{"id":415415,"date":"2024-08-08T11:29:43","date_gmt":"2024-08-08T03:29:43","guid":{"rendered":"https:\/\/www.idc.net\/help\/?p=415415"},"modified":"2024-08-08T11:29:43","modified_gmt":"2024-08-08T03:29:43","slug":"%e9%ab%98%e6%80%a7%e8%83%bd-llm-%e6%8e%a8%e7%90%86%e6%a1%86%e6%9e%b6%e7%9a%84%e8%ae%be%e8%ae%a1%e4%b8%8e%e5%ae%9e%e7%8e%b0","status":"publish","type":"post","link":"https:\/\/idc.net\/help\/415415\/","title":{"rendered":"\u9ad8\u6027\u80fd LLM \u63a8\u7406\u6846\u67b6\u7684\u8bbe\u8ba1\u4e0e\u5b9e\u73b0"},"content":{"rendered":"<h2>\u4e00\u3001\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u6982\u8981\u4ecb\u7ecd<\/h2>\n<p>&nbsp;<\/p>\n<p>\u4e0e\u4f20\u7edf\u7684 CNN \u6a21\u578b\u63a8\u7406\u4e0d\u540c\uff0c\u5927\u8bed\u8a00\u6a21\u578b\u7684\u63a8\u7406\u901a\u5e38\u4f1a\u5206\u6210 prefill \u548c decoding \u4e24\u4e2a\u9636\u6bb5\u3002\u6bcf\u4e00\u4e2a\u8bf7\u6c42\u53d1\u8d77\u540e\u4ea7\u751f\u7684\u63a8\u7406\u8fc7\u7a0b\u90fd\u4f1a\u5148\u7ecf\u5386\u4e00\u4e2a Prefill \u8fc7\u7a0b\uff0cprefill \u8fc7\u7a0b\u4f1a\u8ba1\u7b97\u7528\u6237\u6240\u6709\u7684\u8f93\u5165\uff0c\u5e76\u751f\u6210\u5bf9\u5e94\u7684 KV \u7f13\u5b58\uff0c\u518d\u7ecf\u5386\u82e5\u5e72\u4e2a decoding \u8fc7\u7a0b\uff0c\u6bcf\u4e00\u4e2a decoding \u8fc7\u7a0b\uff0c\u670d\u52a1\u5668\u90fd\u4f1a\u751f\u6210\u4e00\u4e2a\u5b57\u7b26\uff0c\u5e76\u5c06\u5176\u653e\u5165\u5230 KV \u7f13\u5b58\u5f53\u4e2d\uff0c\u4e4b\u540e\u4f9d\u6b21\u8fed\u4ee3\u3002<\/p>\n<p>\u7531\u4e8e decoding \u8fc7\u7a0b\u662f\u9010\u4e2a\u5b57\u7b26\u751f\u6210\u7684\uff0c\u6bcf\u4e00\u6bb5\u7b54\u6848\u7684\u751f\u6210\u90fd\u9700\u8981\u5f88\u957f\u65f6\u95f4\uff0c\u4f1a\u751f\u6210\u5f88\u591a\u5b57\u7b26\uff0c\u6240\u4ee5 decoding \u9636\u6bb5\u7684\u6570\u91cf\u975e\u5e38\u591a\uff0c\u5360\u5230\u6574\u4e2a\u63a8\u7406\u8fc7\u7a0b\u7684 90% \u4ee5\u4e0a\u3002<\/p>\n<p>\u5728 Prefill \u8fc7\u7a0b\u4e2d\uff0c\u867d\u7136\u8ba1\u7b97\u91cf\u5f88\u5927\uff0c\u56e0\u4e3a\u8981\u4e00\u6b21\u6027\u5b8c\u6210\u7528\u6237\u8f93\u5165\u7684\u6240\u6709\u8bcd\u7684\u8ba1\u7b97\uff0c\u4f46\u5b83\u53ea\u662f\u4e00\u6b21\u6027\u7684\u8fc7\u7a0b\uff0c\u6240\u4ee5\u5728\u6574\u4e2a\u63a8\u7406\u4e2d\u53ea\u5360\u4e0d\u5230 10% \u7684\u65f6\u95f4\u3002<\/p>\n<p>\u5728\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u4e2d\u5e38\u4f1a\u7528\u5230\u56db\u4e2a\u6307\u6807\uff1aThroughput\uff08\u541e\u5410\u91cf\uff09\u3001First Token Latency\uff08\u9996\u5b57\u5ef6\u8fdf\uff09\u3001Latency\uff08\u5ef6\u8fdf\uff09\u548cQPS\uff08\u6bcf\u79d2\u8bf7\u6c42\u6570\uff09\u3002\u8fd9\u56db\u4e2a\u6027\u80fd\u6307\u6807\u4f1a\u4ece\u56db\u4e2a\u4e0d\u540c\u7684\u65b9\u9762\u6765\u8861\u91cf\u4e00\u4e2a\u7cfb\u7edf\u7684\u670d\u52a1\u63d0\u4f9b\u80fd\u529b\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u9996\u5148\u6765\u4ecb\u7ecd Throughput\uff08\u541e\u5410\u91cf\uff09\u3002\u4ece\u6a21\u578b\u63a8\u7406\u5c42\u9762\u4e0a\u770b\uff0c\u6700\u5148\u5173\u6ce8\u7684\u5c31\u662f\u541e\u5410\u91cf\u3002\u541e\u5410\u91cf\u662f\u6307\u5f53\u7cfb\u7edf\u7684\u8d1f\u8f7d\u8fbe\u5230\u6700\u5927\u7684\u65f6\u5019\uff0c\u5728\u5355\u4f4d\u65f6\u95f4\u5185\uff0c\u80fd\u591f\u6267\u884c\u591a\u5c11\u4e2a decoding\uff0c\u5373\u751f\u6210\u591a\u5c11\u4e2a\u5b57\u7b26\u3002\u6d4b\u8bd5\u541e\u5410\u91cf\u7684\u65b9\u6cd5\u662f\uff0c\u5047\u8bbe\u6240\u6709\u7528\u6237\u90fd\u4f1a\u5728\u540c\u4e00\u65f6\u523b\u5230\u6765\uff0c\u5e76\u4e14\u8fd9\u4e9b\u7528\u6237\u95ee\u7684\u90fd\u662f\u4e00\u6837\u7684\u95ee\u9898\uff0c\u8fd9\u4e9b\u7528\u6237\u53ef\u4ee5\u540c\u65f6\u542f\u52a8\u548c\u7ed3\u675f\uff0c\u4e14\u4ed6\u4eec\u751f\u6210\u7684\u6587\u672c\u7684\u957f\u5ea6\u548c\u8f93\u5165\u7684\u6587\u672c\u957f\u5ea6\u90fd\u662f\u4e00\u6837\u7684\u3002\u901a\u8fc7\u4f7f\u7528\u5b8c\u5168\u76f8\u540c\u7684\u8f93\u5165\uff0c\u7ec4\u6210\u4e00\u4e2a\u5b8c\u6574\u7684 batch\u3002\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u7cfb\u7edf\u7684\u541e\u5410\u91cf\u4f1a\u8fbe\u5230\u6700\u9ad8\u3002\u4f46\u8fd9\u79cd\u60c5\u51b5\u662f\u4e0d\u5408\u5b9e\u9645\u7684\uff0c\u6240\u4ee5\u8fd9\u662f\u4e00\u4e2a\u7406\u8bba\u7684\u6700\u5927\u503c\u3002\u6211\u4eec\u4f1a\u6d4b\u91cf\u5728\u4e00\u79d2\u949f\u4e4b\u5185\uff0c\u7cfb\u7edf\u80fd\u591f\u6267\u884c\u591a\u5c11\u4e2a\u72ec\u7acb\u7684 decoding \u9636\u6bb5\u3002<\/p>\n<p>\u7b2c\u4e8c\u4e2a\u6307\u6807\u662f First Token Latency\uff08\u9996\u5b57\u5ef6\u8fdf\uff09\u3002\u6307\u7684\u662f\u5f53\u4e00\u6279\u7528\u6237\u8fdb\u5165\u5230\u63a8\u7406\u7cfb\u7edf\u4e4b\u540e\uff0c\u7528\u6237\u5b8c\u6210 Prefill \u9636\u6bb5\u7684\u8fc7\u7a0b\u9700\u8981\u82b1\u591a\u957f\u65f6\u95f4\u3002\u8fd9\u4e5f\u662f\u7cfb\u7edf\u751f\u6210\u7b2c\u4e00\u4e2a\u5b57\u7b26\u6240\u9700\u7684\u54cd\u5e94\u65f6\u95f4\u3002\u5f88\u591a\u9700\u6c42\u5173\u6ce8\u8fd9\u4e00\u6307\u6807\uff0c\u5e0c\u671b\u7528\u6237\u5728\u7cfb\u7edf\u4e0a\u8f93\u5165\u95ee\u9898\u540e\u5f97\u5230\u56de\u7b54\u7684\u65f6\u95f4\u5c0f\u4e8e 2~3 \u79d2\u3002<\/p>\n<p>\u7b2c\u4e09\u4e2a\u6307\u6807\u662f Latency\uff08\u5ef6\u8fdf\uff09\u3002\u6307\u7684\u662f\u6bcf\u4e00\u4e2a decoding \u6240\u9700\u8981\u7684\u65f6\u957f\u3002\u5b83\u53cd\u6620\u7684\u662f\u5927\u8bed\u8a00\u6a21\u578b\u7cfb\u7edf\u5728\u7ebf\u4e0a\u5904\u7406\u7684\u8fc7\u7a0b\u4e2d\uff0c\u6bcf\u751f\u6210\u4e00\u4e2a\u5b57\u7b26\u7684\u95f4\u9694\u662f\u591a\u957f\u65f6\u95f4\uff0c\u4e5f\u5c31\u662f\u751f\u6210\u7684\u8fc7\u7a0b\u6709\u591a\u4e48\u6d41\u7545\u3002\u5927\u90e8\u5206\u60c5\u51b5\u4e0b\uff0c\u6211\u4eec\u5e0c\u671b\u751f\u6210\u7684\u5ef6\u8fdf\u5c0f\u4e8e 50 \u6beb\u79d2\uff0c\u4e5f\u5c31\u662f\u4e00\u79d2\u949f\u751f\u6210 20 \u4e2a\u5b57\u7b26\u3002\u8fd9\u6837\u5927\u8bed\u8a00\u6a21\u578b\u7684\u751f\u6210\u662f\u6bd4\u8f83\u6d41\u7545\u7684\u3002<\/p>\n<p>\u6700\u540e\u4e00\u4e2a\u6307\u6807\u662f QPS\uff08\u6bcf\u79d2\u8bf7\u6c42\u6570\uff09\u3002\u53cd\u6620\u4e86\u5728\u7ebf\u4e0a\u7cfb\u7edf\u7684\u670d\u52a1\u5f53\u4e2d\uff0c\u4e00\u79d2\u949f\u80fd\u591f\u5904\u7406\u591a\u5c11\u4e2a\u7528\u6237\u7684\u8bf7\u6c42\u3002\u8fd9\u4e00\u6307\u6807\u7684\u6d4b\u91cf\u65b9\u5f0f\u6bd4\u8f83\u590d\u6742\uff0c\u540e\u9762\u4f1a\u5c55\u5f00\u4ecb\u7ecd\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u5bf9\u4e8e First Token Latency \u548c Latency \u8fd9\u4e24\u4e2a\u6307\u6807\uff0c\u6211\u4eec\u90fd\u8fdb\u884c\u4e86\u76f8\u5bf9\u5b8c\u5584\u7684\u6d4b\u8bd5\u3002\u8fd9\u4e24\u4e2a\u6307\u6807\u4f1a\u56e0\u4e3a\u7528\u6237\u8f93\u5165\u7684\u957f\u5ea6\u4e0d\u540c\u3001batch size \u7684\u4e0d\u540c\u800c\u53d1\u751f\u975e\u5e38\u5927\u7684\u53d8\u5316\u3002<\/p>\n<p>\u5728\u4e0a\u8868\u4e2d\u53ef\u4ee5\u770b\u5230\uff0c\u5bf9\u4e8e\u540c\u6837\u7684 7B \u6a21\u578b\uff0c\u5982\u679c\u7528\u6237\u7684\u8f93\u5165\u957f\u5ea6\u4ece 8 \u53d8\u6210 2048\uff0cPrefill \u7684\u65f6\u95f4\u5c06\u4ece 6.78 \u6beb\u79d2\uff0c\u76f4\u5230\u53d8\u6210 2078 \u6beb\u79d2\uff0c\u5373 2 \u79d2\u7684\u65f6\u95f4\u3002\u5982\u679c\u6709 80 \u4e2a\u7528\u6237\uff0c\u6bcf\u4e00\u4e2a\u7528\u6237\u90fd\u8f93\u5165 1,024 \u4e2a\u8bcd\uff0c\u90a3\u4e48 Prefill \u5728\u670d\u52a1\u7aef\u5c31\u8981\u8dd1 2 \u79d2\u5de6\u53f3\uff0c\u8fd9\u4e2a\u65f6\u95f4\u5df2\u7ecf\u8d85\u51fa\u4e86\u53ef\u4ee5\u63a5\u53d7\u7684\u8303\u56f4\u3002\u4f46\u5982\u679c\u7528\u6237\u8f93\u5165\u957f\u5ea6\u90fd\u5f88\u77ed\uff0c\u6bd4\u5982\u6bcf\u6b21\u8bbf\u95ee\u53ea\u8f93\u5165 8 \u4e2a\u8bcd\uff0c\u54ea\u6015 768 \u4e2a\u7528\u6237\u540c\u65f6\u5230\u6765\uff0c\u9996\u5b57\u5ef6\u8fdf\u4e5f\u53ea\u6709 165 \u6beb\u79d2\u5de6\u53f3\u3002<\/p>\n<p>\u4e0e\u9996\u5b57\u5ef6\u8fdf\u6700\u76f8\u5173\u7684\u5c31\u662f\u7528\u6237\u7684\u8f93\u5165\u957f\u5ea6\uff0c\u7528\u6237\u8f93\u5165\u7684\u957f\u5ea6\u8d8a\u957f\uff0c\u9996\u5b57\u5ef6\u8fdf\u4e5f\u4f1a\u8d8a\u9ad8\u3002\u7528\u6237\u8f93\u5165\u957f\u5ea6\u5982\u679c\u5f88\u77ed\uff0c\u90a3\u4e48\u9996\u5b57\u5ef6\u8fdf\u5728\u6574\u4e2a\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u8fc7\u7a0b\u4e2d\u90fd\u4e0d\u4f1a\u6210\u4e3a\u74f6\u9888\u3002<\/p>\n<p>\u800c\u540e\u9762\u7684 decoding \u5ef6\u8fdf\uff0c\u901a\u5e38\u53ea\u8981\u4e0d\u662f\u5343\u4ebf\u7ea7\u522b\u7684\u6a21\u578b\uff0cdecoding \u7684\u5ef6\u8fdf\u90fd\u4f1a\u63a7\u5236\u5728 50 \u6beb\u79d2\u4ee5\u5185\u3002\u5b83\u4e3b\u8981\u53d7\u5230 batch size \u7684\u5f71\u54cd\uff0cbatch size \u8d8a\u5927\uff0c\u63a8\u7406\u5ef6\u8fdf\u4e5f\u4f1a\u8d8a\u5927\uff0c\u4f46\u57fa\u672c\u4e0a\u589e\u52a0\u7684\u5e45\u5ea6\u4e0d\u4f1a\u5f88\u9ad8\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u541e\u5410\u91cf\u5176\u5b9e\u4e5f\u4f1a\u53d7\u5230\u8fd9\u4e24\u4e2a\u56e0\u7d20\u7684\u5f71\u54cd\u3002\u5982\u679c\u7528\u6237\u8f93\u5165\u7684\u957f\u5ea6\u548c\u751f\u6210\u7684\u957f\u5ea6\u5f88\u957f\uff0c\u90a3\u4e48\u7cfb\u7edf\u541e\u5410\u91cf\u4e5f\u4e0d\u4f1a\u5f88\u9ad8\u3002\u5982\u679c\u7528\u6237\u8f93\u5165\u957f\u5ea6\u548c\u751f\u6210\u957f\u5ea6\u90fd\u4e0d\u662f\u5f88\u957f\uff0c\u90a3\u4e48\u7cfb\u7edf\u541e\u5410\u91cf\u53ef\u80fd\u4f1a\u8fbe\u5230\u4e00\u4e2a\u975e\u5e38\u79bb\u8c31\u7684\u7a0b\u5ea6\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u518d\u6765\u770b QPS\u3002QPS \u662f\u4e00\u4e2a\u975e\u5e38\u5177\u4f53\u7684\u6307\u6807\uff0c\u5b83\u8868\u793a\u7cfb\u7edf\u4e2d\u6bcf\u79d2\u53ef\u4ee5\u5904\u7406\u591a\u5c11\u4e2a\u8bf7\u6c42\uff0c\u5728\u8fdb\u884c\u8fd9\u4e2a\u6d4b\u8bd5\u7684\u65f6\u5019\uff0c\u6211\u4eec\u4f1a\u4f7f\u7528\u5b9e\u9645\u7684\u6570\u636e\u3002\uff08\u5173\u4e8e\u8fd9\u4efd\u6570\u636e\uff0c\u6211\u4eec\u5df2\u7ecf\u505a\u597d\u4e86\u91c7\u6837\uff0c\u5e76\u4e14\u653e\u5728\u4e86 github \u4e0a\u3002\uff09<\/p>\n<p>QPS \u7684\u6d4b\u91cf\u8ddf\u541e\u5410\u91cf\u4e0d\u592a\u4e00\u6837\uff0c\u56e0\u4e3a\u5728\u5b9e\u9645\u4f7f\u7528\u5927\u8bed\u8a00\u6a21\u578b\u7cfb\u7edf\u7684\u65f6\u5019\uff0c\u6bcf\u4e00\u4e2a\u7528\u6237\u5230\u6765\u7684\u65f6\u95f4\u662f\u4e0d\u786e\u5b9a\u7684\u3002\u6709\u7684\u7528\u6237\u53ef\u80fd\u65e9\u6765\uff0c\u6709\u7684\u7528\u6237\u53ef\u80fd\u665a\u6765\uff0c\u5e76\u4e14\u6bcf\u4e00\u4e2a\u7528\u6237\u505a\u5b8c Prefill \u4e4b\u540e\u7684\u751f\u6210\u957f\u5ea6\u4e5f\u662f\u4e0d\u786e\u5b9a\u7684\u3002\u6709\u7684\u7528\u6237\u53ef\u80fd\u751f\u6210 4 \u4e2a\u8bcd\u5c31\u9000\u51fa\uff0c\u6709\u7684\u7528\u6237\u53ef\u80fd\u8981\u751f\u6210 20 \u591a\u4e2a\u8bcd\u3002<\/p>\n<p>\u5728 Prefill \u9636\u6bb5\uff0c\u5728\u5b9e\u9645\u7ebf\u4e0a\u63a8\u7406\u5f53\u4e2d\uff0c\u56e0\u4e3a\u7528\u6237\u5b9e\u9645\u751f\u6210\u957f\u5ea6\u4e0d\u4e00\u6837\uff0c\u6240\u4ee5\u4f1a\u9047\u5230\u4e00\u4e2a\u95ee\u9898\uff1a\u6709\u4e9b\u7528\u6237\u4f1a\u63d0\u524d\u751f\u6210\u5b8c\uff0c\u800c\u6709\u4e9b\u7528\u6237\u8981\u751f\u6210\u5f88\u591a\u957f\u5ea6\u4e4b\u540e\u624d\u4f1a\u7ed3\u675f\u3002\u5728\u8fd9\u6837\u7684\u751f\u6210\u8fc7\u7a0b\u4e2d\uff0c\u6709\u5f88\u591a\u5730\u65b9\u7684 GPU \u4f1a\u7a7a\u95f2\u3002\u56e0\u6b64\u5728\u5b9e\u9645\u7684\u63a8\u7406\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u7684 QPS \u5e76\u4e0d\u80fd\u591f\u53d1\u6325\u5b8c\u5168\u7684\u541e\u5410\u91cf\u4f18\u52bf\u3002\u6211\u4eec\u7684\u541e\u5410\u91cf\u53ef\u80fd\u5f88\u5927\uff0c\u4f46\u5b9e\u9645\u7684\u5904\u7406\u80fd\u529b\u53ef\u80fd\u4f1a\u5f88\u5dee\uff0c\u56e0\u4e3a\u5728\u8fd9\u4e2a\u5904\u7406\u8fc7\u7a0b\u5f53\u4e2d\u5145\u6ee1\u4e86\u65e0\u6cd5\u4f7f\u7528\u663e\u5361\u7684\u7a7a\u6d1e\u3002\u6240\u4ee5\u5728 QPS \u6307\u6807\u4e0a\uff0c\u6211\u4eec\u4f1a\u6709\u975e\u5e38\u591a\u7684\u5177\u4f53\u7684\u4f18\u5316\u65b9\u6848\uff0c\u907f\u514d\u8ba1\u7b97\u7684\u7a7a\u6d1e\u6216\u8005\u65e0\u6cd5\u6709\u6548\u5229\u7528\u663e\u5361\u7684\u73b0\u8c61\u5b58\u5728\uff0c\u4ece\u800c\u4f7f\u5f97\u541e\u5410\u91cf\u80fd\u591f\u5b8c\u5168\u670d\u52a1\u5230\u7528\u6237\u4e0a\u3002<\/p>\n<h2>\u4e8c\u3001\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u6027\u80fd\u4f18\u5316<\/h2>\n<p>\u63a5\u4e0b\u6765\u8fdb\u5165\u5230\u5927\u8bed\u8a00\u6a21\u578b\u7684\u63a8\u7406\u6d41\u7a0b\u5f53\u4e2d\uff0c\u770b\u770b\u6211\u4eec\u7a76\u7adf\u505a\u4e86\u54ea\u4e9b\u4f18\u5316\uff0c\u4f7f\u5f97\u7cfb\u7edf\u5728 QPS \u4ee5\u53ca\u541e\u5410\u91cf\u7b49\u6307\u6807\u4e0a\u90fd\u8fbe\u5230\u6bd4\u8f83\u4f18\u79c0\u7684\u60c5\u51b5\u3002<\/p>\n<h3>1.\u00a0LLM\u00a0\u63a8\u7406\u8fc7\u7a0b<\/h3>\n<p>&nbsp;<\/p>\n<p>\u9996\u5148\u6765\u8be6\u7ec6\u4ecb\u7ecd\u4e00\u4e0b\u5927\u8bed\u8a00\u6a21\u578b\u7684\u63a8\u7406\u8fc7\u7a0b\uff0c\u524d\u6587\u4e2d\u63d0\u5230\u4e86\u6bcf\u4e2a\u8bf7\u6c42\u90fd\u8981\u7ecf\u5386 prefill \u548c decoding \u4e24\u4e2a\u9636\u6bb5\uff0c\u5728 prefill \u9636\u6bb5\uff0c\u81f3\u5c11\u8981\u505a\u56db\u4ef6\u4e8b\u60c5\uff1a<\/p>\n<p>\u7b2c\u4e00\u4ef6\u4e8b\u60c5\u662f\u628a\u7528\u6237\u7684\u8f93\u5165\u8fdb\u884c\u5411\u91cf\u5316\uff0ctokenize \u7684\u8fc7\u7a0b\u6307\u7684\u662f\u5c06\u7528\u6237\u8f93\u5165\u7684\u6587\u672c\u8f6c\u6362\u4e3a\u5411\u91cf\uff0c\u76f8\u5bf9\u4e8e prefill \u6574\u4e2a\u9636\u6bb5\u6765\u8bf4\uff0c\u5927\u6982\u8981\u5360\u6389 10% \u7684\u65f6\u95f4\uff0c\u8fd9\u662f\u6709\u4ee3\u4ef7\u7684\u3002<\/p>\n<p>\u4e4b\u540e\u5c31\u4f1a\u8fdb\u884c\u771f\u6b63\u7684 prefill \u8ba1\u7b97\uff0c\u8fd9\u4e00\u8fc7\u7a0b\u4f1a\u5360\u6389\u5927\u6982 80% \u7684\u65f6\u95f4\u3002<\/p>\n<p>\u8ba1\u7b97\u4e4b\u540e\u4f1a\u8fdb\u884c sampling\uff0c\u8fd9\u4e2a\u8fc7\u7a0b\u5728 Pytorch \u91cc\u9762\u4e00\u822c\u4f1a\u7528 sample\u3001top p\u3002\u5728\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u5f53\u4e2d\u4f1a\u7528 argmax\u3002\u603b\u800c\u8a00\u4e4b\uff0c\u662f\u6839\u636e\u6a21\u578b\u7684\u7ed3\u679c\uff0c\u751f\u6210\u6700\u540e\u8bcd\u7684\u4e00\u4e2a\u8fc7\u7a0b\u3002\u8fd9\u4e2a\u8fc7\u7a0b\u4f1a\u5360\u6389 10% \u7684\u65f6\u95f4\u3002<\/p>\n<p>\u6700\u540e\u5c06 refill \u7684\u7ed3\u679c\u8fd4\u56de\u7ed9\u5ba2\u6237\uff0c\u8fd9\u9700\u8981\u7684\u65f6\u95f4\u4f1a\u6bd4\u8f83\u77ed\uff0c\u5927\u6982\u5360 2% \u5230 5% \u7684\u65f6\u95f4\u3002<\/p>\n<p>Decoding \u9636\u6bb5\u4e0d\u9700\u8981 tokenize\uff0c\u6bcf\u4e00\u6b21\u505a decoding \u90fd\u4f1a\u76f4\u63a5\u4ece\u8ba1\u7b97\u5f00\u59cb\uff0c\u6574\u4e2adecoding \u8fc7\u7a0b\u4f1a\u5360\u6389 80% \u7684\u65f6\u95f4\uff0c\u800c\u540e\u9762\u7684 sampling\uff0c\u4e5f\u5c31\u662f\u91c7\u6837\u751f\u6210\u8bcd\u7684\u8fc7\u7a0b\uff0c\u4e5f\u8981\u5360\u6389 10% \u7684\u65f6\u95f4\u3002\u4f46\u5b83\u4f1a\u6709\u4e00\u4e2a detokenize \u7684\u65f6\u95f4\uff0cdetokenize \u662f\u6307\u751f\u6210\u4e86\u4e00\u4e2a\u8bcd\u4e4b\u540e\uff0c\u8fd9\u4e2a\u751f\u6210\u7684\u8bcd\u662f\u4e2a\u5411\u91cf\uff0c\u9700\u8981\u628a\u5b83\u89e3\u7801\u56de\u6587\u672c\uff0c\u8fd9\u4e00\u64cd\u4f5c\u5927\u6982\u4f1a\u5360\u6389 5% \u7684\u65f6\u95f4\uff0c\u6700\u540e\u5c06\u8fd9\u4e2a\u751f\u6210\u7684\u8bcd\u8fd4\u56de\u7ed9\u7528\u6237\u3002<\/p>\n<p>\u65b0\u7684\u8bf7\u6c42\u8fdb\u6765\uff0c\u5728\u8fdb\u884c\u5b8c prefill \u4e4b\u540e\uff0c\u4f1a\u4e0d\u65ad\u8fed\u4ee3\u8fdb\u884c decoding\uff0c\u6bcf\u4e00\u4e2a decoding \u9636\u6bb5\u7ed3\u675f\u4e4b\u540e\uff0c\u90fd\u4f1a\u5c06\u7ed3\u679c\u5f53\u573a\u8fd4\u56de\u7ed9\u5ba2\u6237\u3002\u8fd9\u6837\u7684\u751f\u6210\u8fc7\u7a0b\u5728\u5927\u8bed\u8a00\u6a21\u578b\u91cc\u9762\u662f\u5f88\u5e38\u89c1\u7684\uff0c\u6211\u4eec\u79f0\u8fd9\u6837\u7684\u65b9\u5f0f\u4e3a\u6d41\u5f0f\u4f20\u8f93\u3002<\/p>\n<h3>2.\u00a0\u4f18\u5316\uff1a\u6d41\u6c34\u7ebf\u524d\u540e\u5904\u7406\u4e0e\u9ad8\u6027\u80fd\u91c7\u6837<\/h3>\n<p>&nbsp;<\/p>\n<p>\u8fd9\u91cc\u8981\u4ecb\u7ecd\u7684\u7b2c\u4e00\u4e2a\u4f18\u5316\u662f\u6d41\u6c34\u7ebf\u4f18\u5316\uff0c\u5176\u76ee\u7684\u662f\u5c3d\u53ef\u80fd\u8ba9\u663e\u5361\u5229\u7528\u7387\u5360\u6ee1\u3002<\/p>\n<p>\u5728\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u8fc7\u7a0b\u4e2d\uff0ctokenize\u3001fast sample \u548c detokenize \u8fd9\u4e9b\u8fc7\u7a0b\u90fd\u4e0e\u6a21\u578b\u7684\u8ba1\u7b97\u65e0\u5173\u3002\u6211\u4eec\u53ef\u4ee5\u628a\u6574\u4e2a\u5927\u8bed\u8a00\u6a21\u578b\u7684\u63a8\u7406\u60f3\u8c61\u6210\u8fd9\u6837\u4e00\u4e2a\u8fc7\u7a0b\uff0c\u5728\u6267\u884c prefill \u7684\u8fc7\u7a0b\u4e2d\uff0c\u5f53\u6211\u62ff\u5230\u4e86 fast sample \u7684\u8bcd\u5411\u91cf\u4e4b\u540e\uff0c\u5c31\u53ef\u4ee5\u7acb\u523b\u5f00\u59cb\u4e0b\u4e00\u4e2a\u9636\u6bb5 decoding\uff0c\u4e0d\u7528\u7b49\u5230\u7ed3\u679c\u8fd4\u56de\uff0c\u56e0\u4e3a\u7ed3\u679c\u5df2\u7ecf\u5728 GPU \u4e0a\u4e86\u3002\u800c\u5f53\u5b8c\u6210\u4e86\u4e00\u6b21 decoding \u540e\uff0c\u4e0d\u7528\u7b49\u5f85 detokenize \u7684\u5b8c\u6210\uff0c\u53ef\u4ee5\u7acb\u523b\u5f00\u59cb\u4e0b\u4e00\u6b21\u7684 decoding\u3002\u56e0\u4e3a detokenize \u662f\u4e2a CPU \u8fc7\u7a0b\uff0c\u540e\u9762\u8fd9\u4e24\u4e2a\u8fc7\u7a0b\uff0c\u53ea\u6d89\u53ca\u5230\u7528\u6237\u7684\u7ed3\u679c\u8fd4\u56de\uff0c\u4e0d\u6d89\u53ca\u4efb\u4f55 GPU \u7684\u8fd0\u7b97\u3002\u5e76\u4e14\u5728\u6267\u884c\u5b8c\u91c7\u6837\u8fc7\u7a0b\u4e4b\u540e\uff0c\u5c31\u5df2\u7ecf\u77e5\u9053\u4e0b\u4e00\u4e2a\u751f\u6210\u7684\u8bcd\u662f\u4ec0\u4e48\u4e86\uff0c\u6211\u4eec\u5df2\u7ecf\u62ff\u5230\u4e86\u6240\u9700\u7684\u6240\u6709\u6570\u636e\uff0c\u53ef\u4ee5\u7acb\u523b\u5f00\u59cb\u4e0b\u4e00\u6b21\u8fd0\u7b97\uff0c\u4e0d\u9700\u8981\u518d\u7b49\u5f85\u540e\u9762\u4e24\u4e2a\u8fc7\u7a0b\u7684\u5b8c\u6210\u3002<\/p>\n<p>\u5728 PPL.LLM \u7684\u5b9e\u73b0\u5f53\u4e2d\u4f7f\u7528\u4e86\u4e09\u4e2a\u7ebf\u7a0b\u6c60\uff1a<\/p>\n<p>\u7b2c\u4e00\u4e2a\u7ebf\u7a0b\u6c60\u8d1f\u8d23\u6267\u884c tokenize \u8fc7\u7a0b\uff1b<\/p>\n<p>\u7b2c\u4e09\u4e2a\u7ebf\u7a0b\u6c60\u8d1f\u8d23\u6267\u884c\u540e\u9762\u7684 fast sample \u4ee5\u53ca\u8fd4\u56de\u7ed3\u679c\u7684\u8fc7\u7a0b\u548c detokenize\uff1b<\/p>\n<p>\u4e2d\u95f4\u7684\u7ebf\u7a0b\u6c60\u7528\u6765\u6267\u884c computing \u7684\u8fc7\u7a0b\u3002<\/p>\n<p>\u8fd9\u4e09\u4e2a\u7ebf\u7a0b\u6c60\u4e92\u76f8\u5f02\u6b65\u5730\u628a\u8fd9\u4e09\u90e8\u5206\u7684\u5ef6\u8fdf\u76f8\u4e92\u9694\u79bb\uff0c\u4ece\u800c\u5c3d\u53ef\u80fd\u5730\u5c06\u8fd9\u4e09\u90e8\u5206\u7684\u5ef6\u8fdf\u63a9\u853d\u6389\u3002\u8fd9\u5c06\u7ed9\u7cfb\u7edf\u5e26\u6765 10% \u5230 20% \u7684 QPS \u63d0\u5347\uff0c\u8fd9\u5c31\u662f\u6211\u4eec\u6240\u505a\u7684\u7b2c\u4e00\u9879\u4f18\u5316\u3002<\/p>\n<h3>3.\u00a0\u4f18\u5316\uff1a\u52a8\u6001\u6279\u5904\u7406<\/h3>\n<p>&nbsp;<\/p>\n<p>\u5728\u8fd9\u4e4b\u540e\uff0cPPL.LLM \u8fd8\u53ef\u4ee5\u6267\u884c\u4e00\u9879\u66f4\u6709\u610f\u601d\u7684\u4f18\u5316\uff0c\u53eb\u505a\u52a8\u6001\u6279\u5904\u7406\u3002<\/p>\n<p>\u524d\u6587\u4e2d\u63d0\u5230\uff0c\u5728\u5b9e\u9645\u7684\u63a8\u7406\u8fc7\u7a0b\u5f53\u4e2d\uff0c\u7528\u6237\u7684\u751f\u6210\u957f\u5ea6\u4e0d\u540c\uff0c\u5e76\u4e14\u7528\u6237\u5230\u8fbe\u7684\u65f6\u95f4\u4e5f\u5e76\u4e0d\u4e00\u6837\u3002\u56e0\u6b64\u4f1a\u5b58\u5728\u8fd9\u6837\u4e00\u79cd\u60c5\u51b5\uff0c\u5982\u679c\u5f53\u524d\u7684 GPU \u5728\u63a8\u7406\u8fc7\u7a0b\u5f53\u4e2d\uff0c\u5df2\u7ecf\u6709\u4e00\u4e2a\u8bf7\u6c42\u5728\u7ebf\u4e0a\u8fdb\u884c\u63a8\u7406\uff0c\u5728\u63a8\u7406\u8fdb\u884c\u5230\u4e00\u534a\u65f6\uff0c\u7b2c\u4e8c\u4e2a\u8bf7\u6c42\u63d2\u5165\u8fdb\u6765\uff0c\u8fd9\u65f6\u7b2c\u4e8c\u4e2a\u8bf7\u6c42\u7684\u751f\u6210\u8fc7\u7a0b\u4f1a\u8ddf\u7b2c\u4e00\u4e2a\u8bf7\u6c42\u7684\u751f\u6210\u8fc7\u7a0b\u76f8\u51b2\u7a81\u3002\u56e0\u4e3a\u6211\u4eec\u53ea\u6709\u4e00\u4e2a GPU\uff0c\u8fd9\u4e2a GPU \u4e0a\u53ea\u80fd\u591f\u4e32\u5f62\u5730\u8dd1\u4efb\u52a1\uff0c\u6240\u4ee5\u4e0d\u80fd\u7b80\u5355\u5730\u628a\u5b83\u4eec\u5728 GPU \u4e0a\u505a\u5e76\u884c\u3002<\/p>\n<p>\u6211\u4eec\u7684\u505a\u6cd5\u662f\uff0c\u5728\u7b2c\u4e8c\u4e2a\u8bf7\u6c42\u8fdb\u5165\u7684\u65f6\u95f4\u70b9\uff0c\u628a\u5b83\u7684 prefill \u9636\u6bb5\u548c\u7b2c\u4e00\u4e2a\u8bf7\u6c42\u5bf9\u5e94\u7684decoding \u9636\u6bb5\u8fdb\u884c\u6df7\u5408\uff0c\u751f\u6210\u4e00\u4e2a\u65b0\u7684\u9636\u6bb5\u79f0\u4e3a Merge Step\u3002\u5728\u8fd9\u4e2a Merge Step \u4e2d\uff0c\u4e0d\u4ec5\u4f1a\u8fdb\u884c\u7b2c\u4e00\u4e2a\u8bf7\u6c42\u7684 decoding\uff0c\u540c\u65f6\u4f1a\u8fdb\u884c\u7b2c\u4e8c\u4e2a\u8bf7\u6c42\u7684 Prefill\u3002\u8fd9\u9879\u529f\u80fd\u5728\u8bb8\u591a\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u7cfb\u7edf\u4e2d\u90fd\u4f1a\u5b58\u5728\uff0c\u5b83\u7684\u5b9e\u73b0\u4f7f\u5f97\u5927\u8bed\u8a00\u6a21\u578b\u7684 QPS \u63d0\u5347\u8fbe\u5230\u4e86 100%\u3002<\/p>\n<p>\u5177\u4f53\u8fc7\u7a0b\u4e3a\uff0c\u7b2c\u4e00\u4e2a\u8bf7\u6c42\u751f\u6210\u8fc7\u7a0b\u8fdb\u884c\u4e86\u4e00\u534a\uff0c\u610f\u5473\u7740\u5b83\u5728\u8fdb\u884c decoding \u65f6\u4f1a\u6709\u4e00\u4e2a\u957f\u5ea6\u4e3a 1 \u7684\u8f93\u5165\uff0c\u800c\u7b2c\u4e8c\u4e2a\u8bf7\u6c42\u662f\u65b0\u8fdb\u5165\u7684\uff0c\u5728\u8fdb\u884c Prefill \u7684\u8fc7\u7a0b\u5f53\u4e2d\uff0c\u4f1a\u6709\u4e00\u4e2a\u957f\u5ea6\u4e3a 48 \u7684\u8f93\u5165\u3002\u5c06\u8fd9\u4e24\u4e2a\u8f93\u5165\u6cbf\u7740\u7b2c\u4e00\u4e2a\u7ef4\u5ea6\u76f8\u4e92\u62fc\u63a5\uff0c\u62fc\u63a5\u5b8c\u7684\u8f93\u5165\u957f\u5ea6\u4e3a 49\uff0c\u5e76\u4e14 hidden dimension \u662f 4096 \u7684\u8f93\u5165\u3002\u5728\u8fd9\u957f\u5ea6\u4e3a 49 \u7684\u8f93\u5165\u5f53\u4e2d\uff0c\u7b2c\u4e00\u4e2a\u8bcd\u662f\u7b2c\u4e00\u4e2a\u8bf7\u6c42\u7684\uff0c\u5269\u4e0b\u7684 48 \u4e2a\u8bcd\u662f\u7b2c\u4e8c\u4e2a\u8bf7\u6c42\u7684\u3002<\/p>\n<p>\u7531\u4e8e\u5728\u5927\u6a21\u578b\u63a8\u7406\u5f53\u4e2d\uff0c\u6240\u9700\u8981\u7ecf\u5386\u7684\u7b97\u5b50\uff0c\u6bd4\u5982 RMSNorm\u3001\u77e9\u9635\u4e58\u548c attention \u7b49\u7b97\u5b50\uff0c\u4e0d\u8bba\u662f\u505a decoding \u8fd8\u662f\u505a prefill\uff0c\u5b83\u4eec\u7684\u7ed3\u6784\u90fd\u662f\u4e0d\u53d8\u7684\u3002\u56e0\u6b64\u62fc\u63a5\u5b8c\u7684\u8f93\u5165\u53ef\u4ee5\u76f4\u63a5\u653e\u5165\u5230\u6574\u4e2a\u7f51\u7edc\u4e2d\u53bb\u8dd1\u3002\u6211\u4eec\u53ea\u9700\u8981\u5728\u4e00\u4e2a\u5730\u65b9\u52a0\u4ee5\u533a\u5206\uff0c\u90a3\u5c31\u662f attention\u3002\u5728 attention \u7684\u8fc7\u7a0b\u5f53\u4e2d\u6216\u8005\u5728\u6267\u884c self attention \u7b97\u5b50\u7684\u8fc7\u7a0b\u5f53\u4e2d\uff0c\u6211\u4eec\u4f1a\u505a\u4e00\u6b21\u6570\u636e\u5206\u6d41\uff0c\u5c06\u6240\u6709\u505a decoding \u7684\u8bf7\u6c42\u5206\u6d41\u6210\u4e00\u6ce2\uff0c\u628a\u6240\u6709\u505a prefill \u7684\u8bf7\u6c42\u5206\u6d41\u5230\u53e6\u5916\u4e00\u6ce2\uff0c\u6267\u884c\u4e24\u4e2a\u4e0d\u540c\u7684\u8fd0\u7b97\u3002\u6240\u6709\u505a prefill \u7684\u8bf7\u6c42\uff0c\u5c06\u4f1a\u6267\u884c Flash Attention\uff1b\u6240\u6709\u505a decoding \u7684\u7528\u6237\uff0c\u5c06\u4f1a\u6267\u884c\u4e00\u4e2a\u975e\u5e38\u7279\u6b8a\u7684\u7b97\u5b50\uff0c\u53eb\u505a Decoding Attention\u3002\u5728\u5206\u6d41\u6267\u884c\u5b8c attention \u7b97\u5b50\u4e4b\u540e\uff0c\u8fd9\u4e9b\u7528\u6237\u7684\u8f93\u5165\u4f1a\u518d\u4e00\u6b21\u88ab\u62fc\u63a5\u5230\u4e00\u8d77\uff0c\u5b8c\u6210\u5176\u5b83\u7b97\u5b50\u7684\u8ba1\u7b97\u3002<\/p>\n<p>\u5bf9\u4e8e Merge Step\uff0c\u5b9e\u9645\u4e0a\u5f53\u6bcf\u4e2a\u8bf7\u6c42\u5230\u6765\u7684\u65f6\u5019\uff0c\u6211\u4eec\u90fd\u4f1a\u628a\u8fd9\u4e2a\u8bf7\u6c42\u8ddf\u7cfb\u7edf\u4e0a\u73b0\u5728\u5df2\u6709\u7684\u6240\u6709\u8bf7\u6c42\u7684\u8f93\u5165\u62fc\u63a5\u5728\u4e00\u8d77\uff0c\u5b8c\u6210\u8fd9\u6b21\u8ba1\u7b97\uff0c\u7136\u540e\u7ee7\u7eed\u5f80\u4e0b\u4e0d\u505c\u5730\u505a decoding\uff0c\u8fd9\u662f\u52a8\u6001\u6279\u5904\u7406\u5728\u5927\u8bed\u8a00\u6a21\u578b\u4e2d\u7684\u5b9e\u73b0\u3002<\/p>\n<h3>4.\u00a0\u4f18\u5316\uff1aDecoding Attention<\/h3>\n<p>&nbsp;<\/p>\n<p>Decoding Attention \u7b97\u5b50\uff0c\u4e0d\u50cf Flash Attention \u7b97\u5b50\u90a3\u6837\u51fa\u540d\uff0c\u4f46\u5176\u5b9e\u5728\u5904\u7406 decoding \u4efb\u52a1\u4e0a\u6bd4 Flash Attention \u8981\u5feb\u5f97\u591a\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u8fd9\u662f\u4e00\u79cd\u4e13\u95e8\u4e3a decoding \u4efb\u52a1\u6240\u8bbe\u8ba1\u7684\u7b97\u5b50\uff0c\u5b8c\u5168\u4f9d\u8d56 Cuda Core\uff0c\u4e0d\u4f1a\u4f9d\u8d56Tensor Core \u5b8c\u6210\u8ba1\u7b97\u3002\u5b83\u975e\u5e38\u7075\u6d3b\u5e76\u4e14\u5bb9\u6613\u4fee\u6539\uff0c\u4f46\u5b83\u6709\u4e00\u4e2a\u9650\u5236\uff0c\u56e0\u4e3a\u5176\u7279\u70b9\u662f\u5728 decoding \u7684 tensor \u7684\u8fd0\u7b97\u5f53\u4e2d\uff0c\u6240\u4ee5\u4f1a\u8981\u6c42\u8f93\u5165\u7684 q \u7684\u957f\u5ea6\u5fc5\u987b\u662f 1\uff0c\u4f46 k \u548c v \u7684\u957f\u5ea6\u662f\u53ef\u53d8\u7684\u3002\u8fd9\u662f Decoding Attention \u7684\u9650\u5236\uff0c\u5728\u8fd9\u79cd\u9650\u5236\u4e0b\uff0c\u6211\u4eec\u53ef\u4ee5\u505a\u4e00\u4e9b\u7279\u5b9a\u7684\u4f18\u5316\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u8fd9\u79cd\u7279\u5b9a\u7684\u4f18\u5316\u4f7f\u5f97\u5728 decoding \u9636\u6bb5\u7684 attention \u7b97\u5b50\u7684\u5b9e\u73b0\uff0c\u4f1a\u6bd4 Flash Attention \u66f4\u5feb\u3002\u8fd9\u4e2a\u5b9e\u73b0\u76ee\u524d\u4e5f\u5df2\u7ecf\u5f00\u6e90\uff0c\u5927\u5bb6\u53ef\u4ee5\u5230\u4e0a\u56fe\u4e2d\u7684\u7f51\u5740\u8fdb\u884c\u8bbf\u95ee\u3002<\/p>\n<h3>5.\u00a0\u4f18\u5316\uff1aVM Allocator<\/h3>\n<p>&nbsp;<\/p>\n<p>\u53e6\u4e00\u9879\u4f18\u5316\u662f Virtual Memory Allocator\uff0c\u5bf9\u5e94 Page Attention \u4f18\u5316\u3002\u5f53\u8bf7\u6c42\u6765\u5230\u4e4b\u540e\uff0c\u8981\u8fdb\u884c prefill \u9636\u6bb5\uff0c\u53c8\u8981\u8fdb\u884c decoding \u9636\u6bb5\uff0c\u5b83\u6240\u6709\u8f93\u5165\u7684 token \u4f1a\u751f\u6210\u4e00\u4e2a KV \u7f13\u5b58\uff0c\u8fd9\u4e2aKV \u7f13\u5b58\u8bb0\u5f55\u4e86\u8fd9\u4e2a\u8bf7\u6c42\u6240\u6709\u7684\u5386\u53f2\u4fe1\u606f\u3002\u90a3\u4e48\u8981\u7ed9\u8fd9\u6837\u4e00\u4e2a\u8bf7\u6c42\u5206\u914d\u591a\u957f\u7684 KV \u7f13\u5b58\u7a7a\u95f4\uff0c\u624d\u80fd\u6ee1\u8db3\u5b83\u5b8c\u6210\u6b64\u6b21\u751f\u6210\u4efb\u52a1\u5462\uff1f\u5982\u679c\u5206\u7684\u592a\u591a\uff0c\u663e\u5b58\u4f1a\u6709\u6d6a\u8d39\uff0c\u5982\u679c\u5206\u7684\u592a\u5c11\uff0c\u5728 decoding \u9636\u6bb5\uff0c\u78b0\u5230\u4e86 KV \u7f13\u5b58\u7684\u622a\u6b62\u4f4d\u7f6e\uff0c\u5c31\u6ca1\u6709\u529e\u6cd5\u7ee7\u7eed\u5f80\u4e0b\u751f\u6210\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u6709 3 \u79cd\u65b9\u6848\u3002<\/p>\n<p>Pytorch \u7684\u663e\u5b58\u7ba1\u7406\u65b9\u5f0f\u662f\u4e3a\u6bcf\u4e00\u4e2a\u8bf7\u6c42\u9884\u7559\u4e00\u7247\u8db3\u591f\u957f\u7684\u7a7a\u95f4\uff0c\u901a\u5e38\u662f 2048 \u6216\u8005 4096\uff0c\u80fd\u591f\u4fdd\u8bc1\u5b8c\u6210 4096 \u4e2a\u8bcd\u7684\u751f\u6210\u3002\u4f46\u5927\u90e8\u5206\u7528\u6237\u5b9e\u9645\u7684\u751f\u6210\u957f\u5ea6\u4e0d\u4f1a\u6709\u90a3\u4e48\u957f\uff0c\u6240\u4ee5\u4f1a\u6709\u5927\u91cf\u7684\u5185\u5b58\u7a7a\u95f4\u88ab\u6d6a\u8d39\u6389\u3002<\/p>\n<p>Page Attention \u91c7\u7528\u7684\u662f\u53e6\u5916\u4e00\u79cd\u663e\u5b58\u7ba1\u7406\u65b9\u5f0f\u3002\u5141\u8bb8\u751f\u6210\u8fc7\u7a0b\u4e2d\u4e0d\u65ad\u4e3a\u7528\u6237\u8ffd\u52a0\u663e\u5b58\u3002\u7c7b\u4f3c\u4e8e\u64cd\u4f5c\u7cfb\u7edf\u4e2d\u7684\u9875\u5f0f\u5b58\u50a8\u6216\u8005\u5185\u5b58\u5206\u9875\u3002\u5f53\u4e00\u4e2a\u8bf7\u6c42\u6765\u5230\u4e4b\u540e\uff0c\u7cfb\u7edf\u4f1a\u4e3a\u8fd9\u4e2a\u8bf7\u6c42\u5206\u914d\u4e00\u5c0f\u5757\u663e\u5b58\uff0c\u8fd9\u4e00\u5c0f\u5757\u663e\u5b58\u901a\u5e38\u53ea\u591f\u751f\u6210 8 \u4e2a\u5b57\u7b26\uff0c\u5f53\u8bf7\u6c42\u751f\u6210\u4e86 8 \u4e2a\u5b57\u7b26\u4e4b\u540e\uff0c\u7cfb\u7edf\u4f1a\u8ffd\u52a0\u4e00\u5757\u663e\u5b58\uff0c\u53ef\u4ee5\u628a\u7ed3\u679c\u518d\u5199\u5230\u8fd9\u5757\u663e\u5b58\u91cc\u9762\uff0c\u540c\u65f6\u7cfb\u7edf\u4f1a\u7ef4\u62a4\u4e00\u4e2a\u663e\u5b58\u5757\u548c\u663e\u5b58\u5757\u4e4b\u95f4\u7684\u94fe\u8868\uff0c\u4ece\u800c\u4f7f\u5f97\u7b97\u5b50\u53ef\u4ee5\u6b63\u5e38\u5730\u8fdb\u884c\u8f93\u51fa\u3002\u5f53\u751f\u6210\u7684\u957f\u5ea6\u4e0d\u65ad\u53d8\u957f\u65f6\uff0c\u4f1a\u4e0d\u65ad\u5730\u7ed9\u7528\u6237\u8ffd\u52a0\u663e\u5b58\u5757\u7684\u5206\u914d\uff0c\u5e76\u4e14\u53ef\u4ee5\u52a8\u6001\u7ef4\u62a4\u663e\u5b58\u5757\u5206\u914d\u7684\u5217\u8868\uff0c\u4f7f\u7cfb\u7edf\u4e0d\u4f1a\u5b58\u5728\u5927\u91cf\u6d6a\u8d39\u7684\u8d44\u6e90\uff0c\u4e0d\u9700\u8981\u4e3a\u8fd9\u4e2a\u8bf7\u6c42\u4fdd\u7559\u592a\u591a\u7684\u663e\u5b58\u7a7a\u95f4\u3002<\/p>\n<p>PPL.LLM \u4f7f\u7528\u7684\u662f Virtual Memory \u7684\u7ba1\u7406\u673a\u5236\uff0c\u4e3a\u6bcf\u4e00\u4e2a\u8bf7\u6c42\u9884\u6d4b\u4e00\u4e2a\u5b83\u6240\u9700\u7684\u751f\u6210\u957f\u5ea6\u3002\u6bcf\u4e2a\u8bf7\u6c42\u8fdb\u6765\u4e4b\u540e\uff0c\u90fd\u4f1a\u76f4\u63a5\u4e3a\u5176\u5206\u914d\u4e00\u4e2a\u8fde\u7eed\u7684\u7a7a\u95f4\uff0c\u8fd9\u4e2a\u8fde\u7eed\u7a7a\u95f4\u7684\u957f\u5ea6\u662f\u9884\u6d4b\u51fa\u6765\u7684\u3002\u4f46\u7406\u8bba\u4e0a\u770b\u53ef\u80fd\u96be\u4ee5\u5b9e\u73b0\uff0c\u5c24\u5176\u5230\u4e86\u7ebf\u4e0a\u63a8\u7406\u9636\u6bb5\uff0c\u4e0d\u592a\u53ef\u80fd\u6e05\u695a\u5730\u77e5\u9053\u6bcf\u4e2a\u8bf7\u6c42\u7a76\u7adf\u8981\u751f\u6210\u591a\u957f\u7684\u5185\u5bb9\u3002\u56e0\u6b64\u6211\u4eec\u63a8\u8350\u8bad\u7ec3\u4e00\u4e2a\u6a21\u578b\u53bb\u505a\u8fd9\u4ef6\u4e8b\u60c5\u3002\u56e0\u4e3a\u5373\u4f7f\u6211\u4eec\u91c7\u7528\u4e86 Page Attention \u8fd9\u6837\u7684\u6a21\u5f0f\uff0c\u4f9d\u7136\u4f1a\u9047\u5230\u95ee\u9898\u3002Page Attention \u5728\u8fd0\u884c\u7684\u8fc7\u7a0b\u4e2d\uff0c\u5177\u4f53\u5230\u4e00\u4e2a\u7279\u5b9a\u7684\u65f6\u95f4\u70b9\uff0c\u6bd4\u5982\u5f53\u524d\u7cfb\u7edf\u4e0a\u5df2\u7ecf\u6709\u4e86\u56db\u4e2a\u8bf7\u6c42\uff0c\u7cfb\u7edf\u91cc\u9762\u8fd8\u5269\u4f59 6 \u5757\u663e\u5b58\u6ca1\u6709\u88ab\u5206\u914d\u3002\u8fd9\u65f6\u6211\u4eec\u65e0\u6cd5\u77e5\u9053\u662f\u5426\u4f1a\u6709\u65b0\u7684\u8bf7\u6c42\u8fdb\u6765\uff0c\u80fd\u5426\u4e3a\u5176\u7ee7\u7eed\u63d0\u4f9b\u670d\u52a1\uff0c\u56e0\u4e3a\u5f53\u524d\u7684\u56db\u4e2a\u8bf7\u6c42\u8fd8\u6ca1\u6709\u7ed3\u675f\uff0c\u53ef\u80fd\u672a\u6765\u8fd8\u8981\u7ee7\u7eed\u4e3a\u5b83\u4eec\u8ffd\u52a0\u65b0\u7684\u663e\u5b58\u5757\u3002\u6240\u4ee5\u5373\u4f7f\u662f Page Attention \u673a\u5236\uff0c\u8fd8\u662f\u9700\u8981\u9884\u6d4b\u6bcf\u4e00\u4e2a\u7528\u6237\u5b9e\u9645\u7684\u751f\u6210\u957f\u5ea6\u3002\u8fd9\u6837\u624d\u77e5\u9053\u5728\u5177\u4f53\u7684\u4e00\u4e2a\u65f6\u95f4\u70b9\u4e0a\u80fd\u4e0d\u80fd\u63a5\u53d7\u4e00\u4e2a\u65b0\u7684\u7528\u6237\u7684\u8f93\u5165\u3002<\/p>\n<p>\u8fd9\u662f\u6211\u4eec\u76ee\u524d\u6240\u6709\u7684\u63a8\u7406\u7cfb\u7edf\u90fd\u6ca1\u6709\u505a\u5230\u7684\u4e8b\u60c5\uff0c\u5305\u62ec PPL \u76ee\u524d\u4e5f\u6ca1\u6709\u5b9e\u73b0\u3002\u4f46 Virtual Memory \u7684\u7ba1\u7406\u673a\u5236\uff0c\u8fd8\u662f\u8ba9\u6211\u4eec\u5f88\u5927\u7a0b\u5ea6\u4e0a\u907f\u514d\u4e86\u663e\u5b58\u7684\u6d6a\u8d39\uff0c\u4ece\u800c\u4f7f\u7cfb\u7edf\u6574\u4f53\u7684 QPS \u63d0\u5347\u8fbe\u5230 200% \u5de6\u53f3\u3002<\/p>\n<h3>6.\u00a0\u4f18\u5316\uff1aKV Cache \u91cf\u5316<\/h3>\n<p>&nbsp;<\/p>\n<p>PPL.LLM \u5728\u505a\u7684\u53e6\u5916\u4e00\u9879\u4f18\u5316\uff0c\u5c31\u662f KV \u7f13\u5b58\u7684\u91cf\u5316\uff0c\u5728\u670d\u52a1\u7aef\u63a8\u7406\u7684\u8fc7\u7a0b\u5f53\u4e2d\uff0cKV \u7f13\u5b58\u4f1a\u5360\u636e\u7edd\u5927\u90e8\u5206\u7684\u663e\u5b58\u7a7a\u95f4\uff0c\u8fd9\u4f1a\u4e25\u91cd\u9650\u5236\u7cfb\u7edf\u7684\u5e76\u53d1\u8bf7\u6c42\u6570\u91cf\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u53ef\u4ee5\u770b\u5230\uff0c\u5728\u670d\u52a1\u7aef\uff0c\u7279\u522b\u662f A100\u3001H100 \u8fd9\u6837\u7684\u5927\u663e\u5b58\u7684\u670d\u52a1\u5668\u4e0a\u8fd0\u884c\u5982 7B \u6a21\u578b\u8fd9\u6837\u7684\u5927\u8bed\u8a00\u6a21\u578b\u65f6\uff0c\u5b83\u7684 KV \u7f13\u5b58\u5c06\u5360\u5230 84% \u7684\u663e\u5b58\u7a7a\u95f4\uff0c\u800c\u5bf9\u4e8e\u5982 176B \u8fd9\u6837\u7684\u5343\u4ebf\u7ea7\u5927\u6a21\u578b\uff0c\u5b83\u7684 KV \u7f13\u5b58\u4e5f\u5c06\u5360\u7528 50% \u4ee5\u4e0a\u7684\u7f13\u5b58\u7a7a\u95f4\u3002\u8fd9\u4f1a\u4e25\u91cd\u5730\u9650\u5236\u6a21\u578b\u7684\u5e76\u53d1\u6570\u91cf\uff0c\u6bcf\u4e00\u4e2a\u8bf7\u6c42\u5230\u6765\u540e\uff0c\u90fd\u9700\u8981\u7ed9\u5b83\u5206\u914d\u5f88\u5927\u7684\u663e\u5b58\u3002\u8fd9\u6837\u8bf7\u6c42\u6570\u91cf\u5c31\u65e0\u6cd5\u63d0\u5347\u4e0a\u53bb\uff0c\u7ee7\u800c\u4f7f\u5f97 QPS \u4ee5\u53ca\u541e\u5410\u91cf\u90fd\u65e0\u6cd5\u63d0\u5347\u3002<\/p>\n<p>PPL.LLM \u4f7f\u7528\u4e86\u4e00\u79cd\u975e\u5e38\u7279\u6b8a\u7684\u91cf\u5316\u65b9\u5f0f\uff0c\u5206\u7ec4\u91cf\u5316\u5bf9 KV \u7f13\u5b58\u7684\u6570\u636e\u8fdb\u884c\u538b\u7f29\u3002\u4e5f\u5c31\u662f\u5bf9\u539f\u6765 FP16 \u7684\u6570\u636e\uff0c\u4f1a\u5c1d\u8bd5\u628a\u5b83\u91cf\u5316\u5230 INT8\u3002\u8fd9\u6837\u4f1a\u4f7f KV \u7f13\u5b58\u7684\u4f53\u79ef\u7f29\u5c0f 50%\uff0c\u5e76\u4f7f\u5f97\u670d\u52a1\u7aef\u80fd\u591f\u5bb9\u7eb3\u7684\u8bf7\u6c42\u6570\u91cf\u589e\u52a0 100%\u3002<\/p>\n<p>\u4e4b\u6240\u4ee5\u76f8\u6bd4 Faster Transformer \u80fd\u591f\u63d0\u5347\u5927\u7ea6 50% \u7684\u541e\u5410\u91cf\uff0c\u6b63\u662f\u5f97\u76ca\u4e8e KV \u7f13\u5b58\u91cf\u5316\u6240\u5e26\u6765\u7684 batch size \u7684\u63d0\u5347\u3002<\/p>\n<h3>7.\u00a0\u4f18\u5316\uff1a\u77e9\u9635\u4e58\u6cd5\u91cf\u5316<\/h3>\n<p>&nbsp;<\/p>\n<p>\u5728 KV \u7f13\u5b58\u91cf\u5316\u4e4b\u540e\uff0c\u6211\u4eec\u8fdb\u884c\u4e86\u66f4\u7ec6\u529b\u5ea6\u7684\u77e9\u9635\u4e58\u6cd5\u7684\u91cf\u5316\u3002\u5728\u6574\u4e2a\u670d\u52a1\u7aef\u63a8\u7406\u7684\u8fc7\u7a0b\u5f53\u4e2d\uff0c\u77e9\u9635\u4e58\u6cd5\u5360\u5230\u6574\u4e2a\u63a8\u7406\u65f6\u95f4\u7684 70% \u4ee5\u4e0a\uff0cPPL.LLM \u4f7f\u7528\u4e86\u4e00\u79cd\u52a8\u6001\u7684 per-channel\/per-token \u4ea4\u66ff\u7684\u6df7\u5408\u91cf\u5316\u65b9\u5f0f\u6765\u52a0\u901f\u77e9\u9635\u4e58\u6cd5\u3002\u8fd9\u4e9b\u91cf\u5316\u540c\u6837\u662f\u7cbe\u5ea6\u6781\u9ad8\u7684\uff0c\u5e76\u4e14\u80fd\u591f\u63d0\u5347\u63a5\u8fd1 100% \u7684\u6027\u80fd\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u5177\u4f53\u505a\u6cd5\u662f\uff0c\u5728 RMSNorm \u7b97\u5b50\u7684\u57fa\u7840\u4e4b\u4e0a\uff0c\u878d\u5408\u4e00\u4e2a\u91cf\u5316\u7b97\u5b50\uff0c\u8fd9\u4e2a\u91cf\u5316\u7b97\u5b50\u4f1a\u5728 RMSNorm \u7b97\u5b50\u7684\u529f\u80fd\u57fa\u7840\u4e4b\u4e0a\u7edf\u8ba1\u5176 Token \u4fe1\u606f\uff0c\u7edf\u8ba1\u6bcf\u4e00\u4e2a token \u7684\u6700\u5927\u6700\u5c0f\u503c\uff0c\u5e76\u4e14\u6cbf\u7740 token \u7684\u7ef4\u5ea6\uff0c\u628a\u8fd9\u4e2a\u6570\u636e\u8fdb\u884c\u91cf\u5316\u3002\u4e5f\u5c31\u662f\u8bf4\u7ecf\u8fc7\u4e86RMSNorm \u4e4b\u540e\u7684\u6570\u636e\u5c06\u4f1a\u4ece FP16 \u8f6c\u6210 INT8\uff0c\u5e76\u4e14\u8fd9\u4e00\u6b21\u91cf\u5316\u662f\u5168\u52a8\u6001\u7684\uff0c\u4e0d\u9700\u8981\u505a calibration\u3002\u800c\u5728\u540e\u9762\u7684 QKV \u77e9\u9635\u4e58\u5f53\u4e2d\uff0c\u8fd9\u4e09\u4e2a\u77e9\u9635\u4e58\u90fd\u5c06\u8fdb\u884c per-channel \u91cf\u5316\u3002\u5b83\u4eec\u63a5\u6536\u7684\u6570\u636e\u662f INT8 \u7684\uff0c\u540c\u6837\u5b83\u4eec\u7684\u6743\u91cd\u4e5f\u662f INT8 \u7684\uff0c\u6240\u4ee5\u8fd9\u4e9b\u77e9\u9635\u4e58\u53ef\u4ee5\u5b8c\u6574\u5730\u6267\u884c INT8 \u7684\u77e9\u9635\u4e58\u6cd5\u3002\u5b83\u4eec\u7684\u8f93\u51fa\u5c06\u4f1a\u88ab Soft Attention \u63a5\u53d7\uff0c\u4f46\u5728\u63a5\u53d7\u4e4b\u524d\u4f1a\u6267\u884c\u4e00\u6b21\u89e3\u91cf\u5316\u8fc7\u7a0b\uff0c\u8fd9\u6b21\u89e3\u91cf\u5316\u8fc7\u7a0b\u5c06\u548c soft attention \u7b97\u5b50\u878d\u5408\u3002<\/p>\n<p>\u800c\u540e\u9762\u7684 O \u77e9\u9635\u4e58\u6cd5\u662f\u4e0d\u505a\u91cf\u5316\u7684\uff0cSoft Attention \u672c\u8eab\u7684\u8ba1\u7b97\u8fc7\u7a0b\u4e5f\u4e0d\u505a\u4efb\u4f55\u91cf\u5316\u3002\u5728\u540e\u7eed\u7684 FeedForward \u8fc7\u7a0b\u5f53\u4e2d\uff0c\u8fd9\u4e24\u4e2a\u77e9\u9635\u540c\u6837\u91c7\u7528\u4e00\u6837\u7684\u65b9\u5f0f\u8fdb\u884c\u91cf\u5316\uff0c\u548c\u4e0a\u9762\u7684 RMSNorm \u8fdb\u884c\u878d\u5408\uff0c\u6216\u8005\u4e0e\u4e0a\u9762\u7684 Silu \u548c Mul \u8fd9\u6837\u7684\u6fc0\u6d3b\u51fd\u6570\u8fdb\u884c\u878d\u5408\u3002\u5b83\u4eec\u7684\u89e3\u91cf\u5316\u7b97\u5b50\u5c06\u548c\u5176\u4e0b\u6e38\u7b97\u5b50\u8fdb\u884c\u878d\u5408\u3002<\/p>\n<h3>8.\u00a0\u4f18\u5316\uff1aINT8 vs INT4<\/h3>\n<p>&nbsp;<\/p>\n<p>\u76ee\u524d\u5b66\u672f\u754c\u5bf9\u4e8e\u5927\u8bed\u8a00\u6a21\u578b\u7684\u91cf\u5316\u5173\u6ce8\u70b9\u53ef\u80fd\u4e3b\u8981\u96c6\u4e2d\u5728 INT4 \u4e0a\uff0c\u4f46\u662f\u5728\u670d\u52a1\u7aef\u63a8\u7406\u7684\u8fc7\u7a0b\u4e2d\uff0c\u5176\u5b9e\u66f4\u9002\u5408\u4f7f\u7528 INT8 \u7684\u91cf\u5316\u3002<\/p>\n<p>INT4 \u7684\u91cf\u5316\u4e5f\u53eb Weight Only \u7684\u91cf\u5316\uff0c\u8fd9\u79cd\u91cf\u5316\u65b9\u5f0f\u51fa\u73b0\u7684\u610f\u4e49\u5728\u4e8e\uff0c\u5f53\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u8fc7\u7a0b\u4e2d batch \u6bd4\u8f83\u5c0f\u65f6\uff0c\u5728\u77e9\u9635\u4e58\u6cd5\u7684\u8ba1\u7b97\u8fc7\u7a0b\u4e2d\uff0c90% \u7684\u65f6\u95f4\u90fd\u4f1a\u7528\u6765\u52a0\u8f7d\u6743\u91cd\u3002\u56e0\u4e3a\u6743\u91cd\u7684\u4f53\u79ef\u975e\u5e38\u5927\uff0c\u800c\u52a0\u8f7d\u8f93\u5165\u7684\u65f6\u95f4\u5f88\u77ed\uff0c\u5b83\u4eec\u7684\u8f93\u5165\uff0c\u5373 activation \u4e5f\u975e\u5e38\u77ed\uff0c\u8ba1\u7b97\u7684\u65f6\u95f4\u4e5f\u4e0d\u4f1a\u5f88\u957f\uff0c\u5199\u56de\u7ed3\u679c\u7684\u65f6\u95f4\u540c\u6837\u4e0d\u4f1a\u5f88\u957f\uff0c\u8fd9\u610f\u5473\u7740\u8fd9\u4e2a\u7b97\u5b50\u662f\u4e00\u4e2a\u8bbf\u5b58\u5bc6\u96c6\u578b\u7684\u7b97\u5b50\u3002\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u6211\u4eec\u4f1a\u9009\u7528 INT4 \u7684\u91cf\u5316\uff0c\u524d\u63d0\u662f batch \u8db3\u591f\u7684\u5c0f\uff0c\u4f7f\u7528 INT4 \u7684\u91cf\u5316\u6bcf\u4e00\u6b21\u52a0\u8f7d\u6743\u91cd\u4e4b\u540e\uff0c\u4f1a\u7d27\u63a5\u7740\u8fdb\u884c\u4e00\u4e2a\u89e3\u91cf\u5316\u7684\u8fc7\u7a0b\u3002\u8fd9\u6b21\u89e3\u91cf\u5316\u4f1a\u628a\u6743\u91cd\u4ece INT4 \u00a0\u89e3\u91cf\u5316\u6210 FP16\uff0c\u7ecf\u5386\u89e3\u91cf\u5316\u8fc7\u7a0b\u4e4b\u540e\uff0c\u540e\u9762\u7684\u8ba1\u7b97\u548c FP16 \u662f\u5b8c\u5168\u4e00\u6837\u7684\uff0c\u4e5f\u5c31\u662f\u8bf4 INT4 Weight Only \u7684\u91cf\u5316\u9002\u7528\u4e8e\u8bbf\u5b58\u5bc6\u96c6\u6027\u7684\u77e9\u9635\u4e58\u6cd5\uff0c\u5176\u8ba1\u7b97\u8fc7\u7a0b\u8fd8\u662f\u7531 FP16 \u7684\u8fd0\u7b97\u5668\u4ef6\u53bb\u5b8c\u6210\u7684\u3002<\/p>\n<p>\u5f53 batch \u8db3\u591f\u5927\uff0c\u6bd4\u5982 64 \u6216\u8005 128 \u65f6\uff0cINT4 \u7684 Weight Only \u91cf\u5316\u5c06\u4e0d\u4f1a\u5e26\u6765\u4efb\u4f55\u6027\u80fd\u63d0\u5347\u3002\u56e0\u4e3a\u5982\u679c batch \u8db3\u591f\u5927\uff0c\u90a3\u8ba1\u7b97\u65f6\u95f4\u4f1a\u88ab\u62c9\u5f97\u5f88\u957f\u3002\u5e76\u4e14 INT4 Weight Only \u91cf\u5316\u6709\u4e00\u4e2a\u975e\u5e38\u4e0d\u597d\u7684\u70b9\uff0c\u5b83\u7684\u89e3\u91cf\u5316\u8fc7\u7a0b\u6240\u9700\u8981\u7684\u8ba1\u7b97\u91cf\u662f\u4f1a\u968f\u7740 batch \u7684\uff08GEMM Batch\uff09\u63d0\u5347\u800c\u63d0\u5347\u7684\uff0c\u968f\u7740\u8f93\u5165 batch \u7684\u63d0\u5347\uff0c\u89e3\u91cf\u5316\u7684\u65f6\u95f4\u4e5f\u4f1a\u8d8a\u6765\u8d8a\u957f\u3002\u5f53 batch \u8fbe\u5230 128 \u7684\u65f6\u5019\uff0c\u89e3\u91cf\u5316\u6240\u5e26\u6765\u7684\u65f6\u95f4\u635f\u8017\u548c\u52a0\u8f7d\u6743\u91cd\u5e26\u6765\u7684\u6027\u80fd\u4f18\u52bf\uff0c\u5c31\u5df2\u7ecf\u76f8\u4e92\u62b5\u6d88\u4e86\u3002\u4e5f\u5c31\u662f\u8bf4\u5f53 batch \u8fbe\u5230 128 \u4e4b\u540e\uff0cINT4 \u7684\u77e9\u9635\u91cf\u5316\u4e0d\u4f1a\u6bd4 FP16 \u77e9\u9635\u91cf\u5316\u5feb\uff0c\u6027\u80fd\u4f18\u52bf\u6781\u5c0f\u3002\u5927\u6982\u5728 batch\u7b49\u4e8e 64 \u7684\u65f6\u5019\uff0cINT4 \u7684 Weight Only \u91cf\u5316\u53ea\u4f1a\u6bd4 FP16 \u7684\u5feb 30%\uff0c\u7b49\u5230 128 \u7684\u65f6\u5019\uff0c\u5927\u7ea6\u53ea\u4f1a\u5feb 20% \u751a\u81f3\u66f4\u5c0f\u3002<\/p>\n<p>\u4f46\u5bf9\u4e8e INT8 \u6765\u8bf4\uff0cINT8 \u7684\u91cf\u5316\u4e0e INT4 \u91cf\u5316\u6700\u4e0d\u540c\u7684\u4e00\u70b9\uff0c\u662f\u5b83\u4e0d\u9700\u8981\u4efb\u4f55\u89e3\u91cf\u5316\u7684\u8fc7\u7a0b\uff0c\u5e76\u4e14\u5b83\u7684\u8ba1\u7b97\u662f\u53ef\u4ee5\u538b\u7f29\u4e00\u500d\u65f6\u95f4\u7684\u3002\u5728 batch \u7b49\u4e8e 128 \u65f6\uff0c\u4ece FP16\u91cf\u5316\u5230 INT8\uff0c\u52a0\u8f7d\u6743\u91cd\u7684\u65f6\u95f4\u5c06\u4f1a\u51cf\u534a\uff0c\u8ba1\u7b97\u7684\u65f6\u95f4\u4e5f\u4f1a\u51cf\u534a\uff0c\u8fd9\u4f1a\u5e26\u6765\u767e\u5206\u4e4b\u767e\u7684\u52a0\u901f\u3002<\/p>\n<p>\u5728\u670d\u52a1\u7aef\u573a\u666f\u4e0b\uff0c\u7279\u522b\u662f\u56e0\u4e3a\u4f1a\u6709\u4e0d\u65ad\u7684\u8bf7\u6c42\u6d8c\u5165\uff0c\u5927\u90e8\u5206\u7684\u77e9\u9635\u4e58\uff0c\u90fd\u4f1a\u662f\u8ba1\u7b97\u5bc6\u96c6\u578b\u7684\u3002\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u5982\u679c\u4e3a\u4e86\u8ffd\u6c42\u6781\u9650\u7684\u541e\u5410\u91cf\uff0cINT8 \u7684\u6548\u7387\u5176\u5b9e\u662f\u9ad8\u4e8e INT4 \u7684\u3002\u8fd9\u4e5f\u662f\u4e3a\u4ec0\u4e48\u6211\u4eec\u76ee\u524d\u5df2\u7ecf\u5b8c\u6210\u7684\u5b9e\u73b0\u91cc\u9762\uff0c\u5728\u670d\u52a1\u7aef\u4e0a\u4e3b\u63a8 INT8 \u7684\u4e00\u4e2a\u539f\u56e0\u3002<\/p>\n<h3>9.\u00a0\u4f18\u5316\uff1aFP8 vx INT8<\/h3>\n<p>&nbsp;<\/p>\n<p>\u5728 H100\u3001H800\u30014090 \u4e0a\u9762\uff0c\u6211\u4eec\u53ef\u80fd\u4f1a\u6267\u884c FP8 \u7684\u91cf\u5316\u3002FP8 \u8fd9\u6837\u7684\u6570\u636e\u683c\u5f0f\uff0c\u5728 Nvidia \u6700\u65b0\u4e00\u4ee3\u7684\u663e\u5361\u5f53\u4e2d\u88ab\u5f15\u5165\u3002INT8 \u7684\u7cbe\u5ea6\u4ece\u7406\u8bba\u4e0a\u662f\u8981\u9ad8\u4e8e FP8 \u7684\uff0c\u4f46\u662f FP8 \u4f1a\u66f4\u597d\u7528\uff0c\u6027\u80fd\u4f1a\u66f4\u9ad8\u4e00\u4e9b\u3002\u6211\u4eec\u5728\u540e\u7eed\u670d\u52a1\u7aef\u7684\u63a8\u7406\u8fc7\u7a0b\u7684\u66f4\u65b0\u5f53\u4e2d\u4e5f\u4f1a\u63a8\u8fdb FP8 \u7684\u843d\u5730\u3002\u4e0a\u56fe\u4e2d\u53ef\u4ee5\u770b\u5230\uff0cFP8 \u7684\u8bef\u5dee\u76f8\u6bd4 INT8 \u4f1a\u5927 10 \u500d\u5de6\u53f3\u3002INT8 \u4f1a\u6709\u4e00\u4e2a\u91cf\u5316\u7684\u5c3a\u5bf8\u56e0\u5b50\uff0c\u53ef\u4ee5\u901a\u8fc7\u8c03\u6574\u5c3a\u5bf8\u56e0\u5b50\uff0c\u964d\u4f4e INT8 \u7684\u91cf\u5316\u8bef\u5dee\u3002\u800c FP8 \u7684\u91cf\u5316\u8bef\u5dee\u8ddf\u5c3a\u5bf8\u56e0\u5b50\u57fa\u672c\u4e0a\u662f\u65e0\u5173\u7684\uff0c\u5b83\u4e0d\u53d7\u5c3a\u5bf8\u56e0\u5b50\u7684\u5f71\u54cd\uff0c\u8fd9\u4f7f\u5f97\u6211\u4eec\u57fa\u672c\u4e0a\u4e0d\u9700\u8981\u5bf9\u5b83\u505a\u4efb\u4f55\u7684\u00a0 calibration\u3002\u4f46\u662f\u5b83\u7684\u8bef\u5dee\u603b\u4f53\u6765\u8bb2\u662f\u8981\u9ad8\u4e8e INT8 \u7684\u3002<\/p>\n<h3>10.\u00a0\u4f18\u5316\uff1aINT4 vs \u975e\u7ebf\u6027\u91cf\u5316<\/h3>\n<p>&nbsp;<\/p>\n<p>PPL.LLM \u5728\u540e\u7eed\u7684\u66f4\u65b0\u4e2d\uff0c\u4e5f\u4f1a\u66f4\u65b0 INT4 \u7684\u77e9\u9635\u91cf\u5316\u3002\u8fd9\u79cd Weight Only \u7684\u77e9\u9635\u91cf\u5316\u4e3b\u8981\u662f\u4e3a\u7aef\u4fa7\u670d\u52a1\u7684\uff0c\u4e3a\u4e86\u624b\u673a\u7aef\u79fb\u52a8\u7aef\u7b49 batch \u56fa\u5b9a\u4e3a 1 \u7684\u8bbe\u5907\u3002\u5728\u540e\u7eed\u7684\u66f4\u65b0\u5f53\u4e2d\u4f1a\u4ece INT4 \u9010\u6e10\u8f6c\u53d8\u4e3a\u975e\u7ebf\u6027\u91cf\u5316\u3002\u56e0\u4e3a\u5728 Weight Only \u7684\u8ba1\u7b97\u8fc7\u7a0b\u5f53\u4e2d\uff0c\u4f1a\u5b58\u5728\u4e00\u4e2a\u89e3\u91cf\u5316\u7684\u8fc7\u7a0b\uff0c\u8fd9\u4e2a\u89e3\u91cf\u5316\u8fc7\u7a0b\u5b9e\u9645\u662f\u53ef\u5b9a\u5236\u7684\uff0c\u672a\u5fc5\u662f\u4e00\u4e2a\u7ebf\u6027\u7684\u89e3\u91cf\u5316\u8fc7\u7a0b\uff0c\u5176\u4f7f\u7528\u5176\u5b83\u89e3\u91cf\u5316\u8fc7\u7a0b\u4ee5\u53ca\u91cf\u5316\u8fc7\u7a0b\uff0c\u4f1a\u4f7f\u5f97\u8fd9\u4e00\u6b21\u8ba1\u7b97\u7684\u7cbe\u5ea6\u66f4\u9ad8\u3002<\/p>\n<p>\u4e00\u4e2a\u6bd4\u8f83\u5178\u578b\u7684\u4f8b\u5b50\uff0c\u5c31\u662f\u5728\u4e00\u7bc7\u8bba\u6587\u5f53\u4e2d\u6240\u63d0\u5230\u7684 NF4 \u7684\u91cf\u5316\uff0c\u8fd9\u79cd\u91cf\u5316\u5b9e\u9645\u4e0a\u4f1a\u901a\u8fc7\u4e00\u79cd\u6253\u8868\u7684\u65b9\u5f0f\u8fdb\u884c\u91cf\u5316\u53ca\u89e3\u91cf\u5316\uff0c\u662f\u4e00\u79cd\u975e\u7ebf\u6027\u7684\u91cf\u5316\u3002PPL.LLM \u7684\u540e\u7eed\u66f4\u65b0\u5f53\u4e2d\u4f1a\u5c1d\u8bd5\u4f7f\u7528\u8fd9\u6837\u7684\u91cf\u5316\u6765\u5b8c\u6210\u7aef\u4fa7\u63a8\u7406\u7684\u4f18\u5316\u3002<\/p>\n<h2>\u4e09\u3001\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u7684\u786c\u4ef6<\/h2>\n<p>\u6700\u540e\uff0c\u4ecb\u7ecd\u4e00\u4e0b\u5927\u8bed\u8a00\u6a21\u578b\u5904\u7406\u7684\u786c\u4ef6\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>\u6a21\u578b\u7ed3\u6784\u4e00\u65e6\u786e\u5b9a\uff0c\u6211\u4eec\u5c31\u4f1a\u77e5\u9053\u5b83\u5177\u4f53\u7684\u8ba1\u7b97\u91cf\uff0c\u5177\u4f53\u9700\u8981\u591a\u5c11\u8bbf\u5b58\uff0c\u9700\u8981\u591a\u5c11\u8ba1\u7b97\u91cf\u3002\u540c\u65f6\u8fd8\u4f1a\u77e5\u9053\u6bcf\u5f20\u663e\u5361\u7684\u5e26\u5bbd\u3001\u7b97\u529b\u3001\u4ef7\u683c\u7b49\u3002\u5728\u786e\u5b9a\u4e86\u6a21\u578b\u7684\u7ed3\u6784\u4ee5\u53ca\u786e\u5b9a\u4e86\u786c\u4ef6\u6307\u6807\u4e4b\u540e\uff0c\u6211\u4eec\u5c31\u53ef\u4ee5\u901a\u8fc7\u8fd9\u4e9b\u6307\u6807\u53bb\u8ba1\u7b97\u51fa\u5728\u8fd9\u5f20\u663e\u5361\u4e0a\u63a8\u7406\u5927\u6a21\u578b\u7684\u6700\u5927\u541e\u5410\u91cf\u4f1a\u662f\u591a\u5c11\u3001\u8ba1\u7b97\u5ef6\u8fdf\u662f\u591a\u5c11\u3001\u8bbf\u5b58\u8bbf\u95ee\u65f6\u95f4\u9700\u8981\u591a\u5c11\uff0c\u53ef\u4ee5\u7b97\u51fa\u4e00\u4e2a\u975e\u5e38\u5177\u4f53\u7684\u8868\u3002\u6211\u4eec\u628a\u8fd9\u4e2a\u8868\u683c\u516c\u5f00\u5728\u540e\u7eed\u7684\u8d44\u6599\u5f53\u4e2d\uff0c\u5927\u5bb6\u53ef\u4ee5\u8bbf\u95ee\u8fd9\u4e2a\u8868\u683c\uff0c\u67e5\u770b\u6700\u9002\u5408\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u7684\u663e\u5361\u578b\u53f7\u6709\u54ea\u4e9b\u3002<\/p>\n<p>\u5bf9\u4e8e\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u6765\u8bf4\uff0c\u56e0\u4e3a\u5927\u90e8\u5206\u7b97\u5b50\u90fd\u662f\u8bbf\u5b58\u5bc6\u96c6\u578b\u7684\uff0c\u8bbf\u5b58\u7684\u5ef6\u8fdf\u603b\u4f1a\u6bd4\u8ba1\u7b97\u5ef6\u8fdf\u8981\u9ad8\u3002\u56e0\u4e3a\u5927\u8bed\u8a00\u6a21\u578b\u7684\u53c2\u6570\u77e9\u9635\u786e\u5b9e\u592a\u5927\u4e86\uff0c\u6240\u4ee5\u54ea\u6015\u662f\u5728 A100\/80G \u4e0a\uff0cbatch size \u5f00\u5230 272 \u7684\u65f6\u5019\uff0c\u5b83\u7684\u8ba1\u7b97\u5ef6\u8fdf\u90fd\u662f\u8f83\u5c0f\u7684\uff0c\u8bbf\u5b58\u5ef6\u8fdf\u53cd\u800c\u4f1a\u66f4\u9ad8\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u7684\u8bb8\u591a\u4f18\u5316\u90fd\u662f\u4ece\u8bbf\u5b58\u4e0a\u7740\u624b\u7684\u3002\u800c\u8fdb\u884c\u786c\u4ef6\u9009\u62e9\u65f6\uff0c\u6211\u4eec\u4e3b\u8981\u7684\u65b9\u5411\u5c31\u662f\u9009\u62e9\u5e26\u5bbd\u6bd4\u8f83\u9ad8\u3001\u663e\u5b58\u6bd4\u8f83\u5927\u7684\u8bbe\u5907\u3002\u4ece\u800c\u4f7f\u5f97\u5927\u8bed\u8a00\u6a21\u578b\u5728\u63a8\u7406\u65f6\uff0c\u53ef\u4ee5\u652f\u6491\u66f4\u591a\u7684\u8bf7\u6c42\uff0c\u652f\u6491\u66f4\u5feb\u7684\u8bbf\u5b58\uff0c\u76f8\u5e94\u7684\u541e\u5410\u91cf\u4e5f\u4f1a\u66f4\u9ad8\u3002<\/p>\n<p>\u4ee5\u4e0a\u5c31\u662f\u672c\u6b21\u5206\u4eab\u7684\u5185\u5bb9\u3002\u6240\u6709\u76f8\u5173\u8d44\u6599\u90fd\u653e\u5728\u4e86\u7f51\u76d8\u4e2d\uff0c\u94fe\u63a5\u53c2\u89c1\u4e0a\u56fe\u3002\u6211\u4eec\u6240\u6709\u7684\u4ee3\u7801\u4e5f\u5df2\u7ecf\u5f00\u6e90\u5728\u4e86 github \u4e0a\u3002\u6b22\u8fce\u5927\u5bb6\u968f\u65f6\u4e0e\u6211\u4eec\u8fdb\u884c\u6c9f\u901a\u3002<\/p>\n<h2>\u56db\u3001Q &amp; A<\/h2>\n<h4>Q1\uff1aPPL.LLM \u4e2d\u6709\u6ca1\u6709\u4f18\u5316\u50cf\u00a0Flash Attention\u00a0\u4e2d\u7684\u00a0Softmax \u8fd9\u79cd\u8bbf\u5b58\u7684\u95ee\u9898\uff1f<\/h4>\n<p>A1\uff1aDecoding Attention \u8fd9\u4e2a\u7b97\u5b50\u975e\u5e38\u7279\u6b8a\uff0c\u5b83\u7684 Q \u7684\u957f\u5ea6\u6c38\u8fdc\u662f 1\uff0c\u6240\u4ee5\u5b83\u4e0d\u4f1a\u50cfFlash Attention \u90a3\u6837\u9762\u4e34 Softmax \u91cc\u6709\u975e\u5e38\u5927\u7684\u8bbf\u5b58\u91cf\u3002\u5b9e\u9645\u4e0a\u5728 Decoding Attention \u7684\u6267\u884c\u8fc7\u7a0b\u5f53\u4e2d\uff0c\u5c31\u662f\u5b8c\u6574\u5730\u6267\u884c\u8fd9\u6b21 Softmax \u7684\u8fc7\u7a0b\uff0c\u5e76\u4e0d\u9700\u8981\u50cf Flash Attention \u90a3\u6837\u66f4\u5feb\u6267\u884c\u3002<\/p>\n<h4>Q2\uff1aINT4 \u7684 Weight Only \u91cf\u5316\u4e3a\u4ec0\u4e48\u548c batch \u7ebf\u6027\u76f8\u5173\uff0c\u8bf7\u95ee\u8fd9\u662f\u56fa\u5b9a\u6570\u91cf\u5417\uff1f<\/h4>\n<p>A2\uff1a\u8fd9\u662f\u4e00\u4e2a\u597d\u95ee\u9898\uff0c\u9996\u5148\u8fd9\u4e2a\u89e3\u91cf\u5316\u4e0d\u662f\u50cf\u5927\u5bb6\u60f3\u7684\u90a3\u6837\uff0c\u53ea\u9700\u8981\u628a\u6743\u91cd\u4ece INT4 \u585e\u56de FP16 \u5c31\u884c\u4e86\uff0c\u5982\u679c\u53ea\u505a\u8fd9\u4ef6\u4e8b\u60c5\uff0c\u90a3\u6743\u91cd\u6709\u591a\u5c11\u5c31\u8981\u89e3\u591a\u5c11\u3002\u5b9e\u9645\u4e0a\u4e0d\u662f\u8fd9\u6837\u7684\uff0c\u56e0\u4e3a\u8fd9\u662f\u4e00\u4e2a\u878d\u5408\u5728\u77e9\u9635\u4e58\u6cd5\u91cc\u9762\u7684\u89e3\u91cf\u5316\uff0c\u4e0d\u80fd\u5728\u6267\u884c\u77e9\u9635\u4e58\u6cd5\u4e4b\u524d\uff0c\u628a\u6240\u6709\u6743\u91cd\u89e3\u91cf\u5316\u51fa\u6765\uff0c\u653e\u5728\u90a3\u7136\u540e\u518d\u53bb\u8bfb\u3002\u8fd9\u6837\u6211\u4eec\u6240\u505a\u7684 INT4 \u7684\u91cf\u5316\u5c31\u6ca1\u6709\u610f\u4e49\u4e86\u3002\u5b83\u662f\u5728\u6267\u884c\u8fc7\u7a0b\u5f53\u4e2d\u4e0d\u505c\u5730\u53bb\u89e3\u91cf\u5316\uff0c\u56e0\u4e3a\u6211\u4eec\u4f1a\u6267\u884c\u5206\u5757\u7684\u77e9\u9635\u4e58\uff0c\u6bcf\u4e00\u4e2a\u6743\u91cd\u6240\u8981\u8bfb\u5199\u7684\u6b21\u6570\u5e76\u4e0d\u662f 1\uff0c\u9700\u8981\u4e0d\u505c\u5730\u62ff\u8fc7\u6765\u8ba1\u7b97\uff0c\u8fd9\u4e2a\u6b21\u6570\u5b9e\u9645\u4e0a\u8ddf batch \u6709\u5173\u3002\u4e5f\u5c31\u662f\u533a\u522b\u4e8e\u4e4b\u524d\u90a3\u4e9b\u4f18\u5316\u91cf\u5316\u7684\u624b\u6bb5\uff0c\u4f1a\u6709\u5355\u72ec\u7684\u91cf\u5316\u7684\u7b97\u5b50\u548c\u89e3\u91cf\u5316\u7b97\u5b50\u3002\u4e24\u4e2a\u7b97\u5b50\u7684\u63d2\u5165\uff0c\u89e3\u91cf\u5316\u8fd8\u662f\u76f4\u63a5\u878d\u5408\u5728\u7b97\u5b50\u4e2d\u7684\u3002\u6211\u4eec\u6267\u884c\u7684\u662f\u77e9\u9635\u4e58\u6cd5\uff0c\u6240\u4ee5\u6240\u8981\u89e3\u91cf\u5316\u7684\u6b21\u6570\u5e76\u4e0d\u662f\u4e00\u6b21\u3002<\/p>\n<h4>Q3\uff1aKV Cache \u4e2d\u7684\u53cd\u91cf\u5316\u8ba1\u7b97\uff0c\u53ef\u4ee5\u88ab\u4eff\u5b58\u63a9\u76d6\uff1f<\/h4>\n<p>A3\uff1a\u6839\u636e\u6211\u4eec\u7684\u6d4b\u8bd5\u662f\u53ef\u4ee5\u88ab\u63a9\u76d6\u7684\uff0c\u800c\u4e14\u5176\u5b9e\u8fd8\u8fdc\u8fdc\u6709\u5269\u4f59\u3002KV \u8ba1\u7b97\u4e2d\u7684\u53cd\u91cf\u5316\u4ee5\u53ca\u91cf\u5316\u90fd\u4f1a\u88ab\u878d\u5408\u8fdb self attention \u7b97\u5b50\u5f53\u4e2d\uff0c\u5177\u4f53\u6765\u8bf4\u5c31\u662f Decoding Attention\u3002\u6839\u636e\u6d4b\u8bd5\uff0c\u8fd9\u4e2a\u7b97\u5b50\u5373\u4f7f\u5728 10 \u500d\u7684\u8ba1\u7b97\u91cf\uff0c\u53ef\u80fd\u90fd\u53ef\u4ee5\u63a9\u76d6\u6389\u3002\u5c31\u662f\u8bbf\u5b58\u7684\u5ef6\u8fdf\u90fd\u63a9\u76d6\u4e0d\u4e86\u5b83\uff0c\u5b83\u4e3b\u8981\u7684\u74f6\u9888\u5728\u4e8e\u8bbf\u5b58\uff0c\u5b83\u8ba1\u7b97\u91cf\u8fd8\u8fdc\u8fdc\u8fbe\u4e0d\u5230\u53ef\u4ee5\u63a9\u76d6\u6389\u5b83\u8bbf\u5b58\u7684\u90a3\u4e2a\u7a0b\u5ea6\u3002\u6240\u4ee5 KV cache \u5f53\u4e2d\u7684\u53cd\u91cf\u5316\u8ba1\u7b97\uff0c\u5bf9\u4e8e\u8fd9\u4e2a\u7b97\u5b50\u6765\u8bf4\uff0c\u57fa\u672c\u4e0a\u662f\u4e00\u4e2a\u5f88\u597d\u88ab\u63a9\u76d6\u7684\u4e1c\u897f\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u4e00\u3001\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u6982\u8981\u4ecb\u7ecd &nbsp; \u4e0e\u4f20\u7edf\u7684 CNN \u6a21\u578b\u63a8\u7406\u4e0d\u540c\uff0c\u5927\u8bed\u8a00\u6a21\u578b\u7684\u63a8\u7406\u901a\u5e38\u4f1a\u5206\u6210 pref [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[7023],"tags":[],"class_list":["post-415415","post","type-post","status-publish","format-standard","hentry","category-chatgpt"],"_links":{"self":[{"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/posts\/415415","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/comments?post=415415"}],"version-history":[{"count":0,"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/posts\/415415\/revisions"}],"wp:attachment":[{"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/media?parent=415415"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/categories?post=415415"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/tags?post=415415"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}