\u63a8\u7406\u6548\u7387\u6307\u6807<\/strong><\/h2>\n

\u5ef6\u8fdf\u4e0e\u541e\u5410\u91cf<\/h3>\n

\u4f60\u9700\u8981\u8861\u91cf\u5ef6\u8fdf\u548c\u541e\u5410\u91cf\uff0c\u6765\u4e86\u89e3 AI \u5de5\u4f5c\u8d1f\u8f7d\u5728\u7f8e\u56fd\u670d\u52a1\u5668\u4e0a\u7684\u8868\u73b0\u3002\u5ef6\u8fdf\u8868\u793a\u6a21\u578b\u5bf9\u7528\u6237\u8bf7\u6c42\u505a\u51fa\u54cd\u5e94\u6240\u9700\u7684\u65f6\u95f4\u3002\u5c24\u5176\u5728\u5b9e\u65f6\u5e94\u7528\u4e2d\uff0c\u5982\u679c\u4f60\u6709\u4e25\u683c\u7684\u5ef6\u8fdf\u8981\u6c42\uff0c\u5c31\u5e94\u540c\u65f6\u5173\u6ce8\u7aef\u5230\u7aef\u5ef6\u8fdf\u548c\u6a21\u578b\u5ef6\u8fdf\u3002\u5728\u505a\u4efb\u4f55\u66f4\u6539\u4e4b\u524d\uff0c\u5148\u5efa\u7acb\u63a8\u7406\u5ef6\u8fdf\u7684\u57fa\u7ebf\u3002\u4f60\u53ef\u4ee5\u901a\u8fc7\u4f7f\u7528\u66f4\u5c0f\u7684\u6a21\u578b\u6216\u91cf\u5316\u6a21\u578b\u6765\u6539\u5584\u6a21\u578b\u5ef6\u8fdf\u3002\u91cd\u70b9\u4f18\u5316\u9884\u586b\u5145\uff08prefill\uff09\u9636\u6bb5\uff0c\u5e76\u4f7f\u7528\u9ad8\u6548\u8def\u7531\uff0c\u4ee5\u964d\u4f4e\u9996 token \u65f6\u95f4\uff08time to first token\uff09\u3002\u541e\u5410\u91cf\u8861\u91cf\u7cfb\u7edf\u6bcf\u79d2\u53ef\u4ee5\u5904\u7406\u591a\u5c11 token \u6216\u8bf7\u6c42\u3002token \u541e\u5410\u91cf\u8868\u793a\u6bcf\u5757 GPU \u6bcf\u79d2\u751f\u6210\u7684 token \u6570\uff0c\u6709\u52a9\u4e8e\u4f60\u4e86\u89e3 AI \u7cfb\u7edf\u7684\u6574\u4f53\u541e\u5410\u80fd\u529b\u3002<\/p>\n

\u63d0\u793a\uff1a\u901a\u8fc7\u8c03\u6574\u6279\u5927\u5c0f\uff0c\u5728\u4e0d\u663e\u8457\u589e\u52a0\u5ef6\u8fdf\u7684\u524d\u63d0\u4e0b\u6700\u5927\u5316\u541e\u5410\u91cf\u3002<\/p><\/blockquote>\n

\u8d44\u6e90\u5229\u7528\u7387<\/h3>\n

\u8d44\u6e90\u5229\u7528\u7387\u53cd\u6620\u4e86 AI \u6a21\u578b\u5728\u786c\u4ef6\u548c\u80fd\u6e90\u4f7f\u7528\u4e0a\u7684\u6548\u7387\u3002\u4f60\u53ef\u4ee5\u4f7f\u7528\u57fa\u51c6\u6d4b\u8bd5\u6846\u67b6\u6765\u5ea6\u91cf\u6bcf\u4e2a\u8bf7\u6c42\u7684\u5ef6\u8fdf\u3001\u541e\u5410\u91cf\u548c GPU \u529f\u8017\u3002\u8fd9\u4e9b\u6846\u67b6\u8fd8\u4f1a\u8003\u8651\u7535\u6e90\u4f7f\u7528\u6548\u7387\uff08PUE\uff09\u3001\u6c34\u8d44\u6e90\u4f7f\u7528\u6548\u7387\uff08WUE\uff09\u4ee5\u53ca\u78b3\u5f3a\u5ea6\u56e0\u5b50\uff08CIF\uff09\u7b49\u6307\u6807\u3002\u5de5\u4f5c\u8d1f\u8f7d\u80fd\u529b\u901a\u5e38\u4ee5\u6bcf\u79d2\u6d6e\u70b9\u8fd0\u7b97\u6b21\u6570\uff08FLOPS\uff09\u8861\u91cf\uff0c\u7528\u6765\u8868\u660e\u670d\u52a1\u5668\u7684\u8ba1\u7b97\u80fd\u529b\u3002AI \u670d\u52a1\u5668\u53ef\u4ee5\u8fbe\u5230\u4e07\u4ebf\u6b21 FLOPS\uff0c\u4f46\u5b9e\u9645\u8868\u73b0\u4ecd\u53d6\u51b3\u4e8e\u786c\u4ef6\u4f18\u5316\u548c\u57fa\u7840\u8bbe\u65bd\u72b6\u51b5\u3002\u4e0b\u8868\u5c55\u793a\u4e86\u4e0d\u540c\u56e0\u7d20\u5bf9\u63a8\u7406\u6548\u7387\u7684\u5f71\u54cd\uff1a<\/p>\n

\n\n\n\n<\/colgroup>\n\n\n\n\n\n\n

\u56e0\u7d20<\/th>\n	\u5bf9\u63a8\u7406\u6548\u7387\u7684\u5f71\u54cd<\/th>\n<\/tr>\n
\u57fa\u7840\u8bbe\u65bd<\/td>\n	\u662f AI \u63a8\u7406\u53ef\u6301\u7eed\u6027\u7684\u5173\u952e\u51b3\u5b9a\u56e0\u7d20\u3002<\/td>\n<\/tr>\n
\u786c\u4ef6\u4f18\u5316<\/td>\n	\u5b9e\u9645\u6548\u679c\u4f1a\u56e0\u90e8\u7f72\u6761\u4ef6\u548c\u786c\u4ef6\u6548\u7387\u800c\u6709\u5dee\u5f02\u3002<\/td>\n<\/tr>\n
\u80fd\u6e90\u4f7f\u7528<\/td>\n	\u6a21\u578b\u8bbe\u8ba1\u63d0\u5347\u7406\u8bba\u6548\u7387\uff0c\u4f46\u771f\u5b9e\u80fd\u8017\u4f1a\u6709\u6240\u4e0d\u540c\u3002<\/td>\n<\/tr>\n
\u6570\u636e\u4e2d\u5fc3\u4f4e\u6548<\/td>\n	\u9ad8\u6c34\u8db3\u8ff9\u5f80\u5f80\u6e90\u4e8e\u6570\u636e\u4e2d\u5fc3\u4f4e\u6548\uff0c\u800c\u4e0d\u4ec5\u662f\u6a21\u578b\u672c\u8eab\u7279\u6027\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n \u51c6\u786e\u6027\u4e0e\u4e00\u81f4\u6027<\/h3>\n \u4f60\u8fd8\u5fc5\u987b\u68c0\u67e5\u51c6\u786e\u6027\u548c\u4e00\u81f4\u6027\uff0c\u4ee5\u786e\u4fdd AI \u63a8\u7406\u53ef\u9760\u3002\u7531 MLCommons \u521b\u5efa\u7684 MLPerf \u57fa\u51c6\uff0c\u4e3a\u786c\u4ef6\u3001\u8f6f\u4ef6\u548c\u670d\u52a1\u7684\u8bad\u7ec3\u4e0e\u63a8\u7406\u6027\u80fd\u63d0\u4f9b\u4e86\u5ba2\u89c2\u8bc4\u4f30\u3002\u8fd9\u4e9b\u57fa\u51c6\u5728\u4e25\u683c\u6761\u4ef6\u4e0b\uff0c\u6d4b\u8bd5\u5305\u62ec\u6df1\u5ea6\u5b66\u4e60\u4e0e\u673a\u5668\u5b66\u4e60\u5728\u5185\u7684\u591a\u79cd\u6a21\u578b\u67b6\u6784\u3002\u4e00\u81f4\u6027\u610f\u5473\u7740\u6a21\u578b\u5728\u591a\u6b21\u8fd0\u884c\u4e2d\u4f1a\u7ed9\u51fa\u76f8\u540c\u6216\u9ad8\u5ea6\u76f8\u4f3c\u7684\u7ed3\u679c\u3002\u4e0b\u56fe\uff08\u5047\u5b9a\u5b58\u5728\uff09\u5bf9\u6bd4\u4e86\u4e0d\u540c\u540e\u7aef\u548c\u8fd0\u884c\u6a21\u5f0f\u4e0b\u63a8\u7406\u7ed3\u679c\u4e0d\u4e00\u81f4\u7684\u60c5\u51b5\uff1a<\/p>\n \u6d4b\u8bd5\u73af\u5883\u642d\u5efa<\/strong><\/h2>\n \u786c\u4ef6\u4e0e\u7f51\u7edc<\/h3>\n \u8981\u5728\u7f8e\u56fd\u670d\u52a1\u5668\u4e0a\u83b7\u5f97\u9ad8\u6548\u7684 AI \u63a8\u7406\uff0c\u4f60\u9700\u8981\u5408\u7406\u7684\u786c\u4ef6\u4e0e\u7f51\u7edc\u914d\u7f6e\u3002\u5bf9\u4e8e\u5927\u591a\u6570\u751f\u4ea7\u90e8\u7f72\u800c\u8a00\uff0c\u4e00\u53f0\u914d\u5907 4 \u5757 GPU \u7684 2U \u670d\u52a1\u5668\u901a\u5e38\u8db3\u591f\u3002\u5982\u679c\u4f60\u8ffd\u6c42\u6700\u5927\u541e\u5410\u91cf\u548c\u66f4\u4f73\u6563\u70ed\u8868\u73b0\uff0c\u5219 4U \u670d\u52a1\u5668\u662f\u66f4\u597d\u7684\u9009\u62e9\u3002\u914d\u5907 96 GB \u663e\u5b58\u7684 RTX PRO 6000 Blackwell GPU \u9002\u5408\u5c0f\u56e2\u961f\uff0c\u800c\u66f4\u5927\u89c4\u6a21\u90e8\u7f72\u5219\u53ef\u91c7\u7528 4 \u6216 8 GPU \u7684\u914d\u7f6e\u3002AMD EPYC 9005 \u5904\u7406\u5668\u62e5\u6709\u9ad8\u6570\u91cf\u7684 PCIe Gen 5 \u901a\u9053\u548c\u5f3a\u5927\u7684\u5185\u5b58\u5e26\u5bbd\uff0c\u975e\u5e38\u9002\u5408\u591a GPU \u63a8\u7406\u573a\u666f\u3002AI \u5e94\u7528\u5bf9\u7f51\u7edc\u6027\u80fd\u8981\u6c42\u4e5f\u5f88\u9ad8\u3002\u5bf9\u4e8e\u5b9e\u65f6\u89c6\u9891\u5206\u6790\u548c\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u8bad\u7ec3\u800c\u8a00\uff0c\u9ad8\u5e26\u5bbd\u81f3\u5173\u91cd\u8981\u3002\u5982\u679c\u5e26\u5bbd\u4e0d\u8db3\uff0c\u4f60\u53ef\u80fd\u4f1a\u9047\u5230\u8bad\u7ec3\u65f6\u95f4\u5ef6\u957f\u6216\u5b9e\u65f6\u6d41\u7a0b\u5931\u8d25\u7684\u95ee\u9898\u3002\u5927\u578b AI \u6a21\u578b\u540c\u6837\u9700\u8981\u53ef\u89c2\u7684\u5185\u5b58\u5e26\u5bbd\uff0c\u4ee5\u5b9e\u73b0\u5feb\u901f\u6570\u636e\u4f20\u8f93\u3002\u5f53\u4f60\u589e\u52a0 GPU \u6570\u91cf\u65f6\uff0c\u7cfb\u7edf\u5ef6\u8fdf\u6709\u673a\u4f1a\u964d\u4f4e\uff0c\u4f46\u524d\u63d0\u662f\u7f51\u7edc\u5e26\u5bbd\u8db3\u591f\u652f\u6491\u3002<\/p>\n \u63d0\u793a\uff1a\u5728\u6269\u5c55 AI \u5de5\u4f5c\u8d1f\u8f7d\u4e4b\u524d\uff0c\u52a1\u5fc5\u5148\u68c0\u67e5\u5e76\u786e\u8ba4\u4f60\u7684\u7f51\u7edc\u5e26\u5bbd\u662f\u5426\u5145\u8db3\u3002<\/p><\/blockquote>\n \u8f6f\u4ef6\u3001\u6846\u67b6\u4e0e Amazon SageMaker<\/h3>\n \u5728 AWS \u4e0a\u8fdb\u884c\u53ef\u6269\u5c55\u7684 AI \u63a8\u7406\uff0c\u5efa\u8bae\u4f7f\u7528 SageMaker\u3002SageMaker \u4e0e NVIDIA Triton Inference Server \u6df1\u5ea6\u96c6\u6210\uff0c\u53ef\u5728\u6a21\u578b\u670d\u52a1\u4e2d\u6700\u5927\u5316\u541e\u5410\u91cf\u3001\u6700\u5c0f\u5316\u5ef6\u8fdf\u3002SageMaker \u652f\u6301\u5305\u62ec TensorRT\u3001TensorFlow\u3001PyTorch \u548c ONNX \u5728\u5185\u7684\u4e3b\u6d41 AI \u6846\u67b6\u3002SageMaker Inference Recommender \u80fd\u5e2e\u52a9\u4f60\u9009\u62e9\u5408\u9002\u7684\u8ba1\u7b97\u5b9e\u4f8b\u7c7b\u578b\u3001\u5b9e\u4f8b\u6570\u91cf\u4ee5\u53ca\u63a8\u7406\u4f18\u5316\u65b9\u5f0f\uff0c\u4ece\u800c\u514d\u53bb\u5927\u91cf\u7e41\u7410\u5de5\u4f5c\uff0c\u8ba9\u4f60\u53ef\u4ee5\u66f4\u5feb\u5730\u8fdb\u884c\u5b9e\u9a8c\u548c\u4f18\u5316\u3002\u591a\u7ebf\u7a0b\u53ef\u4ee5\u663e\u8457\u63d0\u5347\u541e\u5410\u91cf\uff0c\u5c24\u5176\u662f\u5728 GPU \u4e0a\u90e8\u7f72 Transformer \u6a21\u578b\u65f6\u3002\u6279\u5904\u7406\uff08batching\uff09\u5219\u662f\u53e6\u4e00\u9879\u5173\u952e\u6280\u672f\u3002SageMaker \u548c NVIDIA Triton Inference Server \u90fd\u652f\u6301\u6279\u5904\u7406\uff0c\u901a\u8fc7\u6279\u5904\u7406\u53ef\u4ee5\u63d0\u9ad8\u541e\u5410\u91cf\u5e76\u964d\u4f4e\u5ef6\u8fdf\u3002MLPerf Inference \u4e3a\u4f60\u7684 AI \u5de5\u4f5c\u8d1f\u8f7d\u63d0\u4f9b\u6807\u51c6\u5316\u57fa\u51c6\u3002\u4f60\u53ef\u4ee5\u4f7f\u7528 SageMaker Inference Recommender \u6d4b\u8bd5\u4e0d\u540c\u914d\u7f6e\uff0c\u4e3a\u4f60\u7684\u5b9e\u65f6\u7aef\u70b9\u627e\u5230\u6700\u4f73\u65b9\u6848\u3002<\/p>\n \n\n\n\n<\/colgroup>\n\n\n\n\n\n\n \u7279\u6027<\/th>\n \u4f18\u52bf<\/th>\n<\/tr>\n SageMaker + NVIDIA Triton<\/td>\n \u9ad8\u541e\u5410\u91cf\u3001\u4f4e\u5ef6\u8fdf\u7684\u6a21\u578b\u670d\u52a1\u80fd\u529b<\/td>\n<\/tr>\n SageMaker Inference Recommender<\/td>\n \u81ea\u52a8\u9009\u62e9\u5b9e\u4f8b\u7c7b\u578b\u548c\u4f18\u5316\u7b56\u7565<\/td>\n<\/tr>\n \u591a\u7ebf\u7a0b\u4e0e\u6279\u5904\u7406<\/td>\n \u63d0\u5347\u541e\u5410\u91cf\u5e76\u964d\u4f4e\u5ef6\u8fdf<\/td>\n<\/tr>\n MLPerf Inference \u57fa\u51c6<\/td>\n \u6807\u51c6\u5316\u7684\u6027\u80fd\u6d4b\u91cf\u65b9\u5f0f<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n \u6570\u636e\u51c6\u5907<\/h3>\n \u4e3a\u4e86\u83b7\u5f97\u51c6\u786e\u7684\u63a8\u7406\u57fa\u51c6\u6d4b\u8bd5\u7ed3\u679c\uff0c\u4f60\u5fc5\u987b\u8ba4\u771f\u51c6\u5907\u6570\u636e\u3002\u7cfb\u7edf\u5316\u7684\u53c2\u6570\u9009\u62e9\uff0c\u6709\u52a9\u4e8e\u4f60\u66f4\u6709\u6548\u5730\u6d4b\u8bd5 AI \u6a21\u578b\u3002\u5bf9\u4e8e\u5b9e\u65f6\u670d\u52a1\uff0c\u53ef\u4ee5\u8bbe\u5b9a\u5ef6\u8fdf\u9608\u503c\uff0c\u4f8b\u5982 LLM \u4ea4\u4e92\u4e2d\u6bcf\u4e2a token 50 ms \u7684\u4e0a\u9650\u3002\u4f7f\u7528\u5408\u6210\u6570\u636e\u4e0e\u771f\u5b9e\u6570\u636e\u76f8\u7ed3\u5408\u7684\u65b9\u5f0f\uff0c\u8bc4\u4f30\u6a21\u578b\u5728\u4e0d\u540c\u6570\u636e\u7c7b\u578b\u4e0a\u7684\u8868\u73b0\u3002\u826f\u597d\u7684\u5b9e\u9a8c\u8bbe\u8ba1\u548c\u5408\u7406\u7684\u91c7\u6837\u7b56\u7565\uff0c\u80fd\u591f\u5e2e\u52a9\u4f60\u9ad8\u6548\u63a2\u7d22\u5927\u91cf\u53c2\u6570\u7ec4\u5408\u3002\u6570\u636e\u9884\u5904\u7406\u5728\u63a8\u7406\u6548\u7387\u4e2d\u4e5f\u626e\u6f14\u7740\u91cd\u8981\u89d2\u8272\u3002\u5982\u679c\u4f60\u5728 CPU \u4e0a\u8fdb\u884c\u9884\u5904\u7406\uff0c\u53ef\u80fd\u4f1a\u51fa\u73b0\u74f6\u9888\u3002\u50cf NVIDIA DALI \u8fd9\u6837\u7684\u5de5\u5177\u53ef\u4ee5\u5c06\u9884\u5904\u7406\u5de5\u4f5c\u5378\u8f7d\u5230 GPU \u4e0a\uff0c\u4ece\u800c\u964d\u4f4e\u5ef6\u8fdf\u3001\u63d0\u5347\u541e\u5410\u91cf\u3002\u5f53\u4f60\u5c06 DALI \u4e0e NVIDIA Triton Inference Server \u96c6\u6210\u65f6\uff0c\u53ef\u4ee5\u5728 GPU \u4e0a\u5b8c\u6210\u6570\u636e\u89e3\u7801\u548c\u7f29\u653e\uff0c\u5c3d\u91cf\u51cf\u5c11\u901a\u4fe1\u5f00\u9500\u3002SageMaker \u652f\u6301\u8fd9\u4e9b\u5de5\u4f5c\u6d41\uff0c\u5e2e\u52a9\u4f60\u66f4\u8f7b\u677e\u5730\u4f18\u5316 AI \u6a21\u578b\u90e8\u7f72\u3002SageMaker Inference Recommender \u8fd8\u53ef\u4ee5\u534f\u52a9\u4f60\u6d4b\u8bd5\u6279\u5904\u7406\u7b56\u7565\u548c\u9884\u5904\u7406\u6d41\u6c34\u7ebf\uff0c\u4ece\u800c\u4e3a\u7aef\u70b9\u627e\u5230\u6700\u4f73\u914d\u7f6e\u3002<\/p>\n \u6ce8\u610f\uff1a\u5728 AWS \u4e0a\u5b9e\u73b0\u53ef\u9760\u7684 AI \u63a8\u7406\u7ed3\u679c\uff0c\u79bb\u4e0d\u5f00\u5408\u7406\u7684\u6570\u636e\u51c6\u5907\u548c\u6279\u5904\u7406\u7b56\u7565\u3002<\/p><\/blockquote>\n \u8fd0\u884c\u63a8\u7406\u57fa\u51c6\u6d4b\u8bd5<\/strong><\/h2>\n MLPerf Inference \u4e0e\u5176\u4ed6\u5de5\u5177<\/h3>\n \u4f60\u53ef\u4ee5\u901a\u8fc7\u8fd0\u884c MLPerf Inference \u57fa\u51c6\u6d4b\u8bd5\uff0c\u6765\u8861\u91cf\u7f8e\u56fd\u670d\u52a1\u5668\u4e0a\u7684\u63a8\u7406\u6548\u7387\u3002MLPerf Inference \u662f\u9488\u5bf9 AI\u3001\u6df1\u5ea6\u5b66\u4e60\u548c\u673a\u5668\u5b66\u4e60\u5de5\u4f5c\u8d1f\u8f7d\u7684\u9886\u5148\u57fa\u51c6\u6d4b\u8bd5\u5957\u4ef6\uff0c\u5e2e\u52a9\u4f60\u5bf9\u6bd4\u4e0d\u540c\u786c\u4ef6\u4e0e\u8f6f\u4ef6\u7ec4\u5408\u7684\u6027\u80fd\u3002\u8981\u5f00\u59cb\u6d4b\u8bd5\uff0c\u4f60\u9700\u8981\u6309\u5982\u4e0b\u6b65\u9aa4\u51c6\u5907\u73af\u5883\u5e76\u786e\u4fdd\u7ed3\u679c\u51c6\u786e\uff1a<\/p>\n \n \u5b89\u88c5 MLCommons CM \u6846\u67b6\uff0c\u8be5\u6846\u67b6\u63d0\u4f9b\u7528\u4e8e AI \u57fa\u51c6\u6d4b\u8bd5\u7684\u81ea\u52a8\u5316\u914d\u65b9\u3002<\/li>\n \u4f7f\u7528\u56fe\u5f62\u754c\u9762\u751f\u6210 CM \u547d\u4ee4\uff0c\u4ee5\u81ea\u5b9a\u4e49\u5e76\u8fd0\u884c MLPerf Inference \u57fa\u51c6\u6d4b\u8bd5\u3002<\/li>\n \u901a\u8fc7 pip install tensorflow<\/code> \u548c pip install tensorflow-io<\/code> \u5b89\u88c5 TensorFlow \u4f5c\u4e3a\u540e\u7aef\u3002<\/li>\n \u8bbe\u7f6e\u73af\u5883\u53d8\u91cf\u4ee5\u542f\u7528 oneDNN\uff0c\u5e76\u914d\u7f6e 16 \u4f4d\u6d6e\u70b9\u5b58\u50a8\u683c\u5f0f\uff0c\u4ee5\u63d0\u5347\u6548\u7387\u3002<\/li>\n \u4f7f\u7528 wget -q https:\/\/zenodo.org\/record\/2535873\/files\/resnet50_v1.pb<\/code> \u4e0b\u8f7d\u4f60\u8981\u6d4b\u8bd5\u7684\u673a\u5668\u5b66\u4e60\u6a21\u578b\uff0c\u4f8b\u5982 resnet50-v1.5<\/code>\u3002<\/li>\n \u4e0b\u8f7d\u9002\u7528\u4e8e\u8be5\u6a21\u578b\u7684\u6570\u636e\u96c6\uff0c\u4f8b\u5982 imagenet2012 \u9a8c\u8bc1\u96c6\u3002<\/li>\n \u4f7f\u7528\u547d\u4ee4 .\/run_local.sh tf resnet50 cpu<\/code> \u5728\u670d\u52a1\u5668\u4e0a\u8fd0\u884c\u57fa\u51c6\u6d4b\u8bd5\u3002<\/li>\n<\/ol>\n \u4f60\u4e5f\u53ef\u4ee5\u4f7f\u7528\u5176\u4ed6\u5de5\u5177\uff0c\u4f8b\u5982\u652f\u6301 TensorRT \u540e\u7aef\u7684 NVIDIA Triton Inference Server\uff0c\u7528\u4e8e\u9ad8\u541e\u5410\u91cf AI \u5de5\u4f5c\u8d1f\u8f7d\u3002Triton \u5185\u7f6e\u6279\u5904\u7406\u3001\u6a21\u578b\u7248\u672c\u7ba1\u7406\u4ee5\u53ca\u591a\u6846\u67b6\u90e8\u7f72\u7b49\u529f\u80fd\uff0c\u6709\u52a9\u4e8e\u4f60\u4f18\u5316\u6a21\u578b\u5ef6\u8fdf\u548c\u6574\u4f53\u541e\u5410\u91cf\u3002<\/p>\n \u63d0\u793a\uff1a\u52a1\u5fc5\u4f7f\u7528\u6807\u51c6\u5316\u7684\u57fa\u51c6\u6d4b\u8bd5\u5de5\u5177\uff0c\u786e\u4fdd\u7ed3\u679c\u516c\u5e73\u4e14\u53ef\u91cd\u590d\u3002<\/p><\/blockquote>\n \u6267\u884c AI \u5de5\u4f5c\u8d1f\u8f7d<\/h3>\n \u5728\u4e3a\u63a8\u7406\u57fa\u51c6\u6d4b\u8bd5\u6267\u884c AI \u5de5\u4f5c\u8d1f\u8f7d\u65f6\uff0c\u4f60\u9700\u8981\u8bbe\u7f6e\u6d4b\u8bd5\u573a\u666f\uff0c\u4f7f\u4e4b\u5c3d\u53ef\u80fd\u8d34\u8fd1\u771f\u5b9e\u4f7f\u7528\u573a\u666f\u3002\u5e94\u540c\u65f6\u5173\u6ce8\u5b9e\u65f6\u63a8\u7406\u548c\u6279\u91cf\u5904\u7406\u3002\u4e0b\u8868\u5c55\u793a\u4e86\u5e38\u89c1\u6d4b\u8bd5\u7c7b\u578b\u3001\u5173\u952e\u53c2\u6570\u4ee5\u53ca\u9700\u8981\u8ddf\u8e2a\u7684\u6307\u6807\uff1a<\/p>\n \n\n\n\n\n<\/colgroup>\n\n\n\n\n \u6d4b\u8bd5\u7c7b\u578b<\/th>\n \u53c2\u6570<\/th>\n \u6307\u6807<\/th>\n<\/tr>\n \u8bad\u7ec3\u6d4b\u8bd5<\/td>\n \u6279\u5927\u5c0f\u3001\u6a21\u578b\u5927\u5c0f<\/td>\n \u541e\u5410\u91cf\uff08\u6837\u672c\/\u79d2\uff09\u3001GPU \u540d\u79f0<\/td>\n<\/tr>\n \u63a8\u7406\u6d4b\u8bd5<\/td>\n \u6a21\u578b\u540d\u79f0\u3001\u6700\u5927\u5e76\u53d1\u6570<\/td>\n \u6bcf\u79d2 token \u6570\u3001\u5e73\u5747\u5ef6\u8fdf\u3001GPU \u540d\u79f0<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n \u4f60\u53ef\u4ee5\u4f7f\u7528\u968f\u673a\u91c7\u6837\u7b49\u6280\u672f\uff0c\u4e3a\u6d4b\u8bd5\u9009\u62e9\u5177\u4ee3\u8868\u6027\u7684\u6570\u636e\u70b9\u3002\u7ebf\u6027\u4f18\u5316\u6709\u52a9\u4e8e\u4f60\u627e\u5230\u6279\u5904\u7406\u548c TensorRT \u8bbe\u7f6e\u7684\u6700\u4f73\u7ec4\u5408\uff1b\u8d1d\u53f6\u65af\u63a8\u65ad\u5219\u53ef\u4ee5\u5728\u6570\u636e\u6709\u9650\u7684\u60c5\u51b5\u4e0b\u4f30\u8ba1\u6027\u80fd\u8868\u73b0\uff0c\u8fd9\u5728\u5b58\u5728\u4e0d\u786e\u5b9a\u6027\u7684\u573a\u666f\u4e2d\u5c24\u4e3a\u6709\u7528\u3002\u8fd9\u4e9b\u65b9\u6cd5\u80fd\u591f\u5e2e\u52a9\u4f60\u5c06 AI \u5de5\u4f5c\u8d1f\u8f7d\u8c03\u4f18\u5230\u66f4\u9ad8\u7684\u6548\u7387\u548c\u66f4\u597d\u7684\u6027\u80fd\u6c34\u5e73\u3002<\/p>\n \u4f60\u8fd8\u9700\u8981\u91cd\u89c6\u6279\u5904\u7406\u7b56\u7565\u3002\u6279\u5904\u7406\u901a\u8fc7\u5c06\u591a\u4e2a\u63a8\u7406\u8bf7\u6c42\u5408\u5e76\u5904\u7406\uff0c\u6765\u63d0\u5347\u541e\u5410\u91cf\u3002NVIDIA Triton Inference Server \u548c TensorRT \u90fd\u652f\u6301\u52a8\u6001\u6279\u5904\u7406\uff0c\u53ef\u4ee5\u6839\u636e\u8d1f\u8f7d\u53d8\u5316\u8c03\u6574\u6279\u5927\u5c0f\u3002\u5bf9\u4e8e\u5b9e\u65f6\u63a8\u7406\uff0c\u8fd9\u79cd\u65b9\u5f0f\u53ef\u4ee5\u5728\u964d\u4f4e\u6a21\u578b\u5ef6\u8fdf\u7684\u540c\u65f6\u63d0\u5347\u6574\u4f53\u541e\u5410\u91cf\u3002<\/p>\n \u6ce8\u610f\uff1a\u5408\u7406\u7684\u6279\u5904\u7406\u7b56\u7565\u548c TensorRT \u4f18\u5316\uff0c\u662f\u5b9e\u73b0\u9ad8\u6027\u80fd AI \u63a8\u7406\u7684\u5173\u952e\u3002<\/p><\/blockquote>\n \u6536\u96c6\u6027\u80fd\u6570\u636e<\/h3>\n \u5728\u57fa\u51c6\u6d4b\u8bd5\u8fc7\u7a0b\u4e2d\uff0c\u4f60\u9700\u8981\u6536\u96c6\u8be6\u5c3d\u7684\u6027\u80fd\u6570\u636e\u3002Perf Analyzer \u7b49\u5de5\u5177\u53ef\u4ee5\u5e2e\u52a9\u4f60\u76d1\u63a7\u670d\u52a1\u5668\u7aef\u6307\u6807\uff0c\u5305\u62ec GPU \u5229\u7528\u7387\u548c\u529f\u8017\u3002\u4f60\u53ef\u4ee5\u901a\u8fc7 --collect-metrics<\/code> \u547d\u4ee4\u884c\u9009\u9879\u542f\u7528\u6307\u6807\u91c7\u96c6\u3002\u9ed8\u8ba4\u60c5\u51b5\u4e0b\uff0c\u6307\u6807\u4f1a\u66b4\u9732\u5728 localhost:8002\/metrics<\/code>\uff0c\u4f60\u4e5f\u53ef\u4ee5\u4f7f\u7528 --metrics-url<\/code> \u9009\u9879\u66f4\u6539\u8be5\u5730\u5740\u3002\u9ed8\u8ba4\u91c7\u96c6\u95f4\u9694\u4e3a 1000 \u6beb\u79d2\uff0c\u53ef\u5229\u7528 --metrics-interval<\/code> \u53c2\u6570\u8fdb\u884c\u8c03\u6574\u3002<\/p>\n Perf Analyzer \u4f1a\u5728\u591a GPU \u7cfb\u7edf\u4e2d\u6c47\u603b\u6bcf\u5757 GPU \u7684\u6307\u6807\u3002\u4f60\u53ef\u4ee5\u5229\u7528 -f <filename><\/code> \u548c --verbose-csv<\/code> \u9009\u9879\uff0c\u5c06\u7ed3\u679c\u5bfc\u51fa\u4e3a CSV \u6587\u4ef6\uff0c\u65b9\u4fbf\u540e\u7eed\u5206\u6790\u8d8b\u52bf\u5e76\u8bc6\u522b\u6027\u80fd\u74f6\u9888\u3002<\/p>\n \u5728\u8fdb\u884c\u8d1f\u8f7d\u6d4b\u8bd5\u65f6\uff0c\u4f60\u5e94\u5f53\u53d8\u5316\u63a8\u7406\u8bf7\u6c42\u7684\u901f\u7387\u548c\u5206\u5e03\u7c7b\u578b\uff0c\u56e0\u4e3a\u8fd9\u4e9b\u56e0\u7d20\u5bf9\u5ef6\u8fdf\u548c\u541e\u5410\u91cf\u5f71\u54cd\u5f88\u5927\u3002\u4f8b\u5982\uff1a<\/p>\n \n \u8d85\u5927\u89c4\u6a21\u4e91\u73af\u5883\u7684\u57fa\u51c6\u6d4b\u8bd5\u4f1a\u91cd\u70b9\u5173\u6ce8\u6574\u4f53\u96c6\u7fa4\u6548\u7387\u548c\u670d\u52a1\u4e00\u81f4\u6027\u3002<\/li>\n \u672c\u5730\u4f01\u4e1a\u73af\u5883\u66f4\u5f3a\u8c03\u5728\u53d7\u63a7\u57fa\u7840\u8bbe\u65bd\u5185\u7684\u6027\u80fd\u9a8c\u8bc1\u3002<\/li>\n \u8fb9\u7f18\u63a8\u7406\u573a\u666f\u5219\u5c24\u5176\u5173\u6ce8\u4f4e\u5ef6\u8fdf\u548c\u529f\u8017\u9650\u5236\u3002<\/li>\n \u5d4c\u5165\u5f0f\u6216\u5de5\u4e1a\u90e8\u7f72\u66f4\u5f3a\u8c03\u786e\u5b9a\u6027\u548c\u957f\u671f\u7a33\u5b9a\u8fd0\u884c\u3002<\/li>\n<\/ul>\n \u4f60\u53ef\u4ee5\u6839\u636e\u8fd9\u4e9b\u6d1e\u5bdf\uff0c\u9488\u5bf9\u4e0d\u540c\u90e8\u7f72\u573a\u666f\u8c03\u6574 TensorRT \u548c\u6279\u5904\u7406\u914d\u7f6e\u3002\u6027\u80fd\u8c03\u4f18\u9700\u8981\u4f60\u4ed4\u7ec6\u5206\u6790\u6536\u96c6\u5230\u7684\u6570\u636e\uff0c\u5e76\u636e\u6b64\u505a\u51fa\u6539\u52a8\uff0c\u4ee5\u63d0\u5347\u63a8\u7406\u6548\u7387\u3002<\/p>\n \u63d0\u793a\uff1a\u5728\u8d1f\u8f7d\u6d4b\u8bd5\u4e2d\u59cb\u7ec8\u76d1\u63a7\u6a21\u578b\u5ef6\u8fdf\u3001\u541e\u5410\u91cf\u548c\u8d44\u6e90\u4f7f\u7528\u60c5\u51b5\uff0c\u786e\u4fdd AI \u5de5\u4f5c\u8d1f\u8f7d\u8fbe\u5230\u9884\u671f\u6027\u80fd\u76ee\u6807\u3002<\/p><\/blockquote>\n \u7ed3\u679c\u5206\u6790\u4e0e\u4f18\u5316<\/strong><\/h2>\n \u8bc6\u522b\u74f6\u9888<\/h3>\n \u4f60\u9700\u8981\u5bf9\u57fa\u51c6\u6d4b\u8bd5\u7ed3\u679c\u8fdb\u884c\u6df1\u5165\u5206\u6790\uff0c\u4ee5\u53d1\u73b0\u9650\u5236\u63a8\u7406\u6548\u7387\u7684\u74f6\u9888\u3002\u5e38\u89c1\u95ee\u9898\u5305\u62ec\u9ad8\u5ef6\u8fdf\u3001\u4f4e\u541e\u5410\u91cf\u4ee5\u53ca\u5185\u5b58\u9650\u5236\u3002\u5f53\u4f60\u8fd0\u884c AI \u5de5\u4f5c\u8d1f\u8f7d\u65f6\uff0c\u53ef\u80fd\u4f1a\u53d1\u73b0\u5927\u8bed\u8a00\u6a21\u578b\u662f\u4ee5 token \u4e3a\u7c92\u5ea6\u9010\u4e2a\u751f\u6210\u7684\uff0c\u8fd9\u4f1a\u5728\u957f\u63d0\u793a\u8bcd\u573a\u666f\u4e0b\u663e\u8457\u589e\u52a0\u6a21\u578b\u5ef6\u8fdf\u3002\u5728\u8bf7\u6c42\u91cf\u5927\u3001\u5ef6\u8fdf\u9ad8\u7684\u60c5\u51b5\u4e0b\uff0c\u670d\u52a1\u5668\u53ef\u80fd\u51fa\u73b0\u5361\u987f\u6216\u8d85\u65f6\u3002\u6d41\u5f0f\u6279\u5904\u7406\uff08inflight batching\uff09\u53ef\u4ee5\u63d0\u4f9b\u5e2e\u52a9\uff0c\u4f46\u540c\u65f6\u4e5f\u8981\u5bc6\u5207\u5173\u6ce8\u5185\u5b58\u9700\u6c42\u3002\u66f4\u5927\u7684\u6a21\u578b\u548c\u66f4\u957f\u7684\u63d0\u793a\u8bcd\u90fd\u4f1a\u6d88\u8017\u66f4\u591a\u663e\u5b58\u3002\u968f\u7740\u6a21\u578b\u53d8\u5927\uff0c\u80fd\u8017\u4e5f\u4f1a\u4e0a\u5347\uff0c\u800c\u4f7f\u7528\u9ad8\u7aef\u8bbe\u5907\u8fdb\u884c\u6269\u5c55\u4f1a\u5e26\u6765\u66f4\u9ad8\u7684\u6210\u672c\u3002\u7ba1\u7406\u6a21\u578b\u66f4\u65b0\u4e0e\u7248\u672c\u8fed\u4ee3\u540c\u6837\u5173\u952e\uff0c\u6709\u52a9\u4e8e\u5e73\u7a33\u5b8c\u6210\u6a21\u578b\u90e8\u7f72\u3002<\/p>\n \n \u5ef6\u8fdf\uff1aLLM \u6309 token \u9010\u6b65\u751f\u6210\uff0c\u5bfc\u81f4\u9884\u6d4b\u65f6\u95f4\u53d8\u957f\u3002<\/li>\n \u541e\u5410\u91cf\uff1a\u9ad8\u8bf7\u6c42\u91cf\u53e0\u52a0\u9ad8\u5ef6\u8fdf\uff0c\u53ef\u80fd\u5bfc\u81f4\u670d\u52a1\u5668\u5361\u6b7b\u3002<\/li>\n \u5185\u5b58\uff1a\u66f4\u5927\u7684\u6a21\u578b\u548c\u66f4\u957f\u7684\u63d0\u793a\u8bcd\u9700\u8981\u66f4\u591a VRAM\u3002<\/li>\n \u80fd\u8017\uff1a\u5927\u578b\u6a21\u578b\u5bf9\u529f\u8017\u8981\u6c42\u66f4\u9ad8\u3002<\/li>\n \u53ef\u6269\u5c55\u6027\uff1a\u9ad8\u7aef\u673a\u5668\u5728\u6269\u5bb9\u65f6\u6210\u672c\u66f4\u9ad8\u3002<\/li>\n \u6a21\u578b\u66f4\u65b0\uff1a\u7248\u672c\u7ba1\u7406\u662f\u5e73\u7a33\u90e8\u7f72\u7684\u5173\u952e\u4e00\u73af\u3002<\/li>\n<\/ul>\n \u63d0\u793a\uff1a\u5728\u5206\u6790\u74f6\u9888\u65f6\uff0c\u8981\u91cd\u70b9\u5173\u6ce8 GPU \u5229\u7528\u7387\u548c\u7f51\u7edc\u57fa\u7840\u8bbe\u65bd\uff1b\u4ec5\u4f9d\u8d56\u4f20\u7edf CPU \u6307\u6807\uff0c\u5f80\u5f80\u65e0\u6cd5\u771f\u5b9e\u53cd\u6620 AI \u63a8\u7406\u6027\u80fd\u3002<\/p><\/blockquote>\n \u4f18\u5316\u63a8\u7406\u6548\u7387<\/h3>\n \u4f60\u53ef\u4ee5\u901a\u8fc7\u591a\u79cd\u4f18\u5316\u7b56\u7565\u6765\u63d0\u5347\u63a8\u7406\u6548\u7387\u3002\u91cf\u5316\uff08quantization\uff09\u901a\u8fc7\u964d\u4f4e\u6a21\u578b\u7cbe\u5ea6\uff0c\u51cf\u5c11\u5185\u5b58\u5360\u7528\u5e76\u52a0\u5feb\u63a8\u7406\u901f\u5ea6\u3002\u8fde\u7eed\u6279\u5904\u7406\uff08continuous batching\uff09\u5219\u4ee5\u52a8\u6001\u65b9\u5f0f\u805a\u5408\u63a8\u7406\u8bf7\u6c42\uff0c\u4ee5\u6700\u5927\u5316\u541e\u5410\u91cf\u3002KV \u7f13\u5b58\uff08KV caching\uff09\u53ef\u4ee5\u907f\u514d\u91cd\u590d\u8ba1\u7b97\uff0c\u52a0\u5feb token \u751f\u6210\u3002\u63d0\u524d\u9000\u51fa\u673a\u5236\uff08early exit\uff09\u5141\u8bb8\u6a21\u578b\u5728\u65e0\u9700\u5b8c\u6574\u8ba1\u7b97\u6240\u6709\u5c42\u7684\u60c5\u51b5\u4e0b\u63d0\u524d\u7ed9\u51fa\u7ed3\u679c\uff0c\u4ece\u800c\u8282\u7701\u65f6\u95f4\u3002\u4f4e\u79e9\u5206\u89e3\uff08low-rank factorization\uff09\u628a\u5927\u77e9\u9635\u5206\u89e3\u4e3a\u66f4\u5c0f\u7684\u77e9\u9635\uff0c\u4ee5\u83b7\u5f97\u66f4\u9ad8\u7684\u8fd0\u884c\u901f\u5ea6\u548c\u66f4\u4f4e\u7684\u5185\u5b58\u5360\u7528\u3002\u57fa\u7840\u8bbe\u65bd\u4f18\u5316\u5219\u901a\u8fc7\u6539\u8fdb\u7f51\u7edc\u67b6\u6784\uff0c\u964d\u4f4e\u6570\u636e\u4f20\u8f93\u65f6\u95f4\u3002\u7f13\u5b58\u4e0e\u5907\u5fd8\u5f55\uff08caching vs. memoization\uff09\u4f1a\u5b58\u50a8\u4e2d\u95f4\u7ed3\u679c\uff0c\u4e3a\u540e\u7eed\u8bf7\u6c42\u63d0\u4f9b\u66f4\u5feb\u7684\u54cd\u5e94\u3002\u901a\u8fc7\u5e76\u884c\u5316\u548c\u6279\u5904\u7406\uff0c\u4f60\u53ef\u4ee5\u5229\u7528\u591a\u53f0\u670d\u52a1\u5668\u548c\u591a\u5757 GPU\uff0c\u4ee5\u83b7\u5f97\u66f4\u597d\u7684\u6574\u4f53\u6027\u80fd\u3002<\/p>\n \n\n\n\n<\/colgroup>\n\n\n\n\n\n\n\n\n\n\n \u7b56\u7565<\/th>\n \u8bf4\u660e<\/th>\n<\/tr>\n \u91cf\u5316<\/td>\n \u964d\u4f4e\u6a21\u578b\u7cbe\u5ea6\uff0c\u4ee5\u51cf\u5c11\u5185\u5b58\u5360\u7528\u5e76\u52a0\u901f\u63a8\u7406\u3002<\/td>\n<\/tr>\n \u63d0\u524d\u9000\u51fa\u673a\u5236<\/td>\n \u5141\u8bb8\u6a21\u578b\u5728\u5e76\u975e\u6240\u6709\u5c42\u5b8c\u6210\u8ba1\u7b97\u65f6\u5c31\u7ed9\u51fa\u9884\u6d4b\u7ed3\u679c\u3002<\/td>\n<\/tr>\n \u4f4e\u79e9\u5206\u89e3<\/td>\n \u5c06\u5927\u77e9\u9635\u5206\u89e3\u4e3a\u66f4\u5c0f\u77e9\u9635\uff0c\u4ee5\u83b7\u5f97\u66f4\u9ad8\u901f\u5ea6\u548c\u66f4\u4f4e\u5185\u5b58\u5360\u7528\u3002<\/td>\n<\/tr>\n \u8fde\u7eed\u6279\u5904\u7406<\/td>\n \u4ee5\u52a8\u6001\u65b9\u5f0f\u805a\u5408\u8bf7\u6c42\uff0c\u4ece\u800c\u63d0\u5347\u541e\u5410\u91cf\u3002<\/td>\n<\/tr>\n KV \u7f13\u5b58<\/td>\n \u51cf\u5c11\u91cd\u590d\u8ba1\u7b97\uff0c\u52a0\u5feb token \u751f\u6210\u901f\u5ea6\u3002<\/td>\n<\/tr>\n \u57fa\u7840\u8bbe\u65bd\u4f18\u5316<\/td>\n \u6539\u8fdb\u7f51\u7edc\u67b6\u6784\uff0c\u7f29\u77ed\u6570\u636e\u4f20\u8f93\u65f6\u95f4\u3002<\/td>\n<\/tr>\n \u7f13\u5b58\u4e0e\u5907\u5fd8\u5f55<\/td>\n \u5b58\u50a8\u4e2d\u95f4\u7ed3\u679c\uff0c\u4e3a\u540e\u7eed\u63a8\u7406\u8bf7\u6c42\u63d0\u4f9b\u66f4\u5feb\u54cd\u5e94\u3002<\/td>\n<\/tr>\n \u5e76\u884c\u5316\u4e0e\u6279\u5904\u7406<\/td>\n \u5229\u7528\u591a\u53f0\u670d\u52a1\u5668\u548c\u591a\u5757 GPU \u63d0\u5347\u6574\u4f53\u6027\u80fd\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n \u4f60\u8fd8\u5e94\u5f53\u5bf9\u591a\u7ebf\u7a0b\u8fdb\u884c\u8c03\u4f18\u3002\u5982\u679c\u7ebf\u7a0b\u6570\u8bbe\u7f6e\u8fc7\u591a\uff0c\u53ef\u80fd\u4f1a\u5bfc\u81f4\u5ef6\u8fdf\u5347\u9ad8\u548c GPU \u5229\u7528\u7387\u4e0b\u964d\u3002\u8ba9\u7ebf\u7a0b\u6570\u91cf\u4e0e\u5e76\u53d1\u63a8\u7406\u8bf7\u6c42\u6570\u66f4\u597d\u5730\u5339\u914d\uff0c\u53ef\u4ee5\u540c\u65f6\u63d0\u9ad8\u541e\u5410\u91cf\u5e76\u964d\u4f4e\u5ef6\u8fdf\u3002\u501f\u52a9 SageMaker Inference Recommender \u7b49\u5de5\u5177\u8fdb\u884c\u6027\u80fd\u8c03\u4f18\uff0c\u53ef\u4ee5\u5e2e\u52a9\u4f60\u4e3a AI \u5de5\u4f5c\u8d1f\u8f7d\u627e\u5230\u6700\u5408\u9002\u7684\u6279\u5904\u7406\u548c TensorRT \u914d\u7f6e\u3002<\/p>\n \u501f\u52a9 SageMaker \u8fdb\u884c AI \u6269\u5c55<\/h3>\n Amazon SageMaker \u4e3a\u4f60\u5728 AWS \u4e0a\u6269\u5c55 AI \u63a8\u7406\u63d0\u4f9b\u4e86\u5f3a\u5927\u5de5\u5177\u3002\u4f60\u53ef\u4ee5\u521b\u5efa\u5b9e\u65f6\u63a8\u7406\u548c\u6279\u91cf\u63a8\u7406\u7aef\u70b9\uff0c\u9ad8\u6548\u7ba1\u7406\u6df1\u5ea6\u5b66\u4e60\u548c\u673a\u5668\u5b66\u4e60\u5de5\u4f5c\u8d1f\u8f7d\u3002SageMaker \u652f\u6301\u65e0\u670d\u52a1\u5668\u90e8\u7f72\uff08serverless\uff09\u548c\u81ea\u52a8\u6269\u7f29\u5bb9\u673a\u5236\uff0c\u4ece\u800c\u5728\u4fdd\u6301\u8d44\u6e90\u5e73\u8861\u7684\u540c\u65f6\u63a7\u5236\u6210\u672c\u3002\u4f60\u53ef\u4ee5\u901a\u8fc7 Amazon CloudWatch \u76d1\u63a7\u6027\u80fd\uff0c\u5e76\u5feb\u901f\u6392\u67e5\u95ee\u9898\u3002SageMaker \u8d1f\u8d23\u8de8\u591a\u4e2a\u53ef\u7528\u533a\u7ba1\u7406\u6a21\u578b\u90e8\u7f72\uff0c\u4ee5\u786e\u4fdd\u9ad8\u53ef\u9760\u6027\u3002\u591a\u6a21\u578b\u7aef\u70b9\uff08MMEs\uff09\u5141\u8bb8\u4f60\u5728\u540c\u4e00\u4e2a\u7aef\u70b9\u4e0a\u90e8\u7f72\u591a\u4e2a\u6a21\u578b\uff0c\u4ece\u800c\u6700\u5927\u5316 GPU \u5229\u7528\u7387\u5e76\u964d\u4f4e\u6258\u7ba1\u6210\u672c\u3002SageMaker \u4f1a\u6839\u636e\u6d41\u91cf\u60c5\u51b5\u52a8\u6001\u52a0\u8f7d\u548c\u5378\u8f7d\u6a21\u578b\uff0c\u5728\u4fdd\u8bc1\u5b9e\u65f6\u63a8\u7406\u573a\u666f\u4e2d\u4e9a\u79d2\u7ea7\u5ef6\u8fdf\u7684\u540c\u65f6\u5b9e\u73b0\u9ad8\u6027\u80fd\u3002SageMaker Inference Recommender \u80fd\u5e2e\u52a9\u4f60\u6d4b\u8bd5\u4e0d\u540c\u914d\u7f6e\u3001\u4f18\u5316\u6279\u5904\u7406\u5e76\u8c03\u4f18 TensorRT \u8bbe\u7f6e\uff0c\u4e3a\u5b9e\u65f6\u7aef\u70b9\u627e\u5230\u6700\u4f73\u7ec4\u5408\u3002\u4f60\u53ef\u4ee5\u5c06 SageMaker \u4e0e NVIDIA Triton Inference Server \u7ed3\u5408\u4f7f\u7528\uff0c\u83b7\u5f97\u9ad8\u541e\u5410\u7684\u6a21\u578b\u670d\u52a1\u80fd\u529b\u3002SageMaker \u4e5f\u652f\u6301\u6027\u80fd\u8c03\u4f18\u3001\u8d1f\u8f7d\u6d4b\u8bd5\u548c\u57fa\u51c6\u6d4b\u8bd5\uff0c\u4f7f\u4f60\u80fd\u591f\u5728 AWS \u4e0a\u8f7b\u677e\u4f18\u5316 AI \u5de5\u4f5c\u8d1f\u8f7d\u3002<\/p>\n \n\n\n\n<\/colgroup>\n\n\n\n\n\n\n\n \u7279\u6027<\/th>\n \u8bf4\u660e<\/th>\n<\/tr>\n \u5b9e\u65f6\u63a8\u7406\u4e0e\u6279\u91cf\u63a8\u7406<\/td>\n \u521b\u5efa\u7aef\u70b9\u4ee5\u83b7\u5f97\u5373\u65f6\u9884\u6d4b\uff0c\u5e76\u9ad8\u6548\u7ba1\u7406\u5de5\u4f5c\u8d1f\u8f7d\u3002<\/td>\n<\/tr>\n \u65e0\u670d\u52a1\u5668\u4e14\u5177\u6210\u672c\u4f18\u52bf<\/td>\n \u901a\u8fc7\u81ea\u52a8\u6269\u7f29\u5bb9\u548c AWS Lambda \u96c6\u6210\u5b9e\u73b0\u52a8\u6001\u8d44\u6e90\u7ba1\u7406\u3002<\/td>\n<\/tr>\n \u76d1\u63a7\u4e0e\u8c03\u8bd5<\/td>\n \u4f7f\u7528 Amazon CloudWatch \u8fdb\u884c\u5b9e\u65f6\u76d1\u63a7\uff0c\u5e76\u5229\u7528\u8c03\u8bd5\u5de5\u5177\u652f\u6301\u6a21\u578b\u90e8\u7f72\u3002<\/td>\n<\/tr>\n \u90e8\u7f72\u80fd\u529b<\/td>\n \u5229\u7528\u591a\u79cd\u5b9e\u4f8b\u7c7b\u578b\u4e0e\u591a\u4e2a\u53ef\u7528\u533a\u6269\u5c55\u57fa\u7840\u8bbe\u65bd\u3002<\/td>\n<\/tr>\n SageMaker \u591a\u6a21\u578b\u7aef\u70b9\uff08MMEs\uff09<\/td>\n \u5728\u5355\u4e00\u7aef\u70b9\u4e0a\u90e8\u7f72\u591a\u4e2a\u6a21\u578b\uff0c\u5b9e\u73b0\u5177\u6210\u672c\u4f18\u52bf\u4e14\u53ef\u6269\u5c55\u7684\u5b9e\u65f6\u63a8\u7406\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n \u6ce8\u610f\uff1aSageMaker Inference Recommender \u662f\u4f18\u5316\u6a21\u578b\u90e8\u7f72\u3001\u6279\u5904\u7406\u548c TensorRT \u8bbe\u7f6e\u3001\u5b9e\u73b0\u9876\u7ea7\u63a8\u7406\u6548\u7387\u7684\u5173\u952e\u5de5\u5177\u3002<\/p><\/blockquote>\n \n <\/div>\n<\/div>\n \u901a\u8fc7\u9075\u5faa\u6e05\u6670\u7684\u6d41\u7a0b\uff0c\u4f60\u53ef\u4ee5\u5927\u5e45\u63d0\u5347\u7f8e\u56fd\u670d\u52a1\u5668\u4e0a\u7684\u63a8\u7406\u6548\u7387\u3002\u5148\u4f7f\u7528\u6807\u51c6\u5316\u57fa\u51c6\u6846\u67b6\u5ea6\u91cf\u6027\u80fd\uff0c\u7136\u540e\u501f\u52a9 SageMaker \u548c MLPerf \u83b7\u5f97\u51c6\u786e\u7684\u63a8\u7406\u7ed3\u679c\u3002\u518d\u901a\u8fc7 SageMaker Inference Recommender \u5b8c\u6210\u6a21\u578b\u90e8\u7f72\u4e0e\u6027\u80fd\u8c03\u4f18\u3002\u6700\u540e\uff0c\u6301\u7eed\u7ed3\u5408\u5b9e\u65f6\u63a8\u7406\u3001\u8d1f\u8f7d\u6d4b\u8bd5\u548c\u7aef\u70b9\u76d1\u63a7\uff0c\u4e0d\u65ad\u4f18\u5316\u4f60\u7684 AI \u5de5\u4f5c\u8d1f\u8f7d\u3002<\/p>\n \u6301\u7eed\u4f18\u5316\u5bf9\u4e8e\u8d44\u6e90\u7ba1\u7406\u3001\u6027\u80fd\u63d0\u5347\u4ee5\u53ca\u6a21\u578b\u517c\u5bb9\u6027\u90fd\u81f3\u5173\u91cd\u8981\u3002<\/p>\n \n\n\n\n\n<\/colgroup>\n\n\n\n\n\n \u7b56\u7565\u5c42\u7ea7<\/th>\n \u5173\u6ce8\u9886\u57df<\/th>\n \u5173\u952e\u884c\u52a8<\/th>\n<\/tr>\n 1<\/td>\n \u786c\u4ef6\u5229\u7528<\/td>\n \u4f7f\u7528\u6279\u5904\u7406\u3001\u91cf\u5316\u548c CUDA \u56fe\u3002<\/td>\n<\/tr>\n 2<\/td>\n \u5185\u5b58\u4e0e\u7b97\u6cd5<\/td>\n \u6d88\u9664\u5185\u5b58\u74f6\u9888\u3002<\/td>\n<\/tr>\n 3<\/td>\n \u6269\u5c55\u62d3\u6251<\/td>\n \u5c06\u670d\u52a1\u5668\u5c31\u8fd1\u90e8\u7f72\u5e76\u5408\u7406\u7ba1\u7406\u526f\u672c\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n \u7ed3\u5408 SageMaker Inference Recommender\uff0c\u5728 AWS \u4e0a\u63a2\u7d22\u66f4\u9ad8\u7ea7\u7684 AI \u4e0e\u6df1\u5ea6\u5b66\u4e60\u7b56\u7565\uff0c\u6301\u7eed\u8fdb\u884c\u6027\u80fd\u8c03\u4f18\uff0c\u5e76\u9ad8\u6548\u5904\u7406\u4e0d\u65ad\u589e\u957f\u7684\u63a8\u7406\u8bf7\u6c42\u3002<\/p>\n \u5e38\u89c1\u95ee\u7b54<\/strong><\/h2>\n \u5728\u7f8e\u56fd\u670d\u52a1\u5668\u4e0a\u8861\u91cf\u63a8\u7406\u6027\u80fd\u7684\u6700\u4f73\u65b9\u5f0f\u662f\u4ec0\u4e48\uff1f<\/h3>\n \u4f60\u5e94\u8be5\u4f7f\u7528 MLPerf Inference \u57fa\u51c6\u6d4b\u8bd5\u3002\u8fd9\u4e9b\u6d4b\u8bd5\u80fd\u591f\u4e3a AI \u5de5\u4f5c\u8d1f\u8f7d\u63d0\u4f9b\u6e05\u6670\u7684\u6027\u80fd\u7ed3\u679c\u3002SageMaker \u53ef\u4ee5\u5e2e\u52a9\u4f60\u8fd0\u884c\u8fd9\u4e9b\u57fa\u51c6\uff0c\u5bf9\u6bd4\u4e0d\u540c\u786c\u4ef6\u7684\u8868\u73b0\u3002\u4f60\u53ef\u4ee5\u5728\u6bcf\u6b21\u6d4b\u8bd5\u4e2d\u8ddf\u8e2a\u5ef6\u8fdf\u3001\u541e\u5410\u91cf\u548c\u8d44\u6e90\u4f7f\u7528\u60c5\u51b5\u3002<\/p>\n SageMaker \u5982\u4f55\u63d0\u5347 AI \u63a8\u7406\u6027\u80fd\uff1f<\/h3>\n SageMaker \u8ba9\u4f60\u53ef\u4ee5\u9ad8\u6548\u90e8\u7f72\u6a21\u578b\u3002\u4f60\u53ef\u4ee5\u5229\u7528\u6279\u5904\u7406\u3001\u591a\u7ebf\u7a0b\u4ee5\u53ca NVIDIA Triton \u7b49\u5de5\u5177\u6765\u63d0\u5347\u63a8\u7406\u6027\u80fd\u3002SageMaker \u540c\u65f6\u63d0\u4f9b\u76d1\u63a7\u4e0e\u8c03\u4f18\u5de5\u5177\uff0c\u5e2e\u52a9\u4f60\u6301\u7eed\u4f18\u5316 AI \u5de5\u4f5c\u8d1f\u8f7d\uff0c\u83b7\u5f97\u66f4\u597d\u7684\u7ed3\u679c\u3002<\/p>\n \u4e3a\u4ec0\u4e48\u6279\u5904\u7406\u5bf9\u63a8\u7406\u6027\u80fd\u5f88\u91cd\u8981\uff1f<\/h3>\n \u6279\u5904\u7406\u4f1a\u5c06\u591a\u4e2a\u8bf7\u6c42\u7ec4\u5408\u5728\u4e00\u8d77\u5904\u7406\uff0c\u8fd9\u4e00\u65b9\u5f0f\u53ef\u4ee5\u663e\u8457\u63d0\u5347\u541e\u5410\u91cf\u5e76\u964d\u4f4e\u5ef6\u8fdf\u3002SageMaker \u548c NVIDIA Triton \u90fd\u652f\u6301\u52a8\u6001\u6279\u5904\u7406\u3002\u53ea\u8981\u4e3a\u4f60\u7684\u5de5\u4f5c\u8d1f\u8f7d\u9009\u62e9\u5408\u9002\u7684\u6279\u5927\u5c0f\uff0c\u5c31\u80fd\u770b\u5230\u66f4\u597d\u7684 AI \u63a8\u7406\u8868\u73b0\u3002<\/p>\n \u5982\u4f55\u5728 SageMaker \u4e2d\u76d1\u63a7\u63a8\u7406\u6027\u80fd\uff1f<\/h3>\n \u4f60\u53ef\u4ee5\u5c06 Amazon CloudWatch \u4e0e SageMaker \u7ed3\u5408\u4f7f\u7528\u3002CloudWatch \u4f1a\u8ddf\u8e2a\u5ef6\u8fdf\u3001\u541e\u5410\u91cf\u548c GPU \u5229\u7528\u7387\u7b49\u6307\u6807\u3002\u4f60\u8fd8\u53ef\u4ee5\u4e3a\u6027\u80fd\u4e0b\u964d\u8bbe\u7f6e\u544a\u8b66\u3002SageMaker \u540c\u65f6\u5141\u8bb8\u4f60\u6d4b\u8bd5\u4e0d\u540c\u914d\u7f6e\uff0c\u4ee5\u8fdb\u4e00\u6b65\u6539\u8fdb AI \u63a8\u7406\u7ed3\u679c\u3002<\/p>\n \u80fd\u5426\u901a\u8fc7 SageMaker \u7aef\u70b9\u6269\u5c55 AI \u63a8\u7406\uff1f<\/h3>\n \u53ef\u4ee5\u3002\u4f60\u53ef\u4ee5\u4f7f\u7528 SageMaker \u7aef\u70b9\u6765\u6269\u5c55\u63a8\u7406\u80fd\u529b\u3002SageMaker \u652f\u6301\u81ea\u52a8\u6269\u7f29\u5bb9\u548c\u591a\u6a21\u578b\u7aef\u70b9\uff0c\u4ece\u800c\u5728\u5e94\u5bf9\u66f4\u591a\u8bf7\u6c42\u7684\u540c\u65f6\u4fdd\u6301\u9ad8\u6027\u80fd\u3002\u968f\u7740\u4e1a\u52a1\u9700\u6c42\u589e\u957f\uff0cSageMaker \u80fd\u5e2e\u52a9\u4f60\u8f7b\u677e\u7ba1\u7406 AI \u5de5\u4f5c\u8d1f\u8f7d\u3002<\/p>\n","protected":false},"excerpt":{"rendered":" \u5f53\u4f60\u5728\u7f8e\u56fd\u670d\u52a1\u5668\u4e0a\u90e8\u7f72 AI \u5de5\u4f5c\u8d1f\u8f7d\u65f6\uff0c\u9700\u8981\u5bf9\u63a8\u7406\u6548\u7387\u8fdb\u884c\u5ea6\u91cf\u3002MLPerf Inference \u7b49\u884c\u4e1a\u6807\u51c6\u57fa\u51c6\u6d4b\u8bd5\uff0c\u53ef\u4ee5\u4e3a\u4f60\u63d0\u4f9b\u6e05\u6670\u7684\u670d\u52a1\u5668\u6027\u80fd\u5bf9\u6bd4\u3002\u4f8b\u5982\uff0c\u50cf SR650a V4 \u548c SR680a V4 \u8fd9\u6837\u7684\u7f8e\u56fd\u9876\u7ea7\u670d\u52a1\u5668\u578b\u53f7\uff0c\u5728 AI \u548c\u673a\u5668\u5b66\u4e60\u4efb\u52a1\u7684\u6548\u7387\u65b9\u9762\u5904\u4e8e\u9886\u5148\u5730\u4f4d\u3002Amazon SageMaker \u5e2e\u52a9\u4f60\u5728\u5927\u89c4\u6a21\u573a\u666f\u4e0b\u7ba1\u7406\u6df1\u5ea6\u5b66\u4e60\u63a8\u7406\uff0c\u800c NVIDIA Triton Inference Server \u5219\u652f\u6301\u9ad8\u541e\u5410\u91cf AI\u3002\u4f60\u4f1a\u53d1\u73b0\uff0cSageMaker \u80fd\u4e0e AWS \u65e0\u7f1d\u534f\u540c\uff0c\u7528\u4e8e\u53ef\u6269\u5c55\u7684 AI \u63a8\u7406\u3002\u5f53\u4f60\u4f7f\u7528 SageMaker \u65f6\uff0c\u5c31\u53ef\u4ee5\u6539\u8fdb\u63a8\u7406\u6027\u80fd\uff0c\u5e76\u4f18\u5316\u673a\u5668\u5b66\u4e60\u5de5\u4f5c\u8d1f\u8f7d\u3002 \u5173\u952e\u8981\u70b9 \u4f7f\u7528 MLPerf Inference \u57fa\u51c6\u6d4b\u8bd5\u6765\u8861\u91cf\u7f8e\u56fd\u670d\u52a1\u5668\u4e0a\u7684 AI \u6027\u80fd\u3002 \u901a\u8fc7\u8c03\u6574\u6279\u5927\u5c0f\u548c\u4f7f\u7528\u66f4\u5c0f\u7684\u6a21\u578b\u6765\u4f18\u5316\u5ef6\u8fdf\u4e0e\u541e\u5410\u91cf\u3002 \u76d1\u63a7\u8d44\u6e90\u5229\u7528\u7387\uff0c\u786e\u4fdd\u786c\u4ef6\u548c\u80fd\u6e90\u4f7f\u7528\u7684\u9ad8\u6548\u7387\u3002 \u5229\u7528 Amazon SageMaker \u5b9e\u73b0\u53ef\u6269\u5c55\u7684 AI \u63a8\u7406\u548c\u6027\u80fd\u8c03\u4f18\u3002 \u8ba4\u771f\u51c6\u5907\u6570\u636e\uff0c\u4ee5\u63d0\u5347\u63a8\u7406\u7684\u51c6\u786e\u6027\u4e0e\u4e00\u81f4\u6027\u3002 \u63a8\u7406\u6548\u7387\u6307\u6807 \u5ef6\u8fdf\u4e0e\u541e\u5410\u91cf \u4f60\u9700\u8981\u8861\u91cf\u5ef6\u8fdf\u548c\u541e\u5410\u91cf\uff0c\u6765\u4e86\u89e3 AI \u5de5\u4f5c\u8d1f\u8f7d\u5728\u7f8e\u56fd\u670d\u52a1\u5668\u4e0a\u7684\u8868\u73b0\u3002\u5ef6\u8fdf\u8868\u793a\u6a21\u578b\u5bf9\u7528\u6237\u8bf7\u6c42\u505a\u51fa\u54cd\u5e94\u6240\u9700\u7684\u65f6\u95f4\u3002\u5c24\u5176\u5728\u5b9e\u65f6\u5e94\u7528\u4e2d\uff0c\u5982\u679c\u4f60\u6709\u4e25\u683c\u7684\u5ef6\u8fdf\u8981\u6c42\uff0c\u5c31\u5e94\u540c\u65f6\u5173\u6ce8\u7aef\u5230\u7aef\u5ef6\u8fdf\u548c\u6a21\u578b\u5ef6\u8fdf\u3002\u5728\u505a\u4efb\u4f55\u66f4\u6539\u4e4b\u524d\uff0c\u5148\u5efa\u7acb\u63a8\u7406\u5ef6\u8fdf\u7684\u57fa\u7ebf\u3002\u4f60\u53ef\u4ee5\u901a\u8fc7\u4f7f\u7528\u66f4\u5c0f\u7684\u6a21\u578b\u6216\u91cf\u5316\u6a21\u578b\u6765\u6539\u5584\u6a21\u578b\u5ef6\u8fdf\u3002\u91cd\u70b9\u4f18\u5316\u9884\u586b\u5145\uff08prefill\uff09\u9636\u6bb5\uff0c\u5e76\u4f7f\u7528\u9ad8\u6548\u8def\u7531\uff0c\u4ee5\u964d\u4f4e\u9996 token \u65f6\u95f4\uff08time to first token\uff09\u3002\u541e\u5410\u91cf\u8861\u91cf\u7cfb\u7edf\u6bcf\u79d2\u53ef\u4ee5\u5904\u7406\u591a\u5c11 token […]<\/p>\n Read More…<\/a><\/p>\n","protected":false},"author":2,"featured_media":33335,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[4816],"tags":[12197,12198,12199,12200,4886],"class_list":["post-33340","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-america-dedicated-server-sc","tag-inference-efficiency","tag-amazon-sagemaker","tag-mlperf-inference","tag-nvidia-triton-inference-server","tag-us-dedicated-server-sc"],"acf":[],"yoast_head":"\n