{"id":1005,"date":"2025-08-15T09:54:00","date_gmt":"2025-08-15T01:54:00","guid":{"rendered":"http:\/\/www.applicationsofllm.com\/?p=1005"},"modified":"2025-09-17T10:01:57","modified_gmt":"2025-09-17T02:01:57","slug":"thinking-machines-lab%e9%a6%96%e6%ac%a1%e5%8f%91%e9%95%bf%e6%96%87%ef%bc%8c%e6%8f%ad%e5%bc%80llm%e6%8e%a8%e7%90%86%e4%b8%8d%e7%a1%ae%e5%ae%9a%e6%80%a7%e7%9c%9f%e7%9b%b8","status":"publish","type":"post","link":"http:\/\/www.applicationsofllm.com\/index.php\/2025\/08\/15\/thinking-machines-lab%e9%a6%96%e6%ac%a1%e5%8f%91%e9%95%bf%e6%96%87%ef%bc%8c%e6%8f%ad%e5%bc%80llm%e6%8e%a8%e7%90%86%e4%b8%8d%e7%a1%ae%e5%ae%9a%e6%80%a7%e7%9c%9f%e7%9b%b8\/","title":{"rendered":"Thinking Machines Lab\u9996\u6b21\u53d1\u957f\u6587\uff0c\u63ed\u5f00LLM\u63a8\u7406\u4e0d\u786e\u5b9a\u6027\u771f\u76f8"},"content":{"rendered":"\n

\u5c31\u5728\u4eca\u5929\uff0c\u7531 OpenAI \u524d CTO Mira Murati \u6210\u7acb\u4e8e\u4eca\u5e74 2 \u6708\u7684\u4eba\u5de5\u667a\u80fd\u521d\u521b\u516c\u53f8 Thinking Machines Lab\uff0c\u53d1\u4e86\u7b2c\u4e00\u7bc7\u6587\u7ae0 \u2014\u2014\u300a\u514b\u670d LLM \u63a8\u7406\u4e2d\u7684\u4e0d\u786e\u5b9a\u6027<\/strong>\u300b\uff08Defeating Nondeterminism in LLM Inference\uff09\u3002<\/p>\n\n\n\n

\u8fd9\u7bc7\u535a\u5ba2\u5c5e\u4e8e Thinking Machines Lab \u65b0\u63d0\u51fa\u7684\u535a\u5ba2\u680f\u76ee Connectionism\uff0c\u610f\u4e3a\u300c\u8fde\u63a5\u4e3b\u4e49\u300d\u3002\u8be5\u516c\u53f8\u8868\u793a\uff1a\u300c\u6211\u4eec\u76f8\u4fe1\uff0c\u5206\u4eab\u624d\u80fd\u8ba9\u79d1\u5b66\u66f4\u597d\u5730\u53d1\u5c55\u3002Connectionism \u5c06\u6db5\u76d6\u4e0e\u6211\u4eec\u7684\u7814\u7a76\u4e00\u6837\u5e7f\u6cdb\u7684\u4e3b\u9898\uff1a\u4ece\u6838\u51fd\u6570\u6570\u503c\u8ba1\u7b97\u5230\u63d0\u793a\u5de5\u7a0b\u3002Connectionism \u8fd9\u4e00\u540d\u79f0\u53ef\u4ee5\u8ffd\u6eaf\u5230 AI \u7684\u65e9\u671f\u5e74\u4ee3\u3002\u5b83\u66fe\u662f 20 \u4e16\u7eaa 80 \u5e74\u4ee3\u7684\u4e00\u4e2a\u7814\u7a76\u5206\u652f\uff0c\u4e13\u6ce8\u4e8e\u795e\u7ecf\u7f51\u7edc\u53ca\u5176\u4e0e\u751f\u7269\u5927\u8111\u7684\u76f8\u4f3c\u6027\u3002\u300d<\/p>\n\n\n\n

\u6b64\u5916\uff0cThinking Machines Lab \u8054\u5408\u521b\u59cb\u4eba\u3001\u8457\u540d\u6280\u672f\u535a\u4e3b\u7fc1\u8354\uff08Lilian Weng\uff09\u8fd8\u5728\u8f6c\u63a8\u4e2d\u900f\u9732\u4e86\u4e00\u4e2a\u6d88\u606f\uff0cConnection Machine\uff0c\u5373\u300c\u8fde\u63a5\u673a\u300d\uff0c\u96be\u9053\u4ed6\u4eec\u7684\u4ea7\u54c1\u8981\u6765\u4e86\uff1f<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u771f\u662f\u8ba9\u4eba\u671f\u5f85\u5462\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u5730\u5740\uff1ahttps:\/\/thinkingmachines.ai\/blog\/defeating-nondeterminism-in-llm-inference\/<\/p>\n\n\n\n

\u535a\u5ba2\u4e3b\u8981\u4f5c\u8005\u4e3a Horace He\uff0c\u8fd9\u4f4d PyTorch \u6838\u5fc3\u5f00\u53d1\u8005\u4e8e\u4eca\u5e74 3 \u6708\u4ece Meta \u79bb\u804c\uff0c\u52a0\u5165\u4e86 Thinking Machines\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u63a5\u4e0b\u6765\u770b\u535a\u5ba2\u539f\u6587\u5185\u5bb9\u3002<\/strong><\/p>\n\n\n\n

\u53ef\u590d\u73b0\u6027\uff08reproducibility\uff09\u662f\u79d1\u5b66\u8fdb\u6b65\u7684\u57fa\u77f3\u3002\u7136\u800c\uff0c\u4ece\u5927\u8bed\u8a00\u6a21\u578b\u4e2d\u83b7\u5f97\u53ef\u590d\u73b0\u7684\u7ed3\u679c\u5374\u975e\u5e38\u56f0\u96be\u3002<\/p>\n\n\n\n

\u4f8b\u5982\uff0c\u4f60\u53ef\u80fd\u4f1a\u53d1\u73b0\uff1a\u5373\u4f7f\u662f\u5411 ChatGPT \u63d0\u51fa\u540c\u4e00\u4e2a\u95ee\u9898\u591a\u6b21\uff0c\u4e5f\u53ef\u80fd\u5f97\u5230\u4e0d\u540c\u7684\u56de\u7b54\u3002\u8fd9\u672c\u8eab\u5e76\u4e0d\u4ee4\u4eba\u610f\u5916\uff0c\u56e0\u4e3a\u8bed\u8a00\u6a21\u578b\u751f\u6210\u7ed3\u679c\u7684\u8fc7\u7a0b\u6d89\u53ca\u91c7\u6837 \u2014\u2014 \u8fd9\u4e2a\u8fc7\u7a0b\u4f1a\u5c06\u6a21\u578b\u7684\u8f93\u51fa\u8f6c\u6362\u4e3a\u4e00\u4e2a\u6982\u7387\u5206\u5e03\uff0c\u5e76\u4ee5\u6982\u7387\u65b9\u5f0f\u9009\u62e9\u4e00\u4e2a token\u3002<\/p>\n\n\n\n

\u66f4\u4ee4\u4eba\u60ca\u8bb6\u7684\u662f\uff0c\u5373\u4f7f\u6211\u4eec\u5c06\u6e29\u5ea6\u53c2\u6570\u8c03\u5230 0\uff08\u7406\u8bba\u4e0a\u4f7f\u91c7\u6837\u8fc7\u7a0b\u53d8\u4e3a\u786e\u5b9a\u6027\uff09\uff0c\u5927\u8bed\u8a00\u6a21\u578b\u7684 API \u5728\u5b9e\u9645\u4e2d\u4ecd\u7136\u4e0d\u662f\u786e\u5b9a\u6027\u7684\u3002\u7814\u7a76\u8005\u5df2\u7ecf\u5bf9\u6b64\u6709\u8bf8\u591a\u8ba8\u8bba\u3002<\/p>\n\n\n\n

\u5373\u4f7f\u662f\u5728\u4f60\u81ea\u5df1\u7684\u786c\u4ef6\u4e0a\uff0c\u4f7f\u7528\u5f00\u6e90\u63a8\u7406\u5e93\uff08\u5982 vLLM \u6216 SGLang\uff09\u8fd0\u884c\u63a8\u7406\uff0c\u91c7\u6837\u8fc7\u7a0b\u4f9d\u7136\u4e0d\u662f\u786e\u5b9a\u6027\u7684\u3002<\/p>\n\n\n\n

\u4e3a\u4ec0\u4e48\u5927\u8bed\u8a00\u6a21\u578b\u7684\u63a8\u7406\u5f15\u64ce\u4e0d\u662f\u786e\u5b9a\u6027\u7684\u5462\uff1f<\/p>\n\n\n\n

\u4e00\u4e2a\u5e38\u89c1\u7684\u5047\u8bbe\u662f\uff1a\u6d6e\u70b9\u8fd0\u7b97\u7684\u975e\u7ed3\u5408\u6027\uff08non-associativity\uff09\u4e0e\u5e76\u53d1\u6267\u884c\u7684\u67d0\u79cd\u7ec4\u5408\u4f1a\u5bfc\u81f4\u4e0d\u786e\u5b9a\u6027\uff0c\u8fd9\u53d6\u51b3\u4e8e\u54ea\u4e2a\u5e76\u53d1\u6838\u5fc3\u9996\u5148\u5b8c\u6210\u3002\u6211\u4eec\u5c06\u8fd9\u79cd\u89e3\u91ca\u79f0\u4e3a\u300cLLM \u63a8\u7406\u4e0d\u786e\u5b9a\u6027\u7684\u300e\u5e76\u53d1 + \u6d6e\u70b9\u300f\u5047\u8bbe<\/strong>\u300d\u3002\u4f8b\u5982\uff0c\u4e00\u7bc7\u6700\u8fd1\u7684 arXiv \u8bba\u6587\uff08arXiv:2506.09501\uff09\u5199\u9053\uff1a<\/p>\n\n\n\n

GPU \u4e2d\u7684\u6d6e\u70b9\u8fd0\u7b97\u5177\u6709\u975e\u7ed3\u5408\u6027\uff08non-associativity\uff09\uff0c\u610f\u5473\u7740 (a+b)+c\u2260a+(b+c)\uff0c\u8fd9\u662f\u7531\u4e8e\u7cbe\u5ea6\u6709\u9650\u548c\u820d\u5165\u8bef\u5dee\u6240\u81f4\u3002\u8fd9\u4e00\u7279\u6027\u4f1a\u76f4\u63a5\u5f71\u54cd transformer \u67b6\u6784\u4e2d\u6ce8\u610f\u529b\u5f97\u5206\u548c logit \u7684\u8ba1\u7b97\uff0c\u56e0\u4e3a\u5728\u591a\u7ebf\u7a0b\u4e2d\u8fdb\u884c\u7684\u5e76\u884c\u64cd\u4f5c\uff0c\u5176\u6267\u884c\u987a\u5e8f\u4e0d\u540c\u4f1a\u5bfc\u81f4\u7ed3\u679c\u5dee\u5f02\u3002<\/p>\n\n\n\n

\u867d\u7136\u8fd9\u4e2a\u5047\u8bbe\u5e76\u4e0d\u5b8c\u5168\u9519\u8bef\uff0c\u4f46\u5b83\u5e76\u6ca1\u6709\u63ed\u793a\u4e8b\u60c5\u7684\u5168\u8c8c\u3002<\/p>\n\n\n\n

\u4f8b\u5982\uff0c\u5373\u4f7f\u5728 GPU \u4e0a\uff0c\u5bf9\u76f8\u540c\u7684\u6570\u636e\u53cd\u590d\u8fdb\u884c\u76f8\u540c\u7684\u77e9\u9635\u4e58\u6cd5\u8fd0\u7b97\uff0c\u6bcf\u6b21\u7684\u7ed3\u679c\u4e5f\u90fd\u662f\u6bcf\u4e00\u4f4d\u90fd\u76f8\u540c\u7684\u3002\u6211\u4eec\u786e\u5b9e\u5728\u4f7f\u7528\u6d6e\u70b9\u6570\uff0cGPU \u4e5f\u786e\u5b9e\u5177\u6709\u9ad8\u5ea6\u5e76\u53d1\u6027\u3002<\/p>\n\n\n\n

 \u90a3\u4e3a\u4ec0\u4e48\u5728\u8fd9\u4e2a\u6d4b\u8bd5\u4e2d\u5374\u770b\u4e0d\u5230\u4e0d\u786e\u5b9a\u6027\u5462\uff1f<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u8981\u7406\u89e3\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u4e0d\u786e\u5b9a\u6027\u7684\u771f\u6b63\u539f\u56e0\uff0c\u6211\u4eec\u5fc5\u987b\u66f4\u6df1\u5165\u5730\u63a2\u7a76\u3002<\/p>\n\n\n\n

\u4e0d\u5e78\u7684\u662f\uff0c\u751a\u81f3\u8fde\u300cLLM \u63a8\u7406\u662f\u786e\u5b9a\u6027\u300d\u7684\u8fd9\u4e00\u8bf4\u6cd5\u7684\u5b9a\u4e49\u90fd\u5f88\u96be\u660e\u786e\u3002\u6216\u8bb8\u4ee4\u4eba\u56f0\u60d1\u7684\u662f\uff0c\u4ee5\u4e0b\u8fd9\u4e9b\u770b\u4f3c\u77db\u76fe\u7684\u8bf4\u6cd5\u5b9e\u9645\u4e0a\u540c\u65f6\u90fd\u662f\u771f\u5b9e\u7684\uff1a<\/p>\n\n\n\n

    \n
  • GPU \u4e0a\u7684\u4e00\u4e9b\u6838\uff08kernel\uff09\u662f\u4e0d\u786e\u5b9a\u6027\u7684\u3002<\/li>\n\n\n\n
  • \u7136\u800c\uff0c\u8bed\u8a00\u6a21\u578b\u5728\u524d\u5411\u4f20\u64ad\u8fc7\u7a0b\u4e2d\u4f7f\u7528\u7684\u6240\u6709\u6838\u90fd\u662f\u786e\u5b9a\u6027\u7684\u3002<\/li>\n\n\n\n
  • \u6b64\u5916\uff0c\u50cf vLLM \u8fd9\u6837\u7684 LLM \u63a8\u7406\u670d\u52a1\u5668\u7684\u524d\u5411\u4f20\u64ad\u8fc7\u7a0b\uff0c\u4e5f\u53ef\u4ee5\u88ab\u8ba4\u4e3a\u662f\u786e\u5b9a\u6027\u7684\u3002<\/li>\n\n\n\n
  • \u5c3d\u7ba1\u5982\u6b64\uff0c\u4ece\u4f7f\u7528\u63a8\u7406\u670d\u52a1\u5668\u7684\u4efb\u4f55\u7528\u6237\u7684\u89d2\u5ea6\u6765\u770b\uff0c\u7ed3\u679c\u5374\u662f\u4e0d\u786e\u5b9a\u6027\u7684\u3002<\/li>\n<\/ul>\n\n\n\n

    \u5728\u8fd9\u7bc7\u6587\u7ae0\u4e2d\uff0c\u6211\u4eec\u5c06\u89e3\u91ca\u4e3a\u4ec0\u4e48\u300c\u5e76\u53d1 + \u6d6e\u70b9\u300d\u5047\u8bbe\u6ca1\u6709\u8fbe\u5230\u76ee\u7684\uff0c\u63ed\u9732 LLM \u63a8\u7406\u4e0d\u786e\u5b9a\u6027\u80cc\u540e\u7684\u771f\u6b63\u7f6a\u9b41\u7978\u9996\uff0c\u5e76\u89e3\u91ca\u5982\u4f55\u514b\u670d\u4e0d\u786e\u5b9a\u6027\u5e76\u5728 LLM \u63a8\u7406\u4e2d\u83b7\u5f97\u771f\u6b63\u53ef\u91cd\u590d\u7684\u7ed3\u679c\u3002<\/p>\n\n\n\n

    \u539f\u7f6a\uff1a\u6d6e\u70b9\u6570\u7684\u975e\u7ed3\u5408\u6027<\/strong><\/p>\n\n\n\n

    \u5728\u8ba8\u8bba\u4e0d\u786e\u5b9a\u6027\u4e4b\u524d\uff0c\u6709\u5fc5\u8981\u5148\u89e3\u91ca\u4e00\u4e0b\u4e3a\u4ec0\u4e48\u5b58\u5728\u6570\u503c\u5dee\u5f02\u3002\u6bd5\u7adf\uff0c\u6211\u4eec\u901a\u5e38\u5c06\u673a\u5668\u5b66\u4e60\u6a21\u578b\u89c6\u4e3a\u9075\u5faa\u4ea4\u6362\u5f8b\u6216\u7ed3\u5408\u5f8b\u7b49\u7ed3\u6784\u6027\u89c4\u5219\u7684\u6570\u5b66\u51fd\u6570\u3002\u6211\u4eec\u7684\u673a\u5668\u5b66\u4e60\u5e93\u96be\u9053\u4e0d\u5e94\u8be5\u63d0\u4f9b\u6570\u5b66\u4e0a\u6b63\u786e\u7684\u7ed3\u679c\u5417\uff1f<\/p>\n\n\n\n

    \u7f6a\u9b41\u7978\u9996\u662f\u6d6e\u70b9\u975e\u7ed3\u5408\u6027\uff08floating-point non-associativity\uff09\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u5bf9\u4e8e\u6d6e\u70b9\u6570 a\u3001b\u3001c\uff0c\u6709\uff1a<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u8bbd\u523a\u7684\u662f\uff0c\u6b63\u662f\u6253\u7834\u7ed3\u5408\u5f8b\u8ba9\u6d6e\u70b9\u6570\u53d8\u5f97\u6709\u7528\u3002<\/p>\n\n\n\n

    \u6d6e\u70b9\u6570\u4e4b\u6240\u4ee5\u6709\u7528\uff0c\u662f\u56e0\u4e3a\u5b83\u4eec\u5141\u8bb8\u52a8\u6001\u7684\u7cbe\u5ea6\u3002\u4e3a\u4e86\u4fbf\u4e8e\u89e3\u91ca\uff0c\u6211\u4eec\u5c06\u4f7f\u7528\u5341\u8fdb\u5236\uff08\u800c\u4e0d\u662f\u4e8c\u8fdb\u5236\uff09\uff0c\u5176\u4e2d\u6d6e\u70b9\u6570\u7684\u683c\u5f0f\u4e3a\uff1a\u5c3e\u6570 * 10^ \u6307\u6570\u3002\u8fd9\u91cc\u8fd8\u5c06\u4f7f\u7528 3 \u4f4d\u6570\u5b57\u4f5c\u4e3a\u5c3e\u6570\uff0c1 \u4f4d\u6570\u5b57\u4f5c\u4e3a\u6307\u6570\u3002\uff08\u6ce8\uff1a\u5728\u8ba1\u7b97\u673a\u79d1\u5b66\u4e2d\uff0c\u5c3e\u6570\uff08mantissa\uff0c\u6216\u6709\u6548\u6570\uff09\u662f\u6d6e\u70b9\u6570\u4e2d\u7528\u6765\u8868\u793a\u7cbe\u5ea6\u7684\u90e8\u5206\uff0c\u5b83\u51b3\u5b9a\u4e86\u6570\u5b57\u7684\u6709\u6548\u6570\u5b57\u4f4d\u6570\u548c\u7cbe\u5ea6\u3002\uff09<\/p>\n\n\n\n

    \u4f8b\u5982\uff0c\u5bf9\u4e8e\u503c 3450\uff0c\u6211\u4eec\u53ef\u4ee5\u5c06\u5176\u7cbe\u786e\u8868\u793a\u4e3a 3.45*10^3\u3002\u6211\u4eec\u4e5f\u53ef\u4ee5\u5c06\u66f4\u5c0f\u7684\u503c\uff08\u4f8b\u5982 0.486\uff09\u8868\u793a\u4e3a 4.86*10^-1\u3002\u8fd9\u6837\uff0c\u6d6e\u70b9\u6570\u65e2\u53ef\u4ee5\u8868\u793a\u975e\u5e38\u5c0f\u7684\u503c\uff0c\u4e5f\u53ef\u4ee5\u8868\u793a\u975e\u5e38\u5927\u7684\u503c\u3002\u5728\u79d1\u5b66\u9886\u57df\uff0c\u6211\u4eec\u53ef\u4ee5\u8bf4\u6d6e\u70b9\u6570\u4f7f\u6211\u4eec\u80fd\u591f\u4fdd\u6301\u6709\u6548\u6570\u7684\u4e2a\u6570\u6052\u5b9a\u3002<\/p>\n\n\n\n

    \u5982\u679c\u4e24\u4e2a\u6d6e\u70b9\u6570\u7684\u6307\u6570\u76f8\u540c\uff0c\u5b83\u4eec\u7684\u52a0\u6cd5\u8fd0\u7b97\u770b\u8d77\u6765\u4e0e\u6574\u6570\u52a0\u6cd5\u7c7b\u4f3c\u3002\u4f8b\u5982\uff1a<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u4f46\u662f\uff0c\u5982\u679c\u4e24\u4e2a\u6d6e\u70b9\u6570\u7684\u6307\u6570\u4e0d\u540c\uff0c\u4f8b\u5982 1230 \u548c 23.4\uff0c\u53c8\u4f1a\u53d1\u751f\u4ec0\u4e48\u60c5\u51b5\u5462\uff1f\u7406\u8bba\u4e0a\uff0c\u5b83\u4eec\u7684\u548c\u5e94\u8be5\u662f 1253.4\u3002\u7136\u800c\uff0c\u7531\u4e8e\u6d6e\u70b9\u6570\u8fd0\u7b97\u53ea\u80fd\u4fdd\u7559 3 \u4f4d\u6709\u6548\u6570\u5b57\uff0c\u56e0\u6b64\u7ed3\u679c\u4f1a\u88ab\u820d\u5165\u4e3a 1.25\u00d710\u00b3\uff08\u6216 1250\uff09\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u8868\u793a 1230 \u9700\u8981 3 \u4f4d\u6709\u6548\u6570\u5b57\uff0c\u8868\u793a 23.4 \u4e5f\u9700\u8981 3 \u4f4d\u6709\u6548\u6570\u5b57\u3002\u4f46\u662f\uff0c\u8fd9\u4e24\u4e2a\u6570\u76f8\u52a0\u7684\u7ed3\u679c\uff081253.4\uff09\u5374\u9700\u8981 5 \u4f4d\u6709\u6548\u6570\u5b57\u624d\u80fd\u7cbe\u786e\u8868\u793a\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u7684\u6d6e\u70b9\u6570\u683c\u5f0f\u5fc5\u987b\u820d\u5f03\u6700\u540e\u4e24\u4f4d\uff0834\uff09\u3002\u67d0\u79cd\u7a0b\u5ea6\u4e0a\uff0c\u8fd9\u76f8\u5f53\u4e8e\u6211\u4eec\u5728\u76f8\u52a0\u4e4b\u524d\uff0c\u5c06\u539f\u6765\u7684 23.4 \u56db\u820d\u4e94\u5165\u4e3a 20.0\u3002<\/sup><\/p>\n\n\n\n

    \u7136\u800c\uff0c\u8fd9\u6837\u505a\u4f1a\u5bfc\u81f4\u4fe1\u606f\u4e22\u5931\u3002\u8bf7\u6ce8\u610f\uff0c\u53ea\u8981\u6211\u4eec\u5bf9\u4e24\u4e2a\u4e0d\u540c\u9636\u4f4d\uff08\u5373\u4e0d\u540c\u6307\u6570\uff09\u7684\u6d6e\u70b9\u6570\u8fdb\u884c\u52a0\u6cd5\u8fd0\u7b97\uff0c\u5c31\u4f1a\u53d1\u751f\u8fd9\u79cd\u60c5\u51b5\u3002\u800c\u5b9e\u9645\u5e94\u7528\u4e2d\uff0c\u6211\u4eec\u7ecf\u5e38\u9700\u8981\u5bf9\u4e0d\u540c\u6307\u6570\u7684\u6d6e\u70b9\u6570\u8fdb\u884c\u52a0\u6cd5\u8fd0\u7b97\u3002\u4e8b\u5b9e\u4e0a\uff0c\u5982\u679c\u6211\u4eec\u80fd\u591f\u4fdd\u8bc1\u6240\u6709\u6d6e\u70b9\u6570\u7684\u6307\u6570\u90fd\u76f8\u540c\uff0c\u90a3\u4e48\u6211\u4eec\u5b8c\u5168\u53ef\u4ee5\u53ea\u4f7f\u7528\u6574\u6570\uff01<\/p>\n\n\n\n

    \u6362\u53e5\u8bdd\u8bf4\uff0c\u6bcf\u6b21\u4ee5\u4e0d\u540c\u987a\u5e8f\u76f8\u52a0\u6d6e\u70b9\u6570\u65f6\uff0c\u7ed3\u679c\u90fd\u6709\u53ef\u80fd\u5b8c\u5168\u4e0d\u540c\u3002\u4e3e\u4e2a\u6781\u7aef\u7684\u4f8b\u5b50\uff0c\u5bf9\u4e8e\u67d0\u4e2a\u6570\u7ec4\uff0c\u6839\u636e\u52a0\u6cd5\u987a\u5e8f\u7684\u4e0d\u540c\uff0c\u5176\u6c42\u548c\u7ed3\u679c\u53ef\u80fd\u51fa\u73b0 102 \u79cd\u4e0d\u540c\u7684\u7ed3\u679c\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u867d\u7136\u8fd9\u662f\u5bfc\u81f4\u8f93\u51fa\u7ed3\u679c\u4e0d\u4e00\u81f4\u7684\u6839\u672c\u539f\u56e0\uff0c\u4f46\u5b83\u5e76\u4e0d\u80fd\u76f4\u63a5\u89e3\u91ca\u4e0d\u786e\u5b9a\u6027\u884c\u4e3a\u7684\u6765\u6e90\u3002\u5b83\u4e5f\u65e0\u6cd5\u5e2e\u52a9\u6211\u4eec\u7406\u89e3\u4e3a\u4ec0\u4e48\u6d6e\u70b9\u6570\u7684\u52a0\u6cd5\u987a\u5e8f\u4f1a\u6539\u53d8\u3001\u8fd9\u79cd\u60c5\u51b5\u5728\u4ec0\u4e48\u65f6\u5019\u53d1\u751f\u3001\u4ee5\u53ca\u6211\u4eec\u5982\u4f55\u907f\u514d\u5b83\u3002<\/p>\n\n\n\n

    \u7b54\u6848\u85cf\u5728\u6838\u51fd\u6570\uff08kernel\uff09\u7684\u5b9e\u73b0\u65b9\u5f0f\u4e2d\u3002<\/p>\n\n\n\n

    \u4e3a\u4ec0\u4e48\u6838\u51fd\u6570\u8ba1\u7b97\u4e2d\u6570\u5b57\u52a0\u6cd5\u987a\u5e8f\u5e76\u975e\u603b\u662f\u56fa\u5b9a\u7684\uff1f<\/strong><\/p>\n\n\n\n

    \u5982\u524d\u6240\u8ff0\uff0c\u89e3\u91ca\u6838\u51fd\u6570\u8ba1\u7b97\u4e2d\u6570\u5b57\u52a0\u6cd5\u987a\u5e8f\u4e0d\u4e00\u81f4\u7684\u4e00\u4e2a\u5e38\u89c1\u539f\u56e0\u662f\u300c\u5e76\u53d1\u6027 + \u6d6e\u70b9\u8fd0\u7b97\u300d\u5047\u8bbe\u3002<\/p>\n\n\n\n

    \u8be5\u5047\u8bbe\u8ba4\u4e3a\uff0c\u5982\u679c\u5e76\u53d1\u7ebf\u7a0b\u7684\u6267\u884c\u987a\u5e8f\u662f\u4e0d\u53ef\u9884\u6d4b\u7684\uff0c\u5e76\u4e14\u7d2f\u52a0\u64cd\u4f5c\u7684\u987a\u5e8f\u4f9d\u8d56\u4e8e\u5e76\u53d1\u7ebf\u7a0b\u7684\u6267\u884c\u987a\u5e8f\uff08\u4f8b\u5982\u539f\u5b50\u52a0\u6cd5 \/atomic adds\uff09\uff0c\u90a3\u4e48\u6700\u7ec8\u7684\u7d2f\u52a0\u7ed3\u679c\u4e5f\u4f1a\u53d8\u5f97\u4e0d\u53ef\u9884\u6d4b\u3002<\/p>\n\n\n\n

    \u7136\u800c\uff0c\u4ee4\u4eba\u56f0\u60d1\u7684\u662f\uff0c\u5c3d\u7ba1\u8fd9\u79cd\u73b0\u8c61\u4f1a\u5bfc\u81f4\u6838\u51fd\u6570\u8ba1\u7b97\u7ed3\u679c\u7684\u4e0d\u786e\u5b9a\u6027\uff0c\u4f46\u5e76\u53d1\u673a\u5236\uff08\u4ee5\u53ca\u539f\u5b50\u52a0\u6cd5\uff09\u5b9e\u9645\u4e0a\u4e0e\u5927\u578b\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u4e2d\u7684\u4e0d\u786e\u5b9a\u6027\u65e0\u5173\uff01<\/strong><\/p>\n\n\n\n

    \u4e3a\u4e86\u89e3\u91ca\u771f\u6b63\u7684\u7f6a\u9b41\u7978\u9996\u662f\u4ec0\u4e48\uff0c\u6211\u4eec\u9996\u5148\u9700\u8981\u4e86\u89e3\u4e3a\u4ec0\u4e48\u73b0\u4ee3 GPU \u6838\u51fd\u6570\u5f88\u5c11\u9700\u8981\u4f7f\u7528\u539f\u5b50\u52a0\u6cd5\u3002<\/p>\n\n\n\n

    \u4ec0\u4e48\u65f6\u5019\u9700\u8981\u4f7f\u7528\u539f\u5b50\u52a0\u6cd5\u64cd\u4f5c\uff1f<\/strong><\/p>\n\n\n\n

    GPU \u901a\u5e38\u4f1a\u540c\u65f6\u5728\u591a\u4e2a\u6838\u5fc3\uff08\u5373\u6d41\u5904\u7406\u5668\uff09\u4e0a\u5e76\u884c\u8fd0\u884c\u7a0b\u5e8f\u3002\u7531\u4e8e\u8fd9\u4e9b\u6838\u5fc3\u4e4b\u95f4\u6ca1\u6709\u5185\u7f6e\u540c\u6b65\u673a\u5236\uff0c\u56e0\u6b64\u5982\u679c\u5b83\u4eec\u9700\u8981\u76f8\u4e92\u901a\u4fe1\uff0c\u5c31\u4f1a\u5f88\u9ebb\u70e6\u3002\u4f8b\u5982\uff0c\u5982\u679c\u6240\u6709\u6838\u5fc3\u90fd\u9700\u8981\u5bf9\u540c\u4e00\u4e2a\u5143\u7d20\u8fdb\u884c\u7d2f\u52a0\uff0c\u5c31\u53ef\u4ee5\u4f7f\u7528\u539f\u5b50\u52a0\u6cd5\uff08\u6709\u65f6\u4e5f\u79f0\u4e3a fetch-and-add\uff09\u3002\u539f\u5b50\u52a0\u6cd5\u662f\u4e0d\u786e\u5b9a\u6027\u7684\uff0c\u7ed3\u679c\u7684\u7d2f\u52a0\u987a\u5e8f\u5b8c\u5168\u53d6\u51b3\u4e8e\u54ea\u4e2a\u6838\u5fc3\u5148\u5b8c\u6210\u8ba1\u7b97\u3002<\/p>\n\n\n\n

    \u5177\u4f53\u6765\u8bf4\uff0c\u5047\u8bbe\u4f60\u8981\u4f7f\u7528 100 \u4e2a\u6838\u5fc3\u5bf9\u4e00\u4e2a\u5305\u542b 100 \u4e2a\u5143\u7d20\u7684\u5411\u91cf\u8fdb\u884c\u6c42\u548c\uff08\u4f8b\u5982 torch.sum ()\uff09\u3002\u867d\u7136\u53ef\u4ee5\u5e76\u884c\u52a0\u8f7d\u6240\u6709 100 \u4e2a\u5143\u7d20\uff0c\u4f46\u6700\u7ec8\u6211\u4eec\u5fc5\u987b\u5c06\u7ed3\u679c\u6c47\u603b\u4e3a\u4e00\u4e2a\u503c\u3002\u4e00\u79cd\u5b9e\u73b0\u65b9\u6cd5\u662f\u4f7f\u7528\u67d0\u79cd\u539f\u5b50\u52a0\u6cd5\u64cd\u4f5c\uff0c\u786c\u4ef6\u4fdd\u8bc1\u6240\u6709\u52a0\u6cd5\u64cd\u4f5c\u90fd\u4f1a\u6267\u884c\uff0c\u4f46\u5e76\u4e0d\u4fdd\u8bc1\u6267\u884c\u987a\u5e8f\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u539f\u5b50\u52a0\u6cd5\u64cd\u4f5c\u53ef\u4ee5\u786e\u4fdd\u6bcf\u4e2a\u6838\u5fc3\u7684\u8ba1\u7b97\u7ed3\u679c\u90fd\u80fd\u6700\u7ec8\u53cd\u6620\u5728\u603b\u548c\u4e2d\u3002\u4f46\u662f\uff0c\u5b83\u5e76\u4e0d\u80fd\u4fdd\u8bc1\u8fd9\u4e9b\u7ed3\u679c\u7684\u7d2f\u52a0\u987a\u5e8f\u3002\u7d2f\u52a0\u987a\u5e8f\u5b8c\u5168\u53d6\u51b3\u4e8e\u54ea\u4e2a\u6838\u5fc3\u5148\u5b8c\u6210\u8ba1\u7b97\uff0c\u8fd9\u662f\u4e00\u79cd\u4e0d\u786e\u5b9a\u6027\u884c\u4e3a\u3002<\/p>\n\n\n\n

    \u56e0\u6b64\uff0c\u591a\u6b21\u6267\u884c\u76f8\u540c\u7684\u5e76\u884c\u7a0b\u5e8f\u53ef\u80fd\u4f1a\u4ea7\u751f\u4e0d\u540c\u7684\u7ed3\u679c<\/strong>\u3002\u8fd9\u901a\u5e38\u5c31\u662f\u4eba\u4eec\u6240\u8bf4\u7684\u4e0d\u786e\u5b9a\u6027\uff0c\u5373\uff0c\u4f7f\u7528\u5b8c\u5168\u76f8\u540c\u7684\u8f93\u5165\u6570\u636e\u6267\u884c\u4e24\u6b21\u76f8\u540c\u7684\u7a0b\u5e8f\uff0c\u4f46\u6700\u7ec8\u7ed3\u679c\u5374\u53ef\u80fd\u4e0d\u540c\u3002\u8fd9\u88ab\u79f0\u4e3a\u8fd0\u884c\u95f4\u4e0d\u786e\u5b9a\u6027\uff08run-to-run nondeterminism\uff09\uff0c\u4f8b\u5982\uff0c\u8fd0\u884c\u4e24\u6b21\u5b8c\u5168\u76f8\u540c\u7684 Python \u811a\u672c\uff0c\u5373\u4f7f\u4f9d\u8d56\u5e93\u7248\u672c\u5b8c\u5168\u76f8\u540c\uff0c\u7ed3\u679c\u4e5f\u53ef\u80fd\u4e0d\u540c\u3002<\/p>\n\n\n\n

    \u867d\u7136\u5e76\u53d1\u7684\u539f\u5b50\u52a0\u6cd5\u64cd\u4f5c\u4f1a\u4f7f\u6838\u51fd\u6570\u7684\u6267\u884c\u7ed3\u679c\u53d8\u5f97\u4e0d\u53ef\u9884\u6d4b\uff0c\u4f46\u5bf9\u4e8e\u5927\u591a\u6570\u6838\u51fd\u6570\u6765\u8bf4\uff0c\u539f\u5b50\u52a0\u6cd5\u5e76\u975e\u5fc5\u9700\u3002<\/p>\n\n\n\n

    \u4e8b\u5b9e\u4e0a\uff0c\u5728 LLM \u7684\u5178\u578b\u524d\u5411\u4f20\u64ad\u8fc7\u7a0b\u4e2d\uff0c\u901a\u5e38\u6839\u672c\u4e0d\u9700\u8981\u4f7f\u7528\u539f\u5b50\u52a0\u6cd5\u3002<\/strong>\u8fd9\u53ef\u80fd\u4ee4\u4eba\u611f\u5230\u610f\u5916\uff0c\u56e0\u4e3a\u5e76\u884c\u5316\u8ba1\u7b97\u4e2d\u7684\u5f52\u7ea6\u64cd\u4f5c\u901a\u5e38\u53ef\u4ee5\u4ece\u539f\u5b50\u52a0\u6cd5\u4e2d\u83b7\u76ca\u3002\u4f46\u5b9e\u9645\u4e0a\uff0c\u539f\u5b50\u52a0\u6cd5\u5728\u5927\u591a\u6570\u60c5\u51b5\u4e0b\u5e76\u975e\u5fc5\u9700\uff0c\u4e3b\u8981\u539f\u56e0\u6709\u4e24\u70b9\u3002<\/p>\n\n\n\n

    1. \u901a\u5e38\u60c5\u51b5\u4e0b\uff0c\u6279\u5904\u7406\u7ef4\u5ea6\u4e0a\u7684\u5e76\u884c\u6027\u5df2\u7ecf\u8db3\u591f\uff0c\u56e0\u6b64\u6211\u4eec\u65e0\u9700\u5728\u5f52\u7ea6\u7ef4\u5ea6\u4e0a\u8fdb\u884c\u5e76\u884c\u5316\u3002<\/p>\n\n\n\n

    2. \u968f\u7740\u65f6\u95f4\u7684\u63a8\u79fb\uff0c\u5927\u591a\u6570\u795e\u7ecf\u7f51\u7edc\u5e93\u90fd\u91c7\u7528\u4e86\u5404\u79cd\u7b56\u7565\uff0c\u4ee5\u5728\u4e0d\u727a\u7272\u6027\u80fd\u7684\u60c5\u51b5\u4e0b\u5b9e\u73b0\u7ed3\u679c\u7684\u53ef\u9884\u6d4b\u6027\u3002<\/p>\n\n\n\n

    \u7531\u4e8e\u4e0a\u8ff0\u4e24\u4e2a\u56e0\u7d20\uff0c\u5bf9\u4e8e\u7edd\u5927\u591a\u6570\u795e\u7ecf\u7f51\u7edc\u64cd\u4f5c\u6765\u8bf4\uff0c\u4e0d\u4f7f\u7528\u539f\u5b50\u52a0\u6cd5\u51e0\u4e4e\u4e0d\u4f1a\u5e26\u6765\u6027\u80fd\u635f\u5931\u3002<\/p>\n\n\n\n

    \u5f53\u7136\uff0c\u4ecd\u7136\u6709\u5c11\u6570\u5e38\u89c1\u64cd\u4f5c\u5728\u4e0d\u4f7f\u7528\u539f\u5b50\u52a0\u6cd5\u65f6\u4f1a\u906d\u9047\u663e\u8457\u7684\u6027\u80fd\u4e0b\u964d\u3002\u4f8b\u5982\uff0cPyTorch \u4e2d\u7684 scatter_add\uff08\u5373 a [b] += c\uff09\u3002\u4e0d\u8fc7\uff0c\u5728\u5927\u8bed\u8a00\u6a21\u578b\u4e2d\u552f\u4e00\u5e38\u7528\u4e14\u4f9d\u8d56\u539f\u5b50\u52a0\u6cd5\u7684\u64cd\u4f5c\uff0c\u662f FlashAttention \u7684\u53cd\u5411\u4f20\u64ad\uff08backward\uff09\u3002<\/p>\n\n\n\n

    \u7136\u800c\uff0cLLM \u7684\u524d\u5411\u4f20\u64ad\u8fc7\u7a0b\u4e2d\u5e76\u4e0d\u6d89\u53ca\u4efb\u4f55\u9700\u8981\u539f\u5b50\u52a0\u6cd5\u7684\u64cd\u4f5c\u3002\u56e0\u6b64\uff0cLLM \u7684\u524d\u5411\u8fc7\u7a0b\u672c\u8d28\u4e0a\u662f\u8fd0\u884c\u95f4\u786e\u5b9a\u7684\uff08\u5373\u6bcf\u6b21\u8fd0\u884c\u7ed3\u679c\u4e00\u81f4\uff09\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u7ef4\u57fa\u767e\u79d1\u4e0a\u5199\u9053\uff1a\u4e00\u4e2a\u786e\u5b9a\u6027\u7b97\u6cd5\u662f\u5728\u7ed9\u5b9a\u7279\u5b9a\u8f93\u5165\u7684\u60c5\u51b5\u4e0b\uff0c\u59cb\u7ec8\u4ea7\u751f\u76f8\u540c\u8f93\u51fa\u7684\u7b97\u6cd5\u3002\u800c\u5728\u8fd9\u91cc\uff0c\u53ea\u8981\u8f93\u5165\u5b8c\u5168\u76f8\u540c\uff08\u5373\u63a8\u7406\u670d\u52a1\u5668\u5904\u7406\u7684\u8bf7\u6c42\u5b8c\u5168\u4e00\u81f4\uff09\uff0c\u524d\u5411\u4f20\u64ad\u5c31\u603b\u662f\u4f1a\u751f\u6210\u5b8c\u5168\u76f8\u540c\u7684\u8f93\u51fa\u3002<\/p>\n\n\n\n

    \u7136\u800c\uff0c\u524d\u5411\u4f20\u64ad\u672c\u8eab\u662f\u786e\u5b9a\u6027\u7684\u5e76\u4e0d\u610f\u5473\u7740\u6574\u4e2a\u7cfb\u7edf\u4e5f\u662f\u786e\u5b9a\u6027\u7684\u3002\u6bd4\u5982\uff0c\u5982\u679c\u67d0\u4e2a\u8bf7\u6c42\u7684\u8f93\u51fa\u4f9d\u8d56\u4e8e\u5e76\u884c\u7528\u6237\u7684\u8bf7\u6c42\uff08\u4f8b\u5982 batch-norm \u8fd9\u6837\u7684\u64cd\u4f5c\uff09\uff0c\u90a3\u4e48\u7531\u4e8e\u6bcf\u4e2a\u8bf7\u6c42\u90fd\u65e0\u6cd5\u9884\u77e5\u5176\u4ed6\u5e76\u53d1\u8bf7\u6c42\u7684\u5185\u5bb9\uff0c\u4ece\u5355\u4e2a\u8bf7\u6c42\u7684\u89c6\u89d2\u6765\u770b\uff0c\u6574\u4e2a LLM \u63a8\u7406\u8fc7\u7a0b\u5c31\u4f1a\u662f\u4e0d\u786e\u5b9a\u6027\u7684\u3002<\/p>\n\n\n\n

    \u4e8b\u5b9e\u8bc1\u660e\uff0c\u6211\u4eec\u7684\u8bf7\u6c42\u8f93\u51fa\u786e\u5b9e\u4f9d\u8d56\u4e8e\u5176\u4ed6\u5e76\u53d1\u7528\u6237\u7684\u8bf7\u6c42\u3002\u4f46\u8fd9\u5e76\u4e0d\u662f\u56e0\u4e3a\u8de8 batch \u6cc4\u9732\u4e86\u4fe1\u606f\uff0c\u800c\u662f\u56e0\u4e3a\u6211\u4eec\u7684\u524d\u5411\u4f20\u64ad\u8fc7\u7a0b\u7f3a\u4e4f\u6279\u6b21\u4e0d\u53d8\u6027\uff08batch invariance\uff09\uff0c\u8fd9\u5bfc\u81f4\u540c\u4e00\u4e2a\u8bf7\u6c42\u7684\u8f93\u51fa\u4f1a\u53d7\u5230\u524d\u5411\u4f20\u64ad\u4e2d batch size\uff08batch size\uff09\u53d8\u5316\u7684\u5f71\u54cd\u3002<\/p>\n\n\n\n

    \u6279\u6b21\u4e0d\u53d8\u6027\u4e0e\u786e\u5b9a\u6027<\/strong><\/p>\n\n\n\n

    \u4e3a\u4e86\u8bf4\u660e\u4ec0\u4e48\u662f\u6279\u6b21\u4e0d\u53d8\u6027\uff0c\u6211\u4eec\u53ef\u4ee5\u7b80\u5316\u95ee\u9898\uff0c\u53ea\u5173\u6ce8\u77e9\u9635\u4e58\u6cd5\uff08matmul\uff09\u3002\u4f60\u53ef\u4ee5\u5047\u8bbe\u6240\u6709\u7684 matmul \u5b9e\u73b0\u90fd\u662f\u8fd0\u884c\u95f4\u786e\u5b9a\u7684\uff0c\u4e5f\u5c31\u662f\u8bf4\uff0c\u540c\u6837\u7684\u8f93\u5165\uff0c\u6bcf\u6b21\u8fd0\u884c\u90fd\u4f1a\u5f97\u5230\u76f8\u540c\u7684\u7ed3\u679c\u3002<\/p>\n\n\n\n

    \u4f46\u5b83\u4eec\u5e76\u4e0d\u662f\u6279\u6b21\u4e0d\u53d8\u7684\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u5f53 batch size \u53d1\u751f\u53d8\u5316\u65f6\uff0cbatch \u4e2d\u7684\u6bcf\u4e2a\u5143\u7d20\u53ef\u80fd\u4f1a\u5f97\u5230\u4e0d\u540c\u7684\u8ba1\u7b97\u7ed3\u679c\u3002<\/p>\n\n\n\n

    \u4ece\u6570\u5b66\u89d2\u5ea6\u6765\u770b\uff0c\u8fd9\u662f\u4e00\u79cd\u76f8\u5f53\u53cd\u5e38\u7684\u6027\u8d28\u3002\u7406\u8bba\u4e0a\uff0c\u77e9\u9635\u4e58\u6cd5\u5728 batch \u7ef4\u5ea6\u4e0a\u5e94\u5f53\u662f\u72ec\u7acb\u7684\uff0cbatch \u4e2d\u5176\u4ed6\u5143\u7d20\u7684\u5b58\u5728\u4e0e\u5426\uff0c\u6216 batch \u7684\u5927\u5c0f\uff0c\u90fd\u4e0d\u5e94\u5f71\u54cd\u67d0\u4e2a\u5177\u4f53\u5143\u7d20\u7684\u8ba1\u7b97\u7ed3\u679c\u3002<\/p>\n\n\n\n

    \u7136\u800c\uff0c\u6211\u4eec\u901a\u8fc7\u5b9e\u9a8c\u8bc1\u636e\u53ef\u4ee5\u53d1\u73b0\uff0c\u73b0\u5b9e\u60c5\u51b5\u5e76\u975e\u5982\u6b64\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u8bf7\u6ce8\u610f\uff0c\u8fd9\u91cc\u7684\u786e\u5b9a\u6027\u662f\u6307\u6bcf\u6b21\u8fd0\u884c\u7ed3\u679c\u90fd\u76f8\u540c\u3002\u5982\u679c\u4f60\u591a\u6b21\u8fd0\u884c\u8be5\u811a\u672c\uff0c\u5b83\u4f1a\u59cb\u7ec8\u8fd4\u56de\u76f8\u540c\u7684\u7ed3\u679c\u3002<\/p>\n\n\n\n

    \u4f46\u662f\uff0c\u5982\u679c\u5c06\u975e\u6279\u5904\u7406\u4e0d\u53d8\u7684\u6838\u51fd\u6570\u7528\u4f5c\u66f4\u5927\u63a8\u7406\u7cfb\u7edf\u7684\u4e00\u90e8\u5206\uff0c\u5219\u6574\u4e2a\u7cfb\u7edf\u53ef\u80fd\u53d8\u5f97\u4e0d\u786e\u5b9a\u6027\u3002\u5f53\u4f60\u5411\u63a8\u7406\u7aef\u70b9\u53d1\u9001\u8bf7\u6c42\u65f6\uff0c\u4ece\u7528\u6237\u89d2\u5ea6\u6765\u770b\uff0c\u670d\u52a1\u5668\u7684\u8d1f\u8f7d\u60c5\u51b5\u662f\u4e0d\u53ef\u9884\u6d4b\u7684\u3002\u8d1f\u8f7d\u51b3\u5b9a\u4e86\u6838\u51fd\u6570\u7684 batch size\uff0c\u4ece\u800c\u5f71\u54cd\u6bcf\u4e2a\u8bf7\u6c42\u7684\u6700\u7ec8\u7ed3\u679c\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u5982\u679c\u4f60\u628a\u67d0\u79cd\u6838\u51fd\u6570\u4e0d\u5177\u5907\u4e0d\u53d8\u6027\u7684\u5c5e\u6027\uff08\u4f8b\u5982\uff1abatch size\uff09\u4e0e\u8be5\u5c5e\u6027\u672c\u8eab\u7684\u4e0d\u786e\u5b9a\u6027\uff08\u4f8b\u5982\uff1a\u670d\u52a1\u5668\u8d1f\u8f7d\u60c5\u51b5\uff09\u7ec4\u5408\u5728\u4e00\u8d77\uff0c\u5c31\u4f1a\u5f97\u5230\u4e00\u4e2a\u4e0d\u786e\u5b9a\u6027\u7684\u7cfb\u7edf\u3002<\/p>\n\n\n\n

    \u6362\u53e5\u8bdd\u8bf4\uff0c\u51e0\u4e4e\u6240\u6709\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u7aef\u70b9\u4e4b\u6240\u4ee5\u662f\u4e0d\u786e\u5b9a\u7684\uff0c\u4e3b\u8981\u539f\u56e0\u5c31\u662f\u8d1f\u8f7d\uff08\u4ee5\u53ca\u7531\u6b64\u51b3\u5b9a\u7684 batch size\uff09\u672c\u8eab\u5177\u6709\u4e0d\u786e\u5b9a\u6027\uff01\u8fd9\u79cd\u4e0d\u786e\u5b9a\u6027\u5e76\u975e\u4ec5\u9650\u4e8e GPU\uff0c\u4f7f\u7528 CPU \u6216 TPU \u8fd0\u884c\u7684 LLM \u63a8\u7406\u7aef\u70b9\u4e5f\u4f1a\u5b58\u5728\u540c\u6837\u7684\u95ee\u9898\u3002\u56e0\u6b64\uff0c\u5982\u679c\u6211\u4eec\u60f3\u907f\u514d\u63a8\u7406\u670d\u52a1\u5668\u4e2d\u7684\u4e0d\u786e\u5b9a\u6027\uff0c\u5c31\u5fc5\u987b\u786e\u4fdd\u6838\u51fd\u6570\u5bf9 batch size \u5177\u6709\u4e0d\u53d8\u6027\u3002<\/p>\n\n\n\n

    \u4e3a\u4e86\u7406\u89e3\u5982\u4f55\u5b9e\u73b0\u8fd9\u4e00\u70b9\uff0c\u6211\u4eec\u9996\u5148\u9700\u8981\u4e86\u89e3\u4e3a\u4ec0\u4e48\u6838\u51fd\u6570\u9ed8\u8ba4\u60c5\u51b5\u4e0b\u5e76\u4e0d\u5177\u5907\u6279\u5904\u7406\u4e0d\u53d8\u6027\u3002<\/p>\n\n\n\n

    \u6211\u4eec\u5982\u4f55\u4f7f\u6838\u5177\u6709\u6279\u6b21\u4e0d\u53d8\u6027\uff1f<\/strong><\/p>\n\n\n\n

    \u4e3a\u4e86\u786e\u4fdd Transformer \u6a21\u578b\u7684\u5b9e\u73b0\u4e0e batch size \u65e0\u5173\uff0c\u6211\u4eec\u5fc5\u987b\u786e\u4fdd\u6a21\u578b\u4e2d\u7684\u6bcf\u4e2a\u6838\u5fc3\u6a21\u5757\u90fd\u4e0e batch size \u65e0\u5173\u3002\u5e78\u8fd0\u7684\u662f\uff0c\u6211\u4eec\u53ef\u4ee5\u5047\u8bbe\u6bcf\u4e2a\u9010\u70b9\u8fd0\u7b97\uff08pointwise operation\uff09\u90fd\u4e0e batch size \u65e0\u5173\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u53ea\u9700\u8981\u62c5\u5fc3\u6d89\u53ca\u7684 3 \u4e2a\u64cd\u4f5c\uff1aRMSNorm\u3001\u77e9\u9635\u4e58\u6cd5\u548c\u6ce8\u610f\u529b\u3002<\/p>\n\n\n\n

    \u5de7\u5408\u7684\u662f\uff0c\u8fd9\u4e9b\u64cd\u4f5c\u7684\u96be\u5ea6\u6b63\u597d\u662f\u4f9d\u6b21\u9012\u589e\u7684\u3002\u8981\u60f3\u5728\u4fdd\u6301\u5408\u7406\u6027\u80fd\u7684\u540c\u65f6\u5b9e\u73b0\u6279\u6b21\u4e0d\u53d8\u6027\uff0c\u6bcf\u4e00\u79cd\u64cd\u4f5c\u90fd\u9700\u8981\u4e00\u4e9b\u989d\u5916\u7684\u8003\u91cf\u3002\u6211\u4eec\u5148\u4ece RMSNorm \u5f00\u59cb\u8c08\u8d77\u3002<\/p>\n\n\n\n

    RMSNorm<\/strong><\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    RMSNorm \u5b9e\u73b0\u65b9\u5f0f\uff1a<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u6279\u6b21\u4e0d\u53d8\u6027\u7684\u8981\u6c42\u662f\uff0c\u65e0\u8bba\u6838\u51fd\u6570\u7684 batch size \u5982\u4f55\uff0c\u6bcf\u4e2a\u5143\u7d20\u7684\u5f52\u7ea6\u987a\u5e8f\u90fd\u5fc5\u987b\u4fdd\u6301\u4e0d\u53d8<\/strong>\u3002\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u8fd9\u5e76\u4e0d\u610f\u5473\u7740\u6211\u4eec\u5fc5\u987b\u59cb\u7ec8\u4f7f\u7528\u76f8\u540c\u7684\u5f52\u7ea6\u7b56\u7565\u3002\u4f8b\u5982\uff0c\u5373\u4f7f\u6211\u4eec\u6539\u53d8\u4e86\u8981\u8fdb\u884c\u5f52\u7ea6\u7684\u5143\u7d20\u6570\u91cf\uff0c\u53ea\u8981\u5f52\u7ea6\u987a\u5e8f\u4e0d\u53d8\uff0c\u6211\u4eec\u7684\u7b97\u6cd5\u4ecd\u7136\u53ef\u4ee5\u6ee1\u8db3\u6279\u5904\u7406\u4e0d\u53d8\u6027\u7684\u8981\u6c42\u3002<\/p>\n\n\n\n

    \u56e0\u6b64\uff0c\u53ea\u6709\u5f53 batch size \u5f71\u54cd\u5230\u5f52\u7ea6\u7b56\u7565\u65f6\uff0c\u6211\u4eec\u624d\u4f1a\u6253\u7834\u6279\u6b21\u4e0d\u53d8\u6027\u3002<\/p>\n\n\n\n

    \u8ba9\u6211\u4eec\u6765\u770b\u4e00\u4e0b RMSNorm \u7684\u6807\u51c6\u5e76\u884c\u5316\u7b56\u7565\u3002\u4e00\u822c\u6765\u8bf4\uff0c\u5e76\u884c\u7b97\u6cd5\u90fd\u4f1a\u4ece\u5c3d\u91cf\u51cf\u5c11\u6838\u5fc3\u4e4b\u95f4\u7684\u901a\u4fe1\u4e2d\u83b7\u76ca\u3002\u5728\u8fd9\u91cc\uff0c\u4e3a\u4e86\u65b9\u4fbf\u8ba8\u8bba\uff0c\u4f60\u53ef\u4ee5\u5047\u8bbe\u6211\u4eec\u6240\u8bf4\u7684\u6838\u5fc3\uff08cores\uff09\u5c31\u662f\u6307 SM\uff08Streaming Multiprocessors\uff0c\u6d41\u5904\u7406\u591a\u5904\u7406\u5668\uff09\u3002\u66f4\u5177\u4f53\u5730\u8bf4\uff0c\u8fd9\u91cc\u91cd\u8981\u7684\u6027\u8d28\u662f\uff1a\u6838\u51fd\u6570\u542f\u52a8\u7684\u7ebf\u7a0b\u5757\uff08threadblocks\uff09\u6570\u91cf\u591a\u4e8e SM \u7684\u6570\u91cf\u3002<\/p>\n\n\n\n

    \u57fa\u4e8e\u8fd9\u4e00\u70b9\uff0c\u4e00\u79cd\u53ef\u884c\u7684\u7b56\u7565\u5c31\u662f\uff1a\u5c06\u6bcf\u4e2a batch \u5143\u7d20\u5206\u914d\u7ed9\u4e00\u4e2a\u6838\u5fc3\uff0c\u5c31\u50cf\u4e0a\u56fe\u5c55\u793a\u7684\u90a3\u6837\u3002<\/p>\n\n\n\n

    \u5f53\u6211\u4eec\u589e\u52a0 batch size \u65f6\uff0c\u5e76\u4e0d\u4f1a\u5f71\u54cd\u5f52\u7ea6\u7b56\u7565\uff1b\u5982\u679c batch size = 200 \u5df2\u7ecf\u80fd\u4e3a\u6838\u51fd\u6570\u63d0\u4f9b\u8db3\u591f\u7684\u5e76\u884c\u6027\uff0c\u90a3\u4e48 batch size = 2000 \u663e\u7136\u4e5f\u540c\u6837\u80fd\u591f\u63d0\u4f9b\u8db3\u591f\u7684\u5e76\u884c\u6027\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u53e6\u4e00\u65b9\u9762\uff0c\u51cf\u5c0f batch size \u4e5f\u4f1a\u5e26\u6765\u4e00\u4e9b\u6311\u6218\u3002\u7531\u4e8e\u6211\u4eec\u4e3a\u6bcf\u4e2a\u6279\u6b21\u5143\u7d20\u5206\u914d\u4e00\u4e2a\u6838\u5fc3\uff0c\u51cf\u5c0f batch size \u4f1a\u5bfc\u81f4\u6838\u5fc3\u6570\u91cf\u5927\u4e8e\u6279\u6b21\u5143\u7d20\u6570\u91cf\uff0c\u4ece\u800c\u9020\u6210\u90e8\u5206\u6838\u5fc3\u95f2\u7f6e\u3002\u9047\u5230\u8fd9\u79cd\u60c5\u51b5\uff0c\u4f18\u79c0\u7684\u6838\u51fd\u6570\u5de5\u7a0b\u5e08\u4f1a\u91c7\u7528\u524d\u9762\u63d0\u5230\u7684\u89e3\u51b3\u65b9\u6848\u4e4b\u4e00\uff08\u539f\u5b50\u52a0\u6cd5\u6216\u5206\u6bb5\u6c42\u548c\uff09\uff0c\u4ece\u800c\u4fdd\u6301\u826f\u597d\u7684\u5e76\u884c\u6027\uff0c\u8fdb\u800c\u63d0\u5347\u6027\u80fd\u3002\u7136\u800c\uff0c\u8fd9\u4f1a\u6539\u53d8\u6c42\u548c\u7b56\u7565\uff0c\u5bfc\u81f4\u8be5\u6838\u51fd\u6570\u4e0d\u518d\u5177\u5907 batch size \u4e0d\u53d8\u7684\u7279\u6027\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u6700\u7b80\u5355\u7684\u89e3\u51b3\u65b9\u6848\u5c31\u662f\u76f4\u63a5\u5ffd\u7565\u8fd9\u4e9b\u60c5\u51b5\u3002\u8fd9\u5e76\u4e0d\u662f\u5b8c\u5168\u4e0d\u5408\u7406\u7684\uff0c\u56e0\u4e3a\u5f53 batch size \u5f88\u5c0f\u65f6\uff0c\u6838\u51fd\u6570\u901a\u5e38\u672c\u6765\u5c31\u80fd\u5f88\u5feb\u6267\u884c\uff0c\u56e0\u6b64\u5373\u4f7f\u51fa\u73b0\u4e00\u4e9b\u51cf\u901f\uff0c\u4e5f\u4e0d\u4f1a\u9020\u6210\u707e\u96be\u6027\u7684\u5f71\u54cd\u3002<\/p>\n\n\n\n

    \u5982\u679c\u6211\u4eec\u5fc5\u987b\u4f18\u5316\u8fd9\u79cd\u573a\u666f\uff0c\u4e00\u79cd\u65b9\u6cd5\u662f\uff1a\u59cb\u7ec8\u4f7f\u7528\u4e00\u79cd\u5728\u6781\u5c0f batch size \u4e0b\u4e5f\u80fd\u63d0\u4f9b\u8db3\u591f\u5e76\u884c\u5ea6\u7684\u5f52\u7ea6\u7b56\u7565\u3002\u8fd9\u6837\u7684\u7b56\u7565\u4f1a\u5728 batch size \u8f83\u5927\u65f6\u5bfc\u81f4\u8fc7\u5ea6\u5e76\u884c\uff0c\u4ece\u800c\u65e0\u6cd5\u8fbe\u5230\u5cf0\u503c\u6027\u80fd\uff0c\u4f46\u5b83\u53ef\u4ee5\u8ba9\u6211\u4eec\u5728\u6574\u4e2a batch size \u8303\u56f4\u5185\u90fd\u83b7\u5f97\u5c1a\u53ef\uff08\u867d\u7136\u4e0d\u662f\u6700\u4f73\uff09\u7684\u6027\u80fd\u8868\u73b0\u3002<\/p>\n\n\n\n

    \u6279\u6b21\u4e0d\u53d8\u77e9\u9635\u4e58\u6cd5<\/strong><\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u4ece\u672c\u8d28\u4e0a\u8bb2\uff0c\u4f60\u53ef\u4ee5\u628a\u77e9\u9635\u4e58\u6cd5\u770b\u4f5c\u662f\u4e00\u6b21\u9010\u70b9\u8fd0\u7b97\u540e\u63a5\u4e00\u6b21\u5f52\u7ea6\u3002\u90a3\u4e48\uff0c\u5982\u679c\u6211\u4eec\u901a\u8fc7\u5c06\u8f93\u51fa\u5212\u5206\u4e3a\u5c0f\u5757\u6765\u5e76\u884c\u5316\u77e9\u9635\u4e58\u6cd5\uff0c\u5c31\u80fd\u5f97\u5230\u4e00\u79cd\u7c7b\u4f3c\u7684\u6570\u636e\u5e76\u884c\u6838\u51fd\u6570\u7b56\u7565\uff0c\u4f7f\u5f97\u6bcf\u4e00\u6b21\u5f52\u7ea6\u90fd\u5728\u5355\u4e2a\u6838\u5fc3\u5185\u5b8c\u6210\u3002<\/p>\n\n\n\n

    \u4e0e RMSNorm \u7c7b\u4f3c\uff0c\u77e9\u9635\u4e58\u6cd5\u7684\u6279\u6b21\u7ef4\u5ea6\uff08M \u548c N\uff09\u4e5f\u53ef\u80fd\u53d8\u5f97\u8fc7\u5c0f\uff0c\u8feb\u4f7f\u6211\u4eec\u5fc5\u987b\u6cbf\u5f52\u7ea6\u7ef4\u5ea6\uff08K\uff09\u8fdb\u884c\u62c6\u5206\u3002\u5c3d\u7ba1\u6709\u4e24\u4e2a\u6279\u6b21\u7ef4\u5ea6\uff0c\u77e9\u9635\u4e58\u6cd5\u4ecd\u7136\u9700\u8981\u6bcf\u4e2a\u6838\u5fc3\u6709\u66f4\u591a\u7684\u5de5\u4f5c\u91cf\u624d\u80fd\u6709\u6548\u5229\u7528\u5f20\u91cf\u6838\u5fc3\u3002\u4f8b\u5982\uff0c\u5bf9\u4e8e\u4e00\u4e2a [1024, K] x [K, 1024] \u7684\u77e9\u9635\u4e58\u6cd5\u548c\u4e00\u4e2a\u6807\u51c6\u7684 [128, 128] \u4e8c\u7ef4 tile \u5927\u5c0f\uff0c\u6570\u636e\u5e76\u884c\u7b56\u7565\u6700\u591a\u53ea\u80fd\u5c06\u5176\u5206\u914d\u5230 64 \u4e2a\u6838\u5fc3\u4e0a\uff0c\u8fd9\u4e0d\u8db3\u4ee5\u4f7f GPU \u8fbe\u5230\u9971\u548c\u3002<\/p>\n\n\n\n

    \u5728\u77e9\u9635\u4e58\u6cd5\u4e2d\u6cbf\u5f52\u7ea6\u7ef4\u5ea6\u8fdb\u884c\u62c6\u5206\u88ab\u79f0\u4e3a Split-K \u77e9\u9635\u4e58\u6cd5\u3002\u4e0e RMSNorm \u7684\u60c5\u51b5\u4e00\u6837\uff0c\u4f7f\u7528\u8fd9\u79cd\u7b56\u7565\u4f1a\u7834\u574f\u6279\u6b21\u4e0d\u53d8\u6027\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u77e9\u9635\u4e58\u6cd5\u8fd8\u6709\u4e00\u4e2a\u989d\u5916\u7684\u590d\u6742\u6027\uff0c\u5373\u5f20\u91cf\u6838\u5fc3\u6307\u4ee4\u3002\u5bf9\u4e8e\u5f52\u7ea6\u64cd\u4f5c\uff0c\u6211\u4eec\u53ef\u4ee5\u4e00\u6b21\u53ea\u5904\u7406\u4e00\u884c\uff1b\u4f46\u9ad8\u6548\u7684\u77e9\u9635\u4e58\u6cd5\u6838\u51fd\u6570\u5fc5\u987b\u4e00\u6b21\u6027\u64cd\u4f5c\u4e00\u6574\u4e2a tile\u3002<\/p>\n\n\n\n

    \u6bcf\u6761\u5f20\u91cf\u6838\u5fc3\u6307\u4ee4\uff08\u4f8b\u5982 wgmma.mma_async.sync.aligned.m64n128k16\uff09\u5728\u5185\u90e8\u53ef\u80fd\u6709\u4e0d\u540c\u7684\u5f52\u7ea6\u987a\u5e8f\u3002\u9009\u62e9\u4e0d\u540c\u5f20\u91cf\u6838\u5fc3\u6307\u4ee4\u7684\u4e00\u4e2a\u539f\u56e0\u53ef\u80fd\u662f batch size \u975e\u5e38\u5c0f\u3002\u4f8b\u5982\uff0c\u5982\u679c\u6211\u4eec\u4f7f\u7528\u7684\u5f20\u91cf\u6838\u5fc3 PTX \u6307\u4ee4\u64cd\u4f5c\u7684\u662f\u4e00\u4e2a\u957f\u5ea6\u4e3a 256 \u7684 tile\uff0c\u4f46 batch size \u53ea\u6709 32\uff0c\u90a3\u6211\u4eec\u51e0\u4e4e\u6d6a\u8d39\u4e86\u6240\u6709\u7684\u8ba1\u7b97\u8d44\u6e90\uff01\u5f53 batch size \u4e3a 1 \u65f6\uff0c\u6700\u5feb\u7684\u6838\u51fd\u6570\u901a\u5e38\u6839\u672c\u4e0d\u4f7f\u7528\u5f20\u91cf\u6838\u5fc3\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u56e0\u6b64\uff0c\u786e\u4fdd\u77e9\u9635\u4e58\u6cd5\u6279\u6b21\u4e0d\u53d8\u6027\u7684\u6700\u7b80\u5355\u65b9\u6cd5\u662f\uff1a\u7f16\u8bd1\u4e00\u4e2a\u56fa\u5b9a\u7684\u6838\u51fd\u6570\u914d\u7f6e\uff0c\u5e76\u5c06\u5176\u7528\u4e8e\u6240\u6709\u5f62\u72b6\u7684\u8ba1\u7b97<\/strong>\u3002\u5c3d\u7ba1\u8fd9\u4f1a\u635f\u5931\u4e00\u4e9b\u6027\u80fd\uff0c\u4f46\u5728 LLM \u63a8\u7406\u573a\u666f\u4e0b\uff0c\u8fd9\u79cd\u635f\u5931\u901a\u5e38\u4e0d\u662f\u707e\u96be\u6027\u7684\u3002\u7279\u522b\u662f\uff0cSplit-K \u7b56\u7565\u5728 M \u548c N \u7ef4\u5ea6\u90fd\u5f88\u5c0f\u65f6\u624d\u6700\u88ab\u9700\u8981\uff0c\u800c\u5e78\u8fd0\u7684\u662f\uff0c\u5728\u6211\u4eec\u7684\u5e94\u7528\u573a\u666f\u4e2d\uff0cN \u7ef4\u5ea6\uff08\u5373\u6a21\u578b\u7ef4\u5ea6\uff09\u901a\u5e38\u90fd\u76f8\u5f53\u5927\uff01<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u6279\u6b21\u4e0d\u53d8\u6027\u6ce8\u610f\u529b\u673a\u5236<\/strong><\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u5728\u5b9e\u73b0\u4e86\u77e9\u9635\u4e58\u6cd5\u7684\u6279\u6b21\u4e0d\u53d8\u6027\u4e4b\u540e\uff0c\u6ce8\u610f\u529b\u673a\u5236\u53c8\u5f15\u5165\u4e86\u4e24\u4e2a\u989d\u5916\u7684\u96be\u9898 \u2014\u2014 \u8fd9\u4e5f\u5f88\u8d34\u5207\uff0c\u56e0\u4e3a\u5b83\u6b63\u597d\u5305\u542b\u4e24\u6b21\u77e9\u9635\u4e58\u6cd5\u3002<\/p>\n\n\n\n

    1. \u4e0e RMSNorm \u548c\u77e9\u9635\u4e58\u6cd5\u4ec5\u5728\u7279\u5f81\u7ef4\u5ea6\u4e0a\u8fdb\u884c\u5f52\u7ea6\u4e0d\u540c\uff0c\u6ce8\u610f\u529b\u673a\u5236\u73b0\u5728\u9700\u8981\u5728\u7279\u5f81\u7ef4\u5ea6\u548c\u5e8f\u5217\u7ef4\u5ea6\u4e0a\u90fd\u8fdb\u884c\u5f52\u7ea6\u3002<\/p>\n\n\n\n

    2. \u56e0\u6b64\uff0c\u6ce8\u610f\u529b\u673a\u5236\u5fc5\u987b\u5904\u7406\u5404\u79cd\u5f71\u54cd\u5e8f\u5217\u5904\u7406\u65b9\u5f0f\u7684\u63a8\u7406\u4f18\u5316\uff08\u4f8b\u5982\u5206\u5757\u9884\u586b\u5145\u3001\u524d\u7f00\u7f13\u5b58\u7b49\uff09\u3002<\/p>\n\n\n\n

    \u56e0\u6b64\uff0c\u4e3a\u4e86\u5728 LLM \u63a8\u7406\u4e2d\u5b9e\u73b0\u786e\u5b9a\u6027\uff0c\u6211\u4eec\u7684\u6570\u503c\u8ba1\u7b97\u5fc5\u987b\u5bf9\u4e24\u4e2a\u56e0\u7d20\u4fdd\u6301\u4e0d\u53d8\uff1a\u4e00\u662f\u5355\u6b21\u5904\u7406\u7684\u8bf7\u6c42\u6570\u91cf\uff0c\u4e8c\u662f\u6bcf\u4e2a\u8bf7\u6c42\u5728\u63a8\u7406\u5f15\u64ce\u4e2d\u7684\u5207\u5206\u65b9\u5f0f\u3002<\/p>\n\n\n\n

    \u6211\u4eec\u9996\u5148\u6765\u4e86\u89e3\u4e00\u4e0b\u6ce8\u610f\u529b\u673a\u5236\u7684\u6807\u51c6\u5e76\u884c\u7b56\u7565\uff0c\u8be5\u7b56\u7565\u6700\u521d\u7531 FlashAttention-2 \u63d0\u51fa\u3002\u4e0e RMSNorm \u548c\u77e9\u9635\u4e58\u6cd5\u7c7b\u4f3c\uff0c\u5176\u9ed8\u8ba4\u7b56\u7565\u662f\u6570\u636e\u5e76\u884c\u7b56\u7565\u3002\u7531\u4e8e\u5f52\u7ea6\u662f\u6cbf\u7740\u952e \/ \u503c\uff08K\/V\uff09\u5f20\u91cf\u8fdb\u884c\u7684\uff0c\u56e0\u6b64\u6570\u636e\u5e76\u884c\u7b56\u7565\u53ea\u80fd\u6cbf\u7740\u67e5\u8be2\uff08Q\uff09\u5f20\u91cf\u8fdb\u884c\u5e76\u884c\u5316\u3002<\/p>\n\n\n\n

    \u4f8b\u5982\uff0c\u6839\u636e\u63a8\u7406\u5f15\u64ce\u7684\u9009\u62e9\uff0c\u4e00\u4e2a\u5e8f\u5217\u53ef\u80fd\u88ab\u5206\u6210\u51e0\u4e2a\u90e8\u5206\u5904\u7406\uff08\u5982\u5728\u5206\u5757\u9884\u586b\u5145\u4e2d\uff09\uff0c\u4e5f\u53ef\u80fd\u4e00\u6b21\u6027\u5904\u7406\u5b8c\u6bd5\uff08\u5982\u679c\u9884\u586b\u5145\u672a\u88ab\u5206\u5272\uff09\u3002\u4e3a\u4e86\u5b9e\u73b0\u6279\u6b21\u4e0d\u53d8\u6027\uff0c\u5bf9\u4e8e\u4e00\u4e2a\u7ed9\u5b9a\u7684 token\uff0c\u5176\u5f52\u7ea6\u987a\u5e8f\u5fc5\u987b\u72ec\u7acb\u4e8e\u5176\u6240\u5728\u5e8f\u5217\u4e2d\u540c\u65f6\u88ab\u5904\u7406\u7684\u5176\u4ed6 token \u7684\u6570\u91cf\u3002<\/p>\n\n\n\n

    \u5982\u679c\u4f60\u5c06 KV \u7f13\u5b58\u4e2d\u7684 K\/V \u503c\u4e0e\u5f53\u524d\u6b63\u5728\u5904\u7406\u7684 token \u7684 K\/V \u503c\u5206\u5f00\u8fdb\u884c\u5f52\u7ea6\uff08\u5c31\u50cf\u5728 vLLM \u7684 Triton \u6ce8\u610f\u529b\u6838\u51fd\u6570\u4e2d\u90a3\u6837\uff09\uff0c\u8fd9\u4e2a\u76ee\u6807\u5c31\u65e0\u6cd5\u5b9e\u73b0\u3002\u4f8b\u5982\uff0c\u5728\u5904\u7406\u5e8f\u5217\u4e2d\u7684\u7b2c 1000 \u4e2a\u67e5\u8be2 token \u65f6\uff0c\u65e0\u8bba KV \u7f13\u5b58\u4e2d\u6709 0 \u4e2a token\uff08\u9884\u586b\u5145\u9636\u6bb5\uff09\u8fd8\u662f 999 \u4e2a token\uff08\u89e3\u7801\u9636\u6bb5\uff09\uff0c\u5176\u5f52\u7ea6\u987a\u5e8f\u90fd\u5fc5\u987b\u5b8c\u5168\u76f8\u540c\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u4e3a\u89e3\u51b3\u6b64\u95ee\u9898\uff0c\u6211\u4eec\u53ef\u4ee5\u5728\u6ce8\u610f\u529b\u6838\u51fd\u6570\u8fd0\u884c\u524d\u5c31\u66f4\u65b0 KV \u7f13\u5b58\u548c\u9875\u8868\uff0c\u4ece\u800c\u786e\u4fdd\u65e0\u8bba\u5904\u7406\u591a\u5c11\u4e2a token\uff0c\u6211\u4eec\u7684\u952e\u548c\u503c\u59cb\u7ec8\u5177\u6709\u4e00\u81f4\u7684\u5185\u5b58\u5e03\u5c40\u3002<\/p>\n\n\n\n

    \u52a0\u4e0a\u8fd9\u4e00\u989d\u5916\u5904\u7406\uff08\u4ee5\u53ca\u524d\u6587\u63d0\u5230\u7684\u6240\u6709\u63aa\u65bd\uff0c\u5982\u4f7f\u7528\u4e00\u81f4\u7684 tile \u5927\u5c0f\uff09\uff0c\u6211\u4eec\u4fbf\u80fd\u5b9e\u73b0\u4e00\u4e2a\u6279\u6b21\u4e0d\u53d8\u6027\u7684\u6ce8\u610f\u529b\u673a\u5236\uff01<\/p>\n\n\n\n

    \u7136\u800c\uff0c\u8fd9\u91cc\u5b58\u5728\u4e00\u4e2a\u91cd\u8981\u95ee\u9898\u3002\u4e0e\u77e9\u9635\u4e58\u6cd5\u4e0d\u540c\uff0cLLM \u63a8\u7406\u4e2d\u7684\u6ce8\u610f\u529b\u8ba1\u7b97\u5f62\u72b6\u901a\u5e38\u786e\u5b9e\u9700\u8981\u4e00\u4e2a\u62c6\u5206 – \u5f52\u7ea6\u6838\u51fd\u6570\uff08split-reduction kernel\uff09\uff0c\u8fd9\u7c7b\u6838\u51fd\u6570\u5e38\u88ab\u79f0\u4e3a Split-KV \u6216 FlashDecoding\u3002\u8fd9\u662f\u56e0\u4e3a\u5982\u679c\u6211\u4eec\u4e0d\u6cbf\u7740\u5f52\u7ea6\u7ef4\u5ea6\u8fdb\u884c\u5e76\u884c\uff0c\u5c31\u53ea\u80fd\u6cbf\u7740\u6279\u6b21\u7ef4\u5ea6\u3001\u5934\u7ef4\u5ea6\u548c\u67e5\u8be2\u957f\u5ea6\u7ef4\u5ea6\u8fdb\u884c\u5e76\u884c\u3002<\/p>\n\n\n\n

    \u5728\u6ce8\u610f\u529b\u7684\u89e3\u7801\u9636\u6bb5\uff0c\u67e5\u8be2\u957f\u5ea6\u975e\u5e38\u5c0f\uff08\u901a\u5e38\u4e3a 1\uff09\uff0c\u56e0\u6b64\u9664\u975e batch size \u975e\u5e38\u5927\uff0c\u5426\u5219\u6211\u4eec\u5f80\u5f80\u65e0\u6cd5\u4f7f GPU \u8fbe\u5230\u9971\u548c\u72b6\u6001\u3002\u4e0d\u5e78\u7684\u662f\uff0c\u8fd9\u79cd\u60c5\u51b5\u4e0d\u50cf\u5728 RMSNorm \u548c\u77e9\u9635\u4e58\u6cd5\u4e2d\u90a3\u6837\u5bb9\u6613\u88ab\u5ffd\u7565\u3002\u4f8b\u5982\uff0c\u5982\u679c\u4f60\u7684 KV \u7f13\u5b58\u975e\u5e38\u957f\uff0c\u5373\u4f7f\u53ea\u5904\u7406\u4e00\u4e2a\u8bf7\u6c42\uff0c\u6ce8\u610f\u529b\u6838\u51fd\u6570\u7684\u8ba1\u7b97\u4e5f\u53ef\u80fd\u8017\u65f6\u5f88\u957f\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u6b64\u5916\uff0c\u5e38\u7528\u4e8e\u6ce8\u610f\u529b\u7684\u62c6\u5206 – \u5f52\u7ea6\u7b56\u7565\u4e5f\u7ed9\u6279\u6b21\u4e0d\u53d8\u6027\u5e26\u6765\u4e86\u6311\u6218\u3002\u4f8b\u5982\uff0cFlashInfer \u7684\u5e73\u8861\u8c03\u5ea6\u7b97\u6cd5\u4f1a\u9009\u62e9\u80fd\u591f\u4f7f GPU \u6240\u6709\u6838\u5fc3\u9971\u548c\u7684\u6700\u5927\u62c6\u5206\u5927\u5c0f\uff0c\u8fd9\u4f7f\u5f97\u5176\u5f52\u7ea6\u7b56\u7565\u5e76\u975e\u6279\u6b21\u4e0d\u53d8\u7684\u3002\u7136\u800c\uff0c\u4e0e RMSNorm \/ \u77e9\u9635\u4e58\u6cd5\u4e0d\u540c\uff0c\u65e0\u8bba batch size \u5982\u4f55\uff0c\u4ec5\u4ec5\u9009\u62e9\u4e00\u4e2a\u56fa\u5b9a\u7684\u62c6\u5206\u6570\u91cf\u662f\u4e0d\u591f\u7684\u3002<\/p>\n\n\n\n

    \u76f8\u53cd\uff0c\u4e3a\u4e86\u5b9e\u73b0\u6279\u6b21\u4e0d\u53d8\u6027\uff0c\u6211\u4eec\u5fc5\u987b\u91c7\u7528\u56fa\u5b9a\u62c6\u5206\u5927\u5c0f\u7b56\u7565\u3002\u6362\u8a00\u4e4b\uff0c\u6211\u4eec\u56fa\u5b9a\u7684\u4e0d\u662f\u62c6\u5206\u7684\u6570\u91cf\uff0c\u800c\u662f\u6bcf\u4e2a\u62c6\u5206\u5757\u7684\u5927\u5c0f\uff0c\u8fd9\u6837\u6700\u7ec8\u4f1a\u5f97\u5230\u4e00\u4e2a\u53ef\u53d8\u7684\u62c6\u5206\u6570\u91cf\u3002\u901a\u8fc7\u8fd9\u79cd\u65b9\u5f0f\uff0c\u6211\u4eec\u53ef\u4ee5\u4fdd\u8bc1\u65e0\u8bba\u6b63\u5728\u5904\u7406\u591a\u5c11\u4e2a token\uff0c\u6211\u4eec\u603b\u662f\u6267\u884c\u5b8c\u5168\u76f8\u540c\u7684\u5f52\u7ea6\u987a\u5e8f\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u5b9e\u73b0<\/strong><\/p>\n\n\n\n

    \u6211\u4eec\u57fa\u4e8e vLLM\uff0c\u901a\u8fc7\u5229\u7528\u5176 FlexAttention \u540e\u7aef\u548c torch.Library\uff0c\u63d0\u4f9b\u4e86\u4e00\u4e2a\u786e\u5b9a\u6027\u63a8\u7406\u7684\u6f14\u793a\u3002\u901a\u8fc7 torch.Library\uff0c\u6211\u4eec\u80fd\u591f\u4ee5\u4e00\u79cd\u975e\u4fb5\u5165\u5f0f\u7684\u65b9\u5f0f\u66ff\u6362\u6389\u5927\u90e8\u5206\u76f8\u5173\u7684 PyTorch \u7b97\u5b50\u3002<\/p>\n\n\n\n

    \u4f60\u53ef\u4ee5\u5728 thinking-machines-lab\/batch-invariant-ops \u627e\u5230\u300c\u6279\u6b21\u4e0d\u53d8\u6027\u300d\u6838\u51fd\u6570\u5e93\uff0c\u4ee5\u53ca\u5728\u300c\u786e\u5b9a\u6027\u300d\u6a21\u5f0f\u4e0b\u8fd0\u884c\u7684 vLLM \u793a\u4f8b\u3002<\/p>\n\n\n\n

    \u5730\u5740\uff1ahttps:\/\/github.com\/thinking-machines-lab\/batch_invariant_ops<\/p>\n\n\n\n

    \u5b9e\u9a8c<\/strong><\/p>\n\n\n\n

    \u5b8c\u6210\u7ed3\u679c\u7684\u4e0d\u786e\u5b9a\u6027\u7a0b\u5ea6\u5982\u4f55\uff1f<\/strong><\/p>\n\n\n\n

    \u6211\u4eec\u4f7f\u7528 Qwen3-235B-A22B-Instruct-2507 \u6a21\u578b\uff0c\u5728\u6e29\u5ea6\u4e3a 0 \u7684\u8bbe\u7f6e\u4e0b\uff0c\u4f7f\u7528\u63d0\u793a\u8bcd\u300cTell me about Richard Feynman\u300d\uff08\u975e\u601d\u8003\u6a21\u5f0f\uff09\u91c7\u6837\u4e86 1000 \u6b21\u5b8c\u6210\u7ed3\u679c\uff0c\u6bcf\u6b21\u751f\u6210 1000 \u4e2a token\u3002<\/p>\n\n\n\n

    \u4ee4\u4eba\u60ca\u8bb6\u7684\u662f\uff0c\u6211\u4eec\u5f97\u5230\u4e86 80 \u4e2a\u4e0d\u540c\u7684\u5b8c\u6210\u7ed3\u679c\uff0c\u5176\u4e2d\u6700\u5e38\u89c1\u7684\u4e00\u4e2a\u51fa\u73b0\u4e86 78 \u6b21\u3002<\/p>\n\n\n\n

    \u901a\u8fc7\u89c2\u5bdf\u8fd9\u4e9b\u7ed3\u679c\u7684\u5dee\u5f02\uff0c\u6211\u4eec\u53d1\u73b0\u5b83\u4eec\u5728\u524d 102 \u4e2a token \u4e0a\u5b9e\u9645\u4e0a\u662f\u5b8c\u5168\u76f8\u540c\u7684\uff01<\/p>\n\n\n\n

    \u9996\u6b21\u51fa\u73b0\u5dee\u5f02\u662f\u5728\u7b2c 103 \u4e2a token\u3002\u6240\u6709\u7684\u7ed3\u679c\u90fd\u751f\u6210\u4e86\u300cFeynman was born on May 11, 1918, in\u300d\u8fd9\u4e2a\u5e8f\u5217\u3002\u7136\u800c\uff0c\u63a5\u4e0b\u6765\uff0c\u5176\u4e2d 992 \u6b21\u7ed3\u679c\u751f\u6210\u4e86\u300cQueens, New York\u300d\uff0c\u800c\u53e6\u5916 8 \u6b21\u5219\u751f\u6210\u4e86\u300cNew York City\u300d\u3002<\/p>\n\n\n\n

    \u7136\u800c\uff0c\u5f53\u6211\u4eec\u542f\u7528\u6279\u6b21\u4e0d\u53d8\u6027\u6838\u51fd\u6570\u540e\uff0c\u5168\u90e8 1000 \u6b21\u7ed3\u679c\u90fd\u53d8\u5f97\u5b8c\u5168\u76f8\u540c\u3002\u8fd9\u6b63\u662f\u6211\u4eec\u671f\u671b\u91c7\u6837\u5668\u5e94\u6709\u7684\u8868\u73b0\uff0c\u4f46\u82e5\u4e0d\u4f7f\u7528\u6211\u4eec\u7684\u6279\u6b21\u4e0d\u53d8\u6027\u6838\u51fd\u6570\uff0c\u5c31\u65e0\u6cd5\u5b9e\u73b0\u786e\u5b9a\u6027\u7ed3\u679c\u3002<\/p>\n\n\n\n

    \u6027\u80fd<\/strong><\/p>\n\n\n\n

    \u76ee\u524d\uff0c\u6211\u4eec\u8fd8\u6ca1\u6709\u6295\u5165\u7cbe\u529b\u4f18\u5316\u6279\u6b21\u4e0d\u53d8\u6027\u6838\u51fd\u6570\u7684\u6027\u80fd\u3002\u4e0d\u8fc7\uff0c\u6211\u4eec\u8fd8\u662f\u8fdb\u884c\u4e86\u4e00\u4e9b\u5b9e\u9a8c\u6765\u9a8c\u8bc1\u5176\u6027\u80fd\u662f\u5426\u4ecd\u5728\u53ef\u7528\u8303\u56f4\u5185\u3002<\/p>\n\n\n\n

    \u6211\u4eec\u642d\u5efa\u4e86\u4e00\u4e2a\u914d\u5907\u5355\u5757 GPU \u7684 API \u670d\u52a1\u5668\uff0c\u8fd0\u884c Qwen-3-8B \u6a21\u578b\uff0c\u5e76\u8bf7\u6c42\u751f\u6210 1000 \u4e2a\u5e8f\u5217\uff0c\u8f93\u51fa\u957f\u5ea6\u63a7\u5236\u5728 90 \u5230 110 \u4e2a token \u4e4b\u95f4\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u6027\u80fd\u4e0b\u964d\u7684\u4e3b\u8981\u539f\u56e0\u5728\u4e8e vLLM \u4e2d\u7684 FlexAttention \u96c6\u6210\u5c1a\u672a\u7ecf\u8fc7\u6df1\u5ea6\u4f18\u5316\u3002\u5c3d\u7ba1\u5982\u6b64\uff0c\u6211\u4eec\u770b\u5230\u5176\u6027\u80fd\u5e76\u672a\u51fa\u73b0\u707e\u96be\u6027\u4e0b\u964d\u3002<\/p>\n\n\n\n

    \u771f\u6b63\u7684\u5728\u7b56\u7565\u5f3a\u5316\u5b66\u4e60<\/strong><\/p>\n\n\n\n

    \u6b63\u5982\u7814\u7a76\u4eba\u5458\u6240\u6307\u51fa\u7684\uff0c\u8bad\u7ec3\u548c\u63a8\u7406\u4e4b\u95f4\u7684\u6570\u503c\u5dee\u5f02\u4f1a\u9690\u5f0f\u5730\u5c06\u6211\u4eec\u7684\u5728\u7b56\u7565\u5f3a\u5316\u5b66\u4e60\uff08on-policy RL\uff09\u8f6c\u53d8\u4e3a\u79bb\u7b56\u7565\u5f3a\u5316\u5b66\u4e60\uff08off-policy RL\uff09\u3002<\/p>\n\n\n\n

    \u5f53\u7136\uff0c\u5982\u679c\u6211\u4eec\u751a\u81f3\u65e0\u6cd5\u4ece\u4e24\u6b21\u76f8\u540c\u7684\u63a8\u7406\u8bf7\u6c42\u4e2d\u83b7\u5f97\u6bcf\u4e00\u4f4d\u90fd\u76f8\u540c\u7684\u7ed3\u679c\uff0c\u90a3\u4e48\u5728\u8bad\u7ec3\u548c\u63a8\u7406\u4e4b\u95f4\u83b7\u5f97\u6bcf\u4e00\u4f4d\u90fd\u76f8\u540c\u7684\u7ed3\u679c\u4e5f\u662f\u4e0d\u53ef\u80fd\u7684\u3002\u56e0\u6b64\uff0c\u786e\u5b9a\u6027\u63a8\u7406\u4f7f\u6211\u4eec\u80fd\u591f\u4fee\u6539\u8bad\u7ec3\u5806\u6808\uff0c\u4ece\u800c\u5728\u91c7\u6837\u548c\u8bad\u7ec3\u4e4b\u95f4\u83b7\u5f97\u6bcf\u4e00\u4f4d\u90fd\u76f8\u540c\u7684\u7ed3\u679c\uff0c\u6700\u7ec8\u5b9e\u73b0\u771f\u6b63\u7684\u5728\u7b56\u7565\u5f3a\u5316\u5b66\u4e60\u3002<\/p>\n\n\n\n

    \u6211\u4eec\u5728 Bigmath \u4e0a\uff0c\u4f7f\u7528 RLVR \u8bbe\u7f6e\u8fdb\u884c\u4e86\u5b9e\u9a8c\uff0c\u5176\u4e2d\u5f3a\u5316\u5b66\u4e60\u7b56\u7565\u7531 Qwen 2.5-VL instruct 8B \u6a21\u578b\u521d\u59cb\u5316\uff0c\u6700\u5927 rollout \u957f\u5ea6\u4e3a 4096\u3002<\/p>\n\n\n\n

    \u5982\u679c\u6211\u4eec\u4e0d\u4f7f\u7528\u79bb\u7b56\u7565\u6821\u6b63\uff08\u5373\u91cd\u8981\u5ea6\u52a0\u6743\uff09\u8fdb\u884c\u8bad\u7ec3\uff0c\u6211\u4eec\u7684\u5956\u52b1\u4f1a\u5728\u8bad\u7ec3\u4e2d\u9014\u5d29\u6e83\uff1b\u800c\u6dfb\u52a0\u79bb\u7b56\u7565\u6821\u6b63\u9879\u5219\u53ef\u4ee5\u4f7f\u8bad\u7ec3\u987a\u5229\u8fdb\u884c\u3002\u4f46\u662f\uff0c\u5982\u679c\u6211\u4eec\u5728\u91c7\u6837\u5668\u548c\u8bad\u7ec3\u5668\u4e4b\u95f4\u5b9e\u73b0\u4e86\u6bcf\u4e00\u4f4d\u90fd\u76f8\u540c\u7684\u7ed3\u679c\uff0c\u6211\u4eec\u5c31\u5b8c\u5168\u5904\u4e8e\u5728\u7b56\u7565\u72b6\u6001\uff08\u5373 KL \u6563\u5ea6\u4e3a 0\uff09\uff0c\u540c\u6837\u53ef\u4ee5\u987a\u5229\u5730\u8fdb\u884c\u8bad\u7ec3\u3002<\/p>\n\n\n\n

    \u6211\u4eec\u8fd8\u53ef\u4ee5\u7ed8\u5236\u91c7\u6837\u5668\u548c\u8bad\u7ec3\u5668\u4e4b\u95f4\u5bf9\u6570\u6982\u7387\u7684 KL \u6563\u5ea6\uff0c\u5176\u4e2d\u6240\u6709 3 \u6b21\u8fd0\u884c\u90fd\u8868\u73b0\u51fa\u663e\u8457\u4e0d\u540c\u7684\u884c\u4e3a\u3002\u5728\u4f7f\u7528\u91cd\u8981\u5ea6\u52a0\u6743\u8fd0\u884c\u65f6\uff0cKL \u6563\u5ea6\u4fdd\u6301\u5728 0.001 \u5de6\u53f3\uff0c\u5e76\u4f34\u6709\u5076\u5c14\u7684\u5cf0\u503c\u3002\u7136\u800c\uff0c\u5728\u4e0d\u4f7f\u7528\u91cd\u8981\u5ea6\u52a0\u6743\u7684\u60c5\u51b5\u4e0b\u8fd0\u884c\uff0c\u6700\u7ec8\u4f1a\u5bfc\u81f4 KL \u6563\u5ea6\u5728\u5927\u7ea6\u4e0e\u5956\u52b1\u5d29\u6e83\u540c\u4e00\u65f6\u95f4\u51fa\u73b0\u5cf0\u503c\u3002\u5f53\u7136\uff0c\u5728\u8fd0\u884c\u300c\u771f\u6b63\u7684\u5728\u7b56\u7565\u5f3a\u5316\u5b66\u4e60\u300d\u65f6\uff0c\u6211\u4eec\u7684 KL \u6563\u5ea6\u59cb\u7ec8\u4fdd\u6301\u4e3a 0\uff0c\u8fd9\u8868\u660e\u8bad\u7ec3\u7b56\u7565\u548c\u91c7\u6837\u7b56\u7565\u4e4b\u95f4\u4e0d\u5b58\u5728\u4efb\u4f55\u5dee\u5f02\u3002<\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n

    \u603b\u7ed3<\/strong><\/p>\n\n\n\n

    \u73b0\u4ee3\u8f6f\u4ef6\u7cfb\u7edf\u5f80\u5f80\u7531\u591a\u5c42\u62bd\u8c61\u6784\u6210\u3002\u5728\u673a\u5668\u5b66\u4e60\u4e2d\uff0c\u5f53\u6211\u4eec\u9047\u5230\u4e0d\u786e\u5b9a\u6027\u548c\u4e00\u4e9b\u5fae\u5999\u7684\u6570\u503c\u5dee\u5f02\u65f6\uff0c\u4eba\u4eec\u5f80\u5f80\u4f1a\u503e\u5411\u4e8e\u89c6\u800c\u4e0d\u89c1\u3002<\/p>\n\n\n\n

    \u6bd5\u7adf\uff0c\u6211\u4eec\u7684\u7cfb\u7edf\u672c\u6765\u5c31\u662f\u300c\u6982\u7387\u6027\u7684\u300d\uff0c\u518d\u591a\u4e00\u70b9\u4e0d\u786e\u5b9a\u6027\u53c8\u6709\u4f55\u59a8\uff1f\u5355\u5143\u6d4b\u8bd5\u6302\u6389\u65f6\uff0c\u628a atol\/rtol \u8c03\u5927\u70b9\u6709\u4ec0\u4e48\u95ee\u9898\uff1f\u8bad\u7ec3\u5668\u548c\u91c7\u6837\u5668\u4e4b\u95f4\u7684\u5bf9\u6570\u6982\u7387\u5dee\u5f02\uff0c\u5e94\u8be5\u4e0d\u662f\u771f\u6b63\u7684 bug \u5427\uff1f<\/p>\n\n\n\n

    \u6211\u4eec\u62d2\u7edd\u8fd9\u79cd\u6d88\u6781\u5fc3\u6001\u3002\u53ea\u8981\u7a0d\u5fae\u591a\u505a\u4e00\u4e9b\u52aa\u529b\uff0c\u6211\u4eec\u5c31\u80fd\u7406\u89e3\u4e0d\u786e\u5b9a\u6027\u7684\u6839\u6e90\uff0c\u751a\u81f3\u771f\u6b63\u89e3\u51b3\u5b83\u4eec\uff01<\/p>\n\n\n\n

    \u6211\u4eec\u5e0c\u671b\u8fd9\u7bc7\u535a\u6587\u80fd\u4e3a\u793e\u533a\u63d0\u4f9b\u4e00\u5957\u53ef\u9760\u7684\u601d\u8def\uff0c\u5e2e\u52a9\u5927\u5bb6\u5728\u63a8\u7406\u7cfb\u7edf\u4e2d\u5e94\u5bf9\u4e0d\u786e\u5b9a\u6027\uff0c\u5e76\u6fc0\u52b1\u66f4\u591a\u4eba\u6df1\u5165\u7406\u89e3\u81ea\u5df1\u7684\u7cfb\u7edf\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

    \u5c31\u5728\u4eca\u5929\uff0c\u7531 OpenAI \u524d CTO Mira Murati \u6210\u7acb\u4e8e\u4eca\u5e74 2 \u6708\u7684\u4eba\u5de5\u667a\u80fd\u521d\u521b\u516c\u53f8 Thin […]<\/p>\n","protected":false},"author":1,"featured_media":1007,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[10],"tags":[],"class_list":["post-1005","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-10"],"_links":{"self":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/1005","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/comments?post=1005"}],"version-history":[{"count":1,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/1005\/revisions"}],"predecessor-version":[{"id":1033,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/1005\/revisions\/1033"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/media\/1007"}],"wp:attachment":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/media?parent=1005"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/categories?post=1005"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/tags?post=1005"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}