e5f\u4e0d\u662f\u9020TPU\u7684\u8c37\u6b4c\uff0c\u662f\u4e00\u5bb6\u53ebInferact\u7684\u63a8\u7406\u521b\u4e1a\u516c\u53f8\u3002<\/p>\n

Inferact\u521b\u59cb\u56e2\u961f\u5c31\u662fvLLM\u7684\u539f\u73ed\u4eba\u9a6c\uff0c\u4eca\u5e74\u62ff\u4e86a16z\u9886\u6295\u76841.5\u4ebf\u7f8e\u5143\u79cd\u5b50\u8f6e\uff0c\u4f30\u503c8\u4ebf\u7f8e\u5143\u3002<\/p>\n

\u4ed6\u4eec\u7ed9TPU\u5199\u4e86\u4e00\u79cd\u53ebmegakernel\u7684\u63a8<\/p>\n","protected":false},"author":1,"featured_media":8510,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-8514","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-xwzx"],"_links":{"self":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts\/8514","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/comments?post=8514"}],"version-history":[{"count":0,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts\/8514\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/media\/8510"}],"wp:attachment":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/media?parent=8514"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/categories?post=8514"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/tags?post=8514"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}