\u5982\u4f55\u4e3a\u6a21\u578b\u52a0\u8f7d\u9009\u62e9 GPU \u670d\u52a1\u5668\u663e\u5b58<\/strong><\/h2>\n

\u4ece\u53c2\u6570\u91cf\u4f30\u7b97\u6743\u91cd\u663e\u5b58<\/h3>\n

\u5148\u4ece\u4e00\u4e2a\u7b80\u5355\u7684\u8ba1\u7b97\u5f00\u59cb\uff1a\u6a21\u578b\u6743\u91cd\u663e\u5b58 = \u53c2\u6570\u6570\u91cf \u00d7 \u6bcf\u4e2a\u53c2\u6570\u5360\u7528\u7684\u5b57\u8282\u6570\u3002\u4e0d\u540c\u7684\u6570\u636e\u7c7b\u578b\u5360\u7528\u7a7a\u95f4\u4e0d\u540c\u3002<\/p>\n

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

\u91cf\u5316\u683c\u5f0f<\/th>\n	\u6bcf\u4e2a\u53c2\u6570\u7684\u5b57\u8282\u6570<\/th>\n<\/tr>\n
FP32<\/td>\n	4 \u5b57\u8282<\/td>\n<\/tr>\n
FP16 \/ BF16<\/td>\n	2 \u5b57\u8282<\/td>\n<\/tr>\n
INT8<\/td>\n	1 \u5b57\u8282<\/td>\n<\/tr>\n
INT4<\/td>\n	0.5 \u5b57\u8282<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n \u4e00\u4e2a 70 \u4ebf\u53c2\u6570\u7684\u6a21\u578b\u5728 FP16 \u4e0b\uff0c\u6743\u91cd\u663e\u5b58\u9700\u6c42\u4e3a 14 GB\uff0c\u4e5f\u5c31\u662f 70 \u4ebf \u00d7 2 \u5b57\u8282\u3002\u5982\u679c\u5207\u6362\u5230 INT8\uff0c\u540c\u6837\u7684\u6a21\u578b\u53ea\u9700\u8981 7 GB\uff1bINT4 \u5219\u53ef\u4ee5\u8fdb\u4e00\u6b65\u964d\u5230 3.5 GB\u3002\u539f\u56e0\u662f\u6bcf\u4e2a\u6743\u91cd\u5360\u7528\u7684\u6bd4\u7279\u6570\u66f4\u5c11\uff1aINT4 \u6bcf\u4e2a\u6743\u91cd\u53ea\u7528 4 \u6bd4\u7279\uff0c\u800c FP32 \u9700\u8981 32 \u6bd4\u7279\uff0c\u663e\u5b58\u9700\u6c42\u56e0\u6b64\u51cf\u5c11 8 \u500d\u3002<\/p>\n \u5728\u8bad\u7ec3\u573a\u666f\u4e2d\uff0cZeRO \u8bba\u6587\u63d0\u51fa\u4e86\u4e00\u4e2a\u7528\u4e8e\u4f30\u7b97\u6a21\u578b\u72b6\u6001\uff08model states\uff09\u7684\u6807\u51c6\u516c\u5f0f\uff1a(p + p + 12) * model_size<\/code>\u3002\u5176\u4e2d\uff0cp<\/code> \u662f\u6bcf\u4e2a\u53c2\u6570\u7684\u5b57\u8282\u6570\uff08FP16 \u4e3a 2\uff0cFP32 \u4e3a 4\uff09\uff0cmodel_size<\/code> \u662f\u4ee5\u5341\u4ebf\u53c2\u6570\u4e3a\u5355\u4f4d\u7684\u6a21\u578b\u89c4\u6a21\u3002\u4e00\u4e2a 100 \u4ebf\u53c2\u6570\u7684\u6a21\u578b\u5728\u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\u4e0b\uff0c\u9700\u8981\u7ea6 16 \u5b57\u8282\/\u53c2\u6570\uff0c\u603b\u5171\u7ea6 160 GB \u663e\u5b58\u3002<\/p>\n \u4f46\u8fd9\u4e2a\u516c\u5f0f\u9002\u7528\u4e8e\u201c\u8bad\u7ec3\u201d\uff0c\u4e0d\u9002\u7528\u4e8e\u201c\u63a8\u7406\u201d\u3002\u5728\u63a8\u7406\u9636\u6bb5\uff0c\u4f60\u4e0d\u518d\u9700\u8981\u4f18\u5316\u5668\u72b6\u6001\uff0c\u53ea\u9700\u8981\u6743\u91cd\u672c\u8eab\u3002<\/p>\n \u8003\u8651\u5f00\u9500\u4e0e CUDA \u4e0a\u4e0b\u6587<\/h3>\n \u4ec5\u4ec5\u8ba1\u7b97\u6743\u91cd\u8fd8\u4e0d\u591f\u51c6\u786e\u3002\u4f60\u7684 GPU \u8fd8\u9700\u8981\u4e3a CUDA \u4e0a\u4e0b\u6587\u3001cuDNN \u4e0e cuBLAS \u5de5\u4f5c\u7a7a\u95f4\u4ee5\u53ca kernel \u542f\u52a8\u5f00\u9500\u9884\u7559\u663e\u5b58\uff0c\u8fd9\u4e9b\u5728\u6a21\u578b\u5c1a\u672a\u771f\u6b63\u6267\u884c\u524d\u5c31\u5df2\u7ecf\u5360\u7528\u7a7a\u95f4\u3002<\/p>\n \u63a8\u7406\u8fd0\u884c\u65f6\u548c\u670d\u52a1\u6846\u67b6\uff08\u5982 vLLM\u3001TensorRT-LLM\uff09\u5f80\u5f80\u4f1a\u5728\u6743\u91cd\u4e0e KV Cache \u4e4b\u5916\uff0c\u989d\u5916\u5360\u7528\u7ea6 0.5\u20132 GB \u7684 CUDA \u4e0a\u4e0b\u6587\u4e0e\u6846\u67b6\u5f00\u9500\u3002\u8fd9\u5728 A100\u3001A6000 \u7b49\u70ed\u95e8 GPU \u4e0a\u90fd\u6bd4\u8f83\u5e38\u89c1\uff0c\u5177\u4f53\u6570\u503c\u53d6\u51b3\u4e8e\u6846\u67b6\u4e0e GPU \u914d\u7f6e\u3002<\/p>\n \u5b9e\u8df5\u4e2d\uff0c\u53ef\u4ee5\u5728\u6743\u91cd\u663e\u5b58\u57fa\u7840\u4e0a\u989d\u5916\u52a0\u4e0a 1\u20132 GB \u4f5c\u4e3a\u5b89\u5168\u7f13\u51b2\u3002\u4ee5 7B FP16 \u6a21\u578b\u4e3a\u4f8b\uff0c\u6574\u4f53\u89c4\u5212\u7ea6 15\u201316 GB \u663e\u5b58\u6bd4\u8f83\u7a33\u59a5\u3002\u8fd9\u610f\u5473\u7740\u5355\u5361 24 GB\uff08\u5982 RTX 4090\uff09\u5c31\u5df2\u7ecf\u6709\u76f8\u5f53\u4f59\u91cf\u5e94\u5bf9\u63a8\u7406\u3002<\/p>\n \u5728\u9009\u62e9 GPU \u670d\u52a1\u5668\u663e\u5b58\u65f6\uff0c\u8981\u8bb0\u4f4f\uff1a\u4f18\u5316\u5668\u72b6\u6001\u53ea\u8ddf\u8bad\u7ec3 \/ \u5fae\u8c03\u76f8\u5173\uff0c\u5bf9\u7eaf\u63a8\u7406\u4e0d\u91cd\u8981\u3002\u4e0a\u7ebf\u90e8\u7f72\u65f6\uff0c\u4f60\u53ea\u9700\u5173\u6ce8\u201c\u6743\u91cd + \u5f00\u9500\u201d\u7684\u7ec4\u5408\u5373\u53ef\uff0c\u8fd9\u6837\u65e2\u80fd\u4fdd\u8bc1\u4f30\u7b97\u8db3\u591f\u51c6\u786e\uff0c\u53c8\u4e0d\u81f3\u4e8e\u9884\u7b97\u5931\u63a7\u3002<\/p>\n \u6279\u5927\u5c0f\u4e0e\u6fc0\u6d3b\u5185\u5b58<\/strong><\/h2>\n \u6fc0\u6d3b\u5982\u4f55\u968f\u6279\u5927\u5c0f\u6269\u5f20<\/h3>\n \u6fc0\u6d3b\u5185\u5b58\u7528\u6765\u5b58\u50a8\u524d\u5411\u4f20\u64ad\u4e2d\u7684\u4e2d\u95f4\u5f20\u91cf\u3002\u4e0e\u6743\u91cd\u4e0d\u540c\uff0c\u8fd9\u4e9b\u5f20\u91cf\u4f1a\u968f\u7740\u8f93\u5165\u53d8\u5316\u800c\u53d8\u5316\uff0c\u5e76\u4e14\u968f\u6279\u5927\u5c0f\u7ebf\u6027\u589e\u957f\u3002\u53ef\u4ee5\u7528\u4e00\u4e2a\u7b80\u5355\u5173\u7cfb\u5f0f\u8868\u8fbe\uff1a\u5355\u6837\u672c\u6fc0\u6d3b\u5185\u5b58 \u00d7 \u6279\u5927\u5c0f = \u603b\u6fc0\u6d3b\u5185\u5b58\u3002<\/p>\n \u4ee5\u4e00\u4e2a\u5177\u4f53\u4f8b\u5b50\u8bf4\u660e\uff1a\u5047\u8bbe\u67d0\u4e00\u5c42\u540c\u65f6\u5904\u7406 32 \u5f20 224\u00d7224 \u5206\u8fa8\u7387\u300164 \u901a\u9053\u7684\u56fe\u7247\uff0c\u9700\u8981\u5b58\u50a8 224 \u00d7 224 \u00d7 64 \u00d7 32 \u4e2a\u6570\u503c\uff0c\u5373 102,760,448 \u4e2a\u5143\u7d20\uff0c\u82e5\u4f7f\u7528 4 \u5b57\u8282\u6d6e\u70b9\u6570\uff0c\u603b\u5171\u7ea6 401.40 MB\u3002\u82e5\u5c06\u6279\u5927\u5c0f\u4ece 32 \u7ffb\u500d\u5230 64\uff0c\u5185\u5b58\u9700\u6c42\u4e5f\u7ffb\u500d\u5230 802.80 MB\u3002\u8fd9\u79cd\u4e25\u683c\u7684\u7ebf\u6027\u5173\u7cfb\u6e90\u4e8e\u6279\u5927\u5c0f\u53ea\u4f5c\u4e3a\u4e00\u4e2a\u7b80\u5355\u7684\u4e58\u6570\u3002<\/p>\n \u4e00\u4e2a\u5e38\u7528\u7684 LLM \u6fc0\u6d3b\u4f30\u7b97\u516c\u5f0f\u4e3a\uff1aActivation Memory = 2 \u00d7 (sequence length) \u00d7 (batch size) \u00d7 (hidden size) \u00d7 (number of layers)<\/code><\/p>\n \u8fd9\u4e2a\u516c\u5f0f\u63ed\u793a\u4e86\u540c\u6837\u7684\u6a21\u5f0f\uff1a\u53ef\u4ee5\u628a\u6279\u5927\u5c0f\u56e0\u5b50\u63d0\u51fa\u6765\uff0c\u770b\u4f5c\u201c\u5355\u6837\u672c\u6fc0\u6d3b\u5185\u5b58 \u00d7 \u6279\u5927\u5c0f\u201d\u3002<\/p>\n \u5e8f\u5217\u957f\u5ea6\u5219\u4f1a\u8ba9\u60c5\u51b5\u53d8\u5f97\u66f4\u590d\u6742\uff1a\u6ce8\u610f\u529b\u77e9\u9635\u7684\u5927\u5c0f\u4e0e\u5e8f\u5217\u957f\u5ea6\u7684\u5e73\u65b9\u6210\u6b63\u6bd4\u3002\u5e8f\u5217\u957f\u5ea6\u7ffb\u500d\uff0c\u663e\u5b58\u9700\u6c42\u4f1a\u53d8\u4e3a\u539f\u6765\u7684 4 \u500d\u3002\u4e00\u4e2a 2048 token \u7684\u5e8f\u5217\uff0c\u5176\u6ce8\u610f\u529b\u77e9\u9635\u5305\u542b 536,870,912 \u4e2a\u5143\u7d20\uff0c\u5728 float32 \u4e0b\u7ea6\u5360\u7528 2.0 GB\uff1b\u5982\u679c\u6269\u5c55\u5230 4096 token\uff0c\u5143\u7d20\u53d8\u4e3a 2,147,483,648 \u4e2a\uff0c\u663e\u5b58\u8dc3\u5347\u5230 8.0 GB\u3002\u8fd9\u79cd\u4e8c\u6b21\u65b9\u589e\u957f\u6b63\u662f\u957f\u4e0a\u4e0b\u6587\u63a8\u7406\u6781\u6613\u201c\u7206\u663e\u5b58\u201d\u7684\u539f\u56e0\u3002<\/p>\n KV Cache \u5728\u663e\u5b58\u4f30\u7b97\u4e2d\u7684\u89d2\u8272<\/h3>\n \u5728 LLM \u63a8\u7406\u4e2d\uff0cKV Cache \u5f80\u5f80\u662f\u5bfc\u81f4\u663e\u5b58\u6ea2\u51fa\u7684\u7f6a\u9b41\u7978\u9996\u3002\u8be5\u7f13\u5b58\u7528\u4e8e\u5b58\u50a8\u6a21\u578b\u5df2\u7ecf\u5904\u7406\u8fc7\u7684\u6bcf\u4e2a token \u7684 key \u548c value \u5f20\u91cf\uff0c\u5176\u89c4\u6a21\u968f\u6279\u5927\u5c0f\u3001\u5e8f\u5217\u957f\u5ea6\u548c\u5c42\u6570\u5171\u540c\u589e\u957f\u3002<\/p>\n \u4f60\u53ef\u4ee5\u7528\u4e00\u4e2a\u7b80\u5355\u8ba1\u7b97\u6765\u4f30\u7b97\u201c\u6bcf token \u6bcf\u5c42\u201d\u7684 KV Cache\uff1a\u7528 2 \u4e58\u4ee5\u6ce8\u610f\u529b\u5934\u6570\uff0c\u518d\u4e58\u4ee5\u6bcf\u4e2a\u5934\u7684\u7ef4\u5ea6\uff0c\u6700\u540e\u4e58\u4ee5\u6bcf\u4e2a\u5143\u7d20\u7684\u5b57\u8282\u6570\u3002\u4ee5\u4e00\u4e2a\u5177\u6709 32 \u4e2a\u6ce8\u610f\u529b\u5934\u3001\u6bcf\u5934\u7ef4\u5ea6\u4e3a 128 \u4e14\u4f7f\u7528 FP16\uff082 \u5b57\u8282\uff09\u7684\u6a21\u578b\u4e3a\u4f8b\uff0c\u6bcf token \u6bcf\u5c42\u7684 KV Cache \u9700\u6c42\u4e3a\uff1a2 \u00d7 32 \u00d7 128 \u00d7 2 = 16,384<\/code> \u5b57\u8282\uff0c\u5373 16 KB\u3002\u518d\u4e58\u4ee5\u5e8f\u5217\u957f\u5ea6\u3001\u6279\u5927\u5c0f\u548c\u5c42\u6570\uff0c\u5c31\u53ef\u4ee5\u5f97\u5230 KV Cache \u7684\u603b\u663e\u5b58\u9700\u6c42\u3002<\/p>\n \u5206\u6790\u5de5\u5177\u53ef\u4ee5\u5e2e\u52a9\u4f60\u9a8c\u8bc1\u8fd9\u4e9b\u4f30\u7b97\u3002vLLM \u7684 profiler \u80fd\u5728\u90e8\u7f72\u524d\u6d4b\u91cf GPU \u663e\u5b58\u4f7f\u7528\u60c5\u51b5\uff0c\u5e76\u5c55\u793a\u5176\u968f\u5e76\u53d1\u53d8\u5316\u7684\u8d8b\u52bf\u3002\u4f8b\u5982\uff0cLlama 8B \u5728 1 \u4e2a\u5e76\u53d1\u8bf7\u6c42\u65f6\u5360\u7528 16 GB\uff0c\u5728 4 \u4e2a\u5e76\u53d1\u8bf7\u6c42\u65f6\u4f1a\u5347\u81f3\u7ea6 23 GB\u3002\u4f60\u4e5f\u53ef\u4ee5\u901a\u8fc7 nvidia-smi \u5b9e\u65f6\u76d1\u63a7\u663e\u5b58\u4f7f\u7528\u60c5\u51b5\uff0c\u901a\u8fc7\u5bf9\u6bd4\u201c\u6574\u4f53 GPU \u663e\u5b58\u201d\u548c\u201c\u521d\u59cb\u5206\u914d\u91cf\u201d\u6765\u63a8\u65ad KV Cache \u7684\u589e\u957f\u3002<\/p>\n \u5e76\u53d1\u63a8\u7406\u7684\u663e\u5b58\u4f30\u7b97<\/strong><\/h2>\n \u5171\u4eab\u6743\u91cd\u4e0e\u53e0\u52a0\u6fc0\u6d3b\u5f00\u9500<\/h3>\n \u5f53\u591a\u4e2a\u7528\u6237\u540c\u65f6\u5411\u6a21\u578b\u53d1\u8d77\u8bf7\u6c42\u65f6\uff0c\u6743\u91cd\u53ea\u4f1a\u5728\u663e\u5b58\u4e2d\u52a0\u8f7d\u4e00\u4efd\uff0c\u5e76\u4e0d\u4f1a\u4e3a\u6bcf\u4e2a\u8bf7\u6c42\u590d\u5236 14 GB \u6743\u91cd\u3002\u771f\u6b63\u4f1a\u968f\u7740\u5e76\u53d1\u6570\u91cf\u589e\u957f\u7684\uff0c\u662f\u6bcf\u4e2a\u8bf7\u6c42\u72ec\u7acb\u5360\u7528\u7684\u6fc0\u6d3b\u5185\u5b58\u548c KV Cache\u3002<\/p>\n \u603b\u663e\u5b58\u7684\u8ba1\u7b97\u53ef\u4ee5\u6982\u62ec\u4e3a\uff1a\u6743\u91cd\u663e\u5b58 +\uff08\u5355\u8bf7\u6c42\u6fc0\u6d3b\u663e\u5b58 \u00d7 \u5e76\u53d1\u8bf7\u6c42\u6570\uff09+ \u6846\u67b6\u4e0e\u4e0a\u4e0b\u6587\u5f00\u9500\u3002\u8fd9\u4e2a\u516c\u5f0f\u80fd\u5e2e\u52a9\u4f60\u76f4\u89c2\u5730\u770b\u5230\u663e\u5361\u9700\u8981\u627f\u62c5\u7684\u538b\u529b\u3002<\/p>\n \u52a8\u6001\u6279\u5904\u7406\uff08Dynamic Batching\uff09\u53ef\u4ee5\u8ba9\u4f60\u66f4\u9ad8\u6548\u5730\u5229\u7528\u663e\u5b58\u3002\u8be5\u6280\u672f\u901a\u8fc7\u5b9e\u65f6\u76d1\u63a7 GPU \u663e\u5b58\u5360\u7528\u60c5\u51b5\uff0c\u81ea\u52a8\u8c03\u6574\u6279\u5927\u5c0f\uff0c\u5728\u907f\u514d OOM \u7684\u540c\u65f6\u5c3d\u53ef\u80fd\u63d0\u5347\u541e\u5410\u91cf\u3002\u53d8\u957f\u6279\u5904\u7406\u4f1a\u5c06\u957f\u5ea6\u76f8\u8fd1\u7684\u5e8f\u5217\u5f52\u4e3a\u4e00\u7ec4\uff0c\u4ece\u800c\u51cf\u5c11 padding \u5e26\u6765\u7684\u6d6a\u8d39\u3002\u8fd9\u4e9b\u65b9\u6cd5\u6839\u636e\u5f53\u524d\u663e\u5b58\u72b6\u51b5\u81ea\u9002\u5e94\u8c03\u5ea6\uff0c\u907f\u514d\u65e0\u8c13\u7684\u663e\u5b58\u5360\u7528\u3002<\/p>\n \u4f60\u8fd8\u9700\u8981\u7406\u89e3\u663e\u5b58\u5728\u4e0d\u540c\u7ec4\u4ef6\u4e4b\u95f4\u7684\u5206\u914d\u65b9\u5f0f\u3002\u4f8b\u5982\uff0c\u5728 vLLM \u4e2d\uff0c\u53ef\u89c1\u663e\u5b58\u901a\u5e38\u7b49\u4e8e\u603b\u663e\u5b58\u4e58\u4ee5 gpu_memory_utilization<\/code> \u53c2\u6570\u503c\uff1b\u5728\u8fd9\u90e8\u5206\u53ef\u7528\u663e\u5b58\u4e2d\uff0c\u518d\u4f9d\u6b21\u51cf\u53bb\u6a21\u578b\u6743\u91cd\u3001\u5cf0\u503c\u6fc0\u6d3b\u5185\u5b58\u548c\u6846\u67b6\u5f00\u9500\uff0c\u5269\u4f59\u90e8\u5206\u624d\u771f\u6b63\u7528\u4e8e KV Cache\u3002\u8fd9\u4e2a\u5206\u533a\u601d\u8def\u89e3\u91ca\u4e86\u4e3a\u4ec0\u4e48\u201c\u5e76\u4e0d\u662f\u6240\u6709 GPU \u663e\u5b58\u90fd\u80fd\u7528\u6765\u8dd1\u63a8\u7406\u201d\u3002<\/p>\n \u5b9e\u7528\u793a\u4f8b\u4e0e\u5b8c\u6574\u8ba1\u7b97\u6f14\u7ec3<\/h3>\n \u4e0b\u9762\u7528\u4e00\u4e2a\u5177\u4f53\u573a\u666f\u6f14\u793a\u8ba1\u7b97\u8fc7\u7a0b\uff1a\u5047\u8bbe\u4f60\u6709\u4e00\u4e2a 130 \u4ebf\u53c2\u6570\uff0813B\uff09\u7684\u6a21\u578b\uff0c\u4f7f\u7528 FP16\uff0c\u6743\u91cd\u5927\u7ea6\u9700\u8981 26 GB \u663e\u5b58\u3002\u6bcf\u4e2a\u5e76\u53d1\u8bf7\u6c42\uff08\u542b KV Cache\uff09\u5927\u7ea6\u9700\u8981 2 GB \u6fc0\u6d3b\u5185\u5b58\uff0c\u4f60\u9884\u671f\u5e76\u53d1\u6570\u4e3a 10\u3002<\/p>\n \u90a3\u4e48\u4f60\u7684\u663e\u5b58\u4f30\u7b97\u4e3a\uff1a26 GB \u6743\u91cd + 2 GB \u00d7 10 \u4e2a\u5e76\u53d1\u8bf7\u6c42 = 20 GB\uff0c\u518d\u52a0\u4e0a\u7ea6 1 GB \u6846\u67b6\u4e0e\u4e0a\u4e0b\u6587\u5f00\u9500\uff0c\u603b\u8ba1\u7ea6 47 GB\u3002\u7531\u6b64\u53ef\u4ee5\u63a8\u65ad\uff0c\u4f60\u81f3\u5c11\u9700\u8981\u4e00\u5757 48 GB \u663e\u5b58\u7684 GPU\uff08\u5982 A6000\uff09\uff0c\u6216\u8005\u8003\u8651\u5c06\u6a21\u578b\u5212\u5206\u5230\u591a\u5757 GPU \u4e0a\u3002<\/p>\n \u63a5\u4e0b\u6765\u8981\u601d\u8003\u6279\u5927\u5c0f\u4e0e\u5e76\u53d1\u4e4b\u95f4\u7684\u6743\u8861\uff1a\u8f83\u5927\u7684\u6279\u5927\u5c0f\u4f1a\u5360\u7528\u5927\u91cf GPU \u663e\u5b58\uff0c\u5c24\u5176\u662f KV Cache\uff0c\u4f46\u541e\u5410\u91cf\u672a\u5fc5\u6210\u6bd4\u4f8b\u63d0\u5347\uff1b\u540c\u65f6\uff0c\u8fd8\u53ef\u80fd\u56e0\u4e3a DRAM \u5e26\u5bbd\u9971\u548c\u800c\u663e\u8457\u62c9\u9ad8\u5ef6\u8fdf\u3002\u4e00\u79cd\u57fa\u4e8e\u5206\u6790\u7684\u8c03\u4f18\u65b9\u5f0f\u2014\u2014Batching Configuration Advisor\u2014\u2014\u53ef\u4ee5\u4e3a\u4f60\u627e\u51fa\u4e00\u4e2a\u5728\u541e\u5410\u91cf\u4e0e\u5ef6\u8fdf\u4e4b\u95f4\u66f4\u5e73\u8861\u7684\u6700\u4f18\u6279\u5927\u5c0f\u3002\u901a\u8fc7\u91c7\u7528\u8fd9\u4e00\u6700\u4f18\u6279\u5927\u5c0f\uff0c\u4f60\u53ef\u4ee5\u91ca\u653e\u51fa\u4e00\u90e8\u5206\u663e\u5b58\uff0c\u5728\u540c\u4e00\u5757 GPU \u4e0a\u8fd0\u884c\u591a\u4e2a\u6a21\u578b\u526f\u672c\u3002\u5bf9 OPT-1.3B \u6765\u8bf4\uff0c\u5f00\u542f 4 \u4e2a\u526f\u672c\u3001\u6bcf\u4e2a\u526f\u672c\u4f7f\u7528\u4f18\u5316\u8fc7\u7684\u663e\u5b58\u914d\u7f6e\u65f6\uff0c\u603b\u541e\u5410\u91cf\u76f8\u6bd4\u201c\u5355\u526f\u672c + \u6700\u5927\u663e\u5b58\u5360\u7528\u201d\u65b9\u6848\u53ef\u4ee5\u63d0\u5347\u7ea6 33.7%\u3002<\/p>\n \u8f93\u51fa\u957f\u5ea6\u5bf9\u663e\u5b58\u7684\u654f\u611f\u6027\u540c\u6837\u4e0d\u5bb9\u5ffd\u89c6\uff1a\u4ee5 OPT-1.3B \u4e3a\u4f8b\uff0c\u5f53\u6279\u91cf\u4e3a 520 \u4e2a\u8bf7\u6c42\u3001\u6bcf\u4e2a\u8bf7\u6c42\u53ea\u751f\u6210 130 \u4e2a\u8f93\u51fa token \u65f6\uff0cKV Cache \u53ea\u4f7f\u7528\u4e86\u7ea6 20% \u7684\u5bb9\u91cf\uff1b\u4f46\u82e5\u6bcf\u4e2a\u8bf7\u6c42\u751f\u6210 520 \u4e2a\u8f93\u51fa token\uff0c\u5219 KV Cache \u4f7f\u7528\u7387\u4f1a\u98d9\u5347\u81f3 80% \u4ee5\u4e0a\u3002\u5728\u751f\u6210\u5f02\u5e38\u957f\u7684\u8f93\u51fa\u65f6\uff0c\u5e76\u53d1\u5e26\u6765\u7684\u6536\u76ca\u4f1a\u660e\u663e\u9012\u51cf\u3002<\/p>\n \u5982\u679c\u5355\u5361\u663e\u5b58\u59cb\u7ec8\u4e0d\u591f\u7528\uff0c\u53ef\u4ee5\u8003\u8651\u663e\u5b58\u5207\u5206\u65b9\u6848\u3002\u6d41\u6c34\u7ebf\u5e76\u884c\uff08pipeline parallelism\uff09\u5c06\u6a21\u578b\u6309\u5c42\u5782\u76f4\u5207\u5206\u5230\u591a\u5757 GPU \u4e0a\uff0c\u4f8b\u5982 4 \u8def\u5207\u5206\u53ef\u4ee5\u5c06\u5355\u5361\u6743\u91cd\u663e\u5b58\u964d\u4f4e\u5230\u539f\u6765\u7684\u56db\u5206\u4e4b\u4e00\u3002\u5f20\u91cf\u5e76\u884c\uff08tensor parallelism\uff09\u5219\u5728\u6c34\u5e73\u65b9\u5411\u4e0a\u5207\u5206\u6bcf\u4e00\u5c42\uff0c\u5c06\u6743\u91cd\u4e0e\u6fc0\u6d3b\u5206\u522b\u644a\u5230\u591a\u5757 GPU \u4e0a\u3002\u8fd9\u4e24\u79cd\u65b9\u5f0f\u90fd\u53ef\u4ee5\u628a\u603b\u4f53\u5185\u5b58\u8d1f\u8f7d\u5206\u6563\u5230\u591a\u5361\uff0c\u4ece\u800c\u5728\u6bcf\u5757 GPU \u4e0a\u4e3a\u66f4\u5927\u7684\u6279\u91cf\u6216\u66f4\u591a\u5e76\u53d1\u8bf7\u6c42\u817e\u51fa\u7a7a\u95f4\u3002<\/p>\n \u5728\u9009\u62e9 GPU \u670d\u52a1\u5668\u663e\u5b58\u65f6\uff0c\u8fd8\u8981\u8bb0\u4f4f\u603b GPU \u6570\u91cf\u4f1a\u968f\u5e76\u53d1\u548c\u6a21\u578b\u5e76\u884c\u89c4\u6a21\u589e\u957f\u3002\u4e00\u4e2a\u5e38\u7528\u7684\u4f30\u7b97\u516c\u5f0f\u662f\uff1aTotal GPUs = concurrency \u00d7 (tensor_parallel_size \u00d7 pipeline_parallel_size)<\/code>\u3002\u8fd9\u4e2a\u5173\u7cfb\u5f0f\u80fd\u5e2e\u52a9\u4f60\u4ece\u6574\u4f53\u89c6\u89d2\u89c4\u5212 GPU \u6570\u91cf\u4e0e\u663e\u5b58\u9700\u6c42\u3002\u4ee5\u8fd9\u4e9b\u8ba1\u7b97\u4e3a\u8d77\u70b9\uff0c\u518d\u7ed3\u5408\u5b9e\u9645\u8d1f\u8f7d\u6d4b\u8bd5\uff0c\u4f60\u5c31\u80fd\u5728\u907f\u514d\u663e\u5b58\u6ea2\u51fa\u7684\u540c\u65f6\uff0c\u6700\u5927\u9650\u5ea6\u964d\u4f4e\u9884\u7b97\u6d6a\u8d39\u3002<\/p>\n \n <\/div>\n<\/div>\n \u73b0\u5728\uff0c\u4f60\u5df2\u7ecf\u638c\u63e1\u4e86\u4e00\u4e2a\u6e05\u6670\u7684\u4e09\u6b65\u65b9\u6cd5\uff1a\u7b2c\u4e00\u6b65\uff0c\u6839\u636e\u53c2\u6570\u91cf\u548c\u7cbe\u5ea6\u8ba1\u7b97\u6743\u91cd\u663e\u5b58\uff1b\u7b2c\u4e8c\u6b65\uff0c\u4f30\u7b97\u5355\u6837\u672c\u6fc0\u6d3b\u5185\u5b58\uff08\u5305\u542b KV Cache\uff09\uff1b\u7b2c\u4e09\u6b65\uff0c\u5c06\u5176\u4e58\u4ee5\u9884\u671f\u5e76\u53d1\u6570\uff0c\u5e76\u52a0\u4e0a\u6846\u67b6\u4e0e CUDA \u4e0a\u4e0b\u6587\u7684\u5f00\u9500\u3002<\/p>\n \u8fd9\u4e2a\u65b9\u6cd5\u9700\u8981\u7ed3\u5408\u8fed\u4ee3\u8c03\u4f18\u3002\u901a\u8fc7 vLLM \u7684 profiler \u6216 nvidia-smi \u7b49\u5de5\u5177\uff0c\u4f60\u53ef\u4ee5\u5c06\u7406\u8bba\u4f30\u7b97\u4e0e\u771f\u5b9e\u5de5\u4f5c\u8d1f\u8f7d\u8fdb\u884c\u5bf9\u6bd4\u9a8c\u8bc1\u3002\u5148\u5728\u5c0f\u663e\u5b58 GPU \u4e0a\u505a\u538b\u529b\u6d4b\u8bd5\uff0c\u518d\u51b3\u5b9a\u6700\u7ec8\u7684\u786c\u4ef6\u65b9\u6848\u3002\u7ebf\u4e0a GPU \u663e\u5b58\u8ba1\u7b97\u5668\u4e5f\u80fd\u4e3a\u4f60\u63d0\u4f9b\u4e00\u4e2a\u5feb\u901f\u7684\u8d77\u70b9\u3002<\/p>\n \u4ee5\u8fd9\u4e9b\u516c\u5f0f\u4e3a\u57fa\u7840\uff0c\u914d\u5408\u4f60\u7684\u771f\u5b9e\u4e1a\u52a1\u8d1f\u8f7d\u8fdb\u884c\u6d4b\u8bd5\uff0c\u4f60\u5c31\u80fd\u540c\u65f6\u907f\u514d\u663e\u5b58\u4e0d\u8db3\u548c\u9884\u7b97\u6d6a\u8d39\u3002\u8bb0\u4f4f\uff1a\u9009\u62e9 GPU \u670d\u52a1\u5668\u663e\u5b58\u8981\u57fa\u4e8e\u6d4b\u91cf\u548c\u6570\u636e\uff0c\u800c\u4e0d\u662f\u62cd\u8111\u888b\u3002\u52a8\u6001\u6279\u5904\u7406\u4e0e\u5bf9 KV Cache \u7684\u7cbe\u7ec6\u7ba1\u7406\uff0c\u5c06\u8ba9\u4f60\u7684\u63a8\u7406\u90e8\u7f72\u66f4\u9ad8\u6548\u3001\u66f4\u7a33\u5065\u3002<\/p>\n \u5e38\u89c1\u95ee\u9898\uff08FAQ\uff09<\/strong><\/h2>\n 7B FP16 \u6a21\u578b\u7684\u6700\u5c0f\u663e\u5b58\u9700\u6c42\u662f\u591a\u5c11\uff1f<\/h3>\n \u901a\u5e38\u9700\u8981\u81f3\u5c11 15\u201316 GB\u3002\u6a21\u578b\u6743\u91cd\u672c\u8eab\u9700\u8981 14 GB\uff0c\u518d\u989d\u5916\u9884\u7559 1\u20132 GB \u7528\u4e8e CUDA \u4e0a\u4e0b\u6587\u548c\u63a8\u7406\u6846\u67b6\u5f00\u9500\u3002\u4e00\u5757 24 GB \u663e\u5b58\u7684 GPU\uff08\u5982 RTX 4090\uff09\u5728\u63a8\u7406\u573a\u666f\u4e0b\u4f1a\u6bd4\u8f83\u5bbd\u88d5\u3002<\/p>\n \u4e3a\u4ec0\u4e48\u957f\u4e0a\u4e0b\u6587\u5bb9\u6613\u5bfc\u81f4\u663e\u5b58\u6ea2\u51fa\uff1f<\/h3>\n \u56e0\u4e3a KV Cache \u548c\u6ce8\u610f\u529b\u77e9\u9635\u968f\u5e8f\u5217\u957f\u5ea6\u5448\u4e8c\u6b21\u65b9\u589e\u957f\u3002\u5c06\u5e8f\u5217\u957f\u5ea6\u4ece\u539f\u6765\u7684\u57fa\u7840\u7ffb\u500d\uff0c\u6ce8\u610f\u529b\u76f8\u5173\u7684\u663e\u5b58\u9700\u6c42\u5c31\u4f1a\u53d8\u6210 4 \u500d\u3002\u4f8b\u5982\uff0c4096 token \u7684\u5e8f\u5217\u5728 float32 \u4e0b\u7ea6\u5360\u7528 8 GB\uff0c\u800c 2048 token \u53ea\u9700\u8981\u7ea6 2 GB\u3002<\/p>\n \u6211\u80fd\u5728\u4e00\u5757 GPU \u4e0a\u670d\u52a1\u591a\u4e2a\u7528\u6237\u5417\uff1f<\/h3>\n \u53ef\u4ee5\u3002\u591a\u7528\u6237\u4f1a\u5171\u4eab\u4e00\u4efd\u6a21\u578b\u6743\u91cd\uff0c\u4f46\u6bcf\u4e2a\u7528\u6237\u90fd\u6709\u81ea\u5df1\u72ec\u7acb\u7684\u6fc0\u6d3b\u548c KV Cache\u3002\u4f60\u53ef\u4ee5\u7528\u4ee5\u4e0b\u65b9\u5f0f\u4f30\u7b97\u663e\u5b58\u9700\u6c42\uff1a\u6743\u91cd\u663e\u5b58 +\uff08\u5355\u7528\u6237\u6fc0\u6d3b\u663e\u5b58 \u00d7 \u7528\u6237\u6570\u91cf\uff09+ \u6846\u67b6\u4e0e\u4e0a\u4e0b\u6587\u5f00\u9500\u3002<\/p>\n \u4e0a\u7ebf\u90e8\u7f72\u65f6\u662f\u5426\u5e94\u8be5\u4f7f\u7528\u91cf\u5316\uff1f<\/h3>\n \u91cf\u5316\u53ef\u4ee5\u663e\u8457\u964d\u4f4e\u663e\u5b58\u5360\u7528\u5e76\u52a0\u901f\u63a8\u7406\u3002INT8 \u6bcf\u4e2a\u53c2\u6570\u53ea\u7528 1 \u5b57\u8282\uff0c\u800c FP16 \u9700\u8981 2 \u5b57\u8282\u3002\u4ee5 7B \u6a21\u578b\u4e3a\u4f8b\uff0c\u4ece FP16 \u5207\u6362\u5230 INT8\uff0c\u6743\u91cd\u663e\u5b58\u4f1a\u4ece 14 GB \u964d\u5230 7 GB\uff0c\u4ece\u800c\u5728\u540c\u7b49\u663e\u5b58\u4e0b\u5bb9\u7eb3\u66f4\u5927\u7684\u6a21\u578b\u6216\u66f4\u591a\u5e76\u53d1\u3002<\/p>\n","protected":false},"excerpt":{"rendered":" \u4f60\u5728\u7f8e\u56fd\u6216\u9999\u6e2f\u670d\u52a1\u5668\u4e0a\u90e8\u7f72\u5927\u8bed\u8a00\u6a21\u578b\u65f6\uff0c\u63a8\u7406\u8fc7\u7a0b\u4e2d\u906d\u9047\u663e\u5b58\u4e0d\u8db3\uff08out-of-memory\uff09\u9519\u8bef\u3002\u4e0a\u4e0b\u6587\u7a97\u53e3\u8d8a\u957f\uff0c\u95ee\u9898\u8d8a\u4e25\u91cd\u3002\u4f60\u4f1a\u7591\u60d1\uff1a\u201c\u9488\u5bf9\u5f53\u524d\u6a21\u578b\u89c4\u6a21\u3001\u6279\u5927\u5c0f\u548c\u5e76\u53d1\u8bf7\u6c42\u6570\uff0c\u6211\u5230\u5e95\u9700\u8981\u591a\u5c11 GPU \u663e\u5b58\uff1f\u201d \u7b54\u6848\u5728\u4e8e\u7efc\u5408\u8ba1\u7b97\u6a21\u578b\u6743\u91cd\u3001\u6fc0\u6d3b\u5185\u5b58\u4ee5\u53ca\u5e76\u53d1\u5e26\u6765\u7684\u989d\u5916\u5f00\u9500\u3002\u4ee5 70 \u4ebf\u53c2\u6570\u6a21\u578b\uff087B\uff09\u4e3a\u4f8b\uff0c\u4f7f\u7528 FP16 \u7cbe\u5ea6\uff0c\u4ec5\u6743\u91cd\u5c31\u9700\u8981 14 GB \u663e\u5b58\u2014\u2014\u8fd9\u662f 70 \u4ebf\u4e2a\u53c2\u6570\u4e58\u4ee5\u6bcf\u4e2a\u53c2\u6570 2 \u5b57\u8282\u5f97\u5230\u7684\u7ed3\u679c\u3002 \u672c\u6307\u5357\u5c06\u5e26\u4f60\u6309\u6b65\u9aa4\u5b8c\u6210\u663e\u5b58\u4f30\u7b97\u3002\u4f60\u4f1a\u4e86\u89e3\u6a21\u578b\u52a0\u8f7d\u65b9\u5f0f\u3001\u6279\u5927\u5c0f\u5bf9\u663e\u5b58\u7684\u5f71\u54cd\uff0c\u4ee5\u53ca\u5e76\u53d1\u63a8\u7406\u5982\u4f55\u653e\u5927\u5185\u5b58\u9700\u6c42\u3002\u8fd9\u6837\uff0c\u4f60\u5c31\u80fd\u5728\u4e0d\u8d85\u652f\u3001\u4e5f\u4e0d\u8fc7\u5ea6\u538b\u7f29\u914d\u7f6e\u7684\u524d\u63d0\u4e0b\uff0c\u4e3a\u7f8e\u56fd\u6216\u9999\u6e2f\u670d\u52a1\u5668\u9009\u62e9\u5408\u9002\u7684 GPU \u663e\u5b58\u5bb9\u91cf\u3002\u5168\u6587\u805a\u7126\u5b9e\u7528\u65b9\u6848\uff0c\u800c\u4e0d\u662f\u7eaf\u7406\u8bba\u63a8\u6f14\u3002 \u6838\u5fc3\u8981\u70b9 \u901a\u8fc7\u201c\u53c2\u6570\u91cf \u00d7 \u6bcf\u4e2a\u53c2\u6570\u7684\u5b57\u8282\u6570\u201d\u6765\u8ba1\u7b97\u6743\u91cd\u663e\u5b58\u3002\u4e00\u4e2a 7B \u6a21\u578b\u5728 FP16 \u4e0b\u9700\u8981 14 GB \u663e\u5b58\u3002 \u989d\u5916\u9884\u7559 1\u20132 GB \u7528\u4e8e CUDA \u4e0a\u4e0b\u6587\u548c\u6846\u67b6\u5f00\u9500\uff0c\u4f5c\u4e3a\u5b89\u5168\u7684\u663e\u5b58\u57fa\u7ebf\u3002 \u6fc0\u6d3b\u5185\u5b58\u968f\u6279\u5927\u5c0f\u7ebf\u6027\u589e\u957f\uff1bKV Cache \u5219\u968f\u5e8f\u5217\u957f\u5ea6\u5448\u4e8c\u6b21\u65b9\uff08\u5e73\u65b9\uff09\u589e\u957f\u3002 \u591a\u5e76\u53d1\u8bf7\u6c42\u5171\u4eab\u540c\u4e00\u4efd\u6743\u91cd\uff0c\u6bcf\u4e2a\u8bf7\u6c42\u53ea\u4f1a\u989d\u5916\u589e\u52a0\u81ea\u5df1\u7684\u6fc0\u6d3b\u548c KV Cache\u3002 \u4f7f\u7528 vLLM \u7b49\u5206\u6790\u5de5\u5177\u6216 nvidia-smi \u5bf9\u4f30\u7b97\u7ed3\u679c\u8fdb\u884c\u9a8c\u8bc1\uff1b\u5148\u5728\u5c0f\u663e\u5b58 GPU \u4e0a\u6d4b\u8bd5\uff0c\u518d\u51b3\u5b9a\u6b63\u5f0f\u91c7\u8d2d\u65b9\u6848\u3002 \u5982\u4f55\u4e3a\u6a21\u578b\u52a0\u8f7d\u9009\u62e9 GPU \u670d\u52a1\u5668\u663e\u5b58 \u4ece\u53c2\u6570\u91cf\u4f30\u7b97\u6743\u91cd\u663e\u5b58 \u5148\u4ece\u4e00\u4e2a\u7b80\u5355\u7684\u8ba1\u7b97\u5f00\u59cb\uff1a\u6a21\u578b\u6743\u91cd\u663e\u5b58 = \u53c2\u6570\u6570\u91cf […]<\/p>\n Read More…<\/a><\/p>\n","protected":false},"author":2,"featured_media":34341,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[62],"tags":[13017,13018,13019,13020,12701],"class_list":["post-34344","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-hong-kong-dedicated-server-sc","tag-gpu-server-vram","tag-model-loading","tag-batch-size","tag-concurrent-requests","tag-llm-inference"],"acf":[],"yoast_head":"\n