{"id":43491,"date":"2023-09-12T15:41:00","date_gmt":"2023-09-12T07:41:00","guid":{"rendered":"https:\/\/www.ioiotimes.com\/?p=43491"},"modified":"2023-09-12T23:45:15","modified_gmt":"2023-09-12T15:45:15","slug":"nvidia-tensorrt-llm%e5%a2%9e%e5%bc%b7%e4%ba%86nvidia-h100-gpu%e4%b8%8a%e5%a4%a7%e5%9e%8b%e8%aa%9e%e8%a8%80%e6%a8%a1%e5%9e%8b%e7%9a%84%e6%8e%a8%e8%ab%96%e8%83%bd%e5%8a%9b","status":"publish","type":"post","link":"https:\/\/www.ioiotimes.com\/?p=43491","title":{"rendered":"NVIDIA TensorRT-LLM\u589e\u5f37\u4e86NVIDIA H100 GPU\u4e0a\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u7684\u63a8\u8ad6\u80fd\u529b"},"content":{"rendered":"\n

\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u63d0\u4f9b\u6975\u70ba\u51fa\u8272\u7684\u65b0\u529f\u80fd\uff0c\u64f4\u5927\u4eba\u5de5\u667a\u6167\u6f5b\u5728\u7684\u61c9\u7528\u9818\u57df\u3002\u4e0d\u904e\u5176\u9f90\u5927\u898f\u6a21\u8207\u7368\u7279\u7684\u57f7\u884c\u7279\u6027\uff0c\u5f88\u96e3\u7528\u5177\u6210\u672c\u6548\u76ca\u7684\u65b9\u5f0f\u4f86\u4f7f\u7528\u5b83\u5011\u3002<\/p>\n\n\n\n

NVIDIA \u4e0d\u65b7\u8207 Meta<\/a>\u3001AnyScale<\/a>\u3001Cohere<\/a>\u3001Deci<\/a>\u3001Grammarly<\/a>\u3001Mistral AI<\/a>\u3001MosaicML<\/a>\uff08\u73fe\u5df2\u6210\u70ba Databricks<\/a> \u7684\u4e00\u54e1\uff09\u3001OctoML<\/a>\u3001Tabnine<\/a>\u53ca Together AI \u7b49\u91cd\u9ede\u4f01\u696d\u5bc6\u5207\u5408\u4f5c\uff0c\u4ee5\u52a0\u5feb\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u7684\u63a8\u8ad6\u901f\u5ea6\u53ca\u53d6\u5f97\u6700\u4f73\u7684\u63a8\u8ad6\u7d50\u679c\u3002<\/p>\n\n\n\n

\u9019\u4e9b\u5275\u65b0\u9805\u76ee\u5df2\u7d93\u6574\u5408\u9032\u958b\u6e90\u578b\u614b\u7684 NVIDIA TensorRT-LLM<\/a> \u8edf\u9ad4\uff0c\u5c07\u5728\u672a\u4f86\u5e7e\u9031\u5167\u767c\u5e03\u3002TensorRT-LLM \u7531 TensorRT \u6df1\u5ea6\u5b78\u7fd2\u7de8\u8b6f\u5668\u7d44\u6210\uff0c\u5305\u62ec\u7d93\u6700\u4f73\u5316\u8abf\u6574\u7684\u5167\u6838\u3001\u524d\u8655\u7406\u548c\u5f8c\u8655\u7406\u6b65\u9a5f\uff0c\u4ee5\u53ca\u591a GPU\uff0f\u591a\u7bc0\u9ede\u901a\u8a0a\u57fa\u5143\uff0c\u53ef\u5728 NVIDIA GPU \u4e0a\u5275\u9020\u51fa\u7a81\u7834\u6027\u7684\u6548\u80fd\u8868\u73fe\u3002\u5b83\u8b93\u958b\u767c\u4eba\u54e1\u80fd\u5920\u5617\u8a66\u65b0\u7684\u5927\u578b\u8a9e\u8a00\u6a21\u578b\uff0c\u63d0\u4f9b\u5cf0\u503c\u6548\u80fd\u548c\u5feb\u901f\u81ea\u8a02\u529f\u80fd\uff0c\u800c\u7121\u9700\u5177\u5099\u6df1\u539a\u7684 C++ \u6216 NVIDIA CUDA \u76f8\u95dc\u77e5\u8b58\u3002<\/p>\n\n\n\n

<\/p>\n\n\n\n

\"\"
\u25b2NVIDIA TensorRT-LLM<\/figcaption><\/figure>\n\n\n\n

<\/p>\n\n\n\n

TensorRT-LLM \u900f\u904e\u958b\u6e90\u6a21\u7d44 Python API\uff0c\u63d0\u9ad8\u4e86\u6613\u7528\u6027\u548c\u64f4\u5145\u6027\uff0c\u53ef\u7528\u65bc\u5b9a\u7fa9\u3001\u6700\u4f73\u5316\u548c\u57f7\u884c\u65b0\u67b6\u69cb\uff0c\u9084\u6703\u96a8\u8457\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u7684\u767c\u5c55\u800c\u589e\u5f37\uff0c\u4e14\u80fd\u8f15\u9b06\u81ea\u8a02\u76f8\u95dc\u5167\u5bb9\u3002<\/p>\n\n\n\n

\u4f8b\u5982 MosaicML \u5728 TensorRT-LLM \u7684\u57fa\u790e\u4e0a\u7121\u7e2b\u589e\u52a0\u6240\u9700\u7684\u7279\u5b9a\u529f\u80fd\uff0c\u4e26\u5c07\u5176\u6574\u5408\u5230\u63a8\u8ad6\u670d\u52d9\u4e2d\u3002Databricks \u5de5\u7a0b\u90e8\u9580\u526f\u7e3d\u88c1 Naveen Rao \u6307\u51fa\uff1a\u300c\u9019\u7d55\u5c0d\u662f\u4e00\u4ef6\u8f15\u800c\u6613\u8209\u7684\u4e8b\u3002\u300d<\/p>\n\n\n\n

\u300cTensorRT-LLM \u7c21\u55ae\u6613\u7528\uff0c\u63d0\u4f9b\u5305\u62ec\u8a5e\u5143\u4e32\u6d41\u8655\u7406\uff08streaming of tokens\uff09\u3001\u52d5\u614b\u6279\u6b21\u8655\u7406\uff08in-flight batching\uff09\u3001paged-attention\u3001\u91cf\u5316\u7b49\u9f4a\u5168\u7684\u529f\u80fd\uff0c\u800c\u4e14\u6548\u7387\u51fa\u8272\u3002\u5b83\u70ba\u4f7f\u7528 NVIDIA GPU \u7684\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u670d\u52d9\u63d0\u4f9b\u4e86\u6700\u5148\u9032\u7684\u6548\u80fd\uff0c\u8b93\u6211\u5011\u80fd\u5920\u628a\u7701\u4e0b\u4f86\u7684\u6210\u672c\u56de\u994b\u7d66\u5ba2\u6236\u3002\u300dRao \u8aaa\u3002<\/p>\n\n\n\n

<\/p>\n\n\n\n

\u6548\u80fd\u6bd4\u8f03<\/strong><\/h3>\n\n\n\n

\u6458\u8981\u6587\u7ae0\u53ea\u662f\u5927\u578b\u8a9e\u8a00\u6a21\u578b<\/a><\/strong>\u7684\u773e\u591a\u61c9\u7528\u9805\u76ee\u4e4b\u4e00\u3002\u4ee5\u4e0b\u57fa\u6e96\u6e2c\u8a66\u986f\u793a\u4e86 TensorRT-LLM \u5728\u6700\u65b0\u7684 NVIDIA Hopper \u67b6\u69cb\u4e0a\u6240\u63d0\u5347\u7684\u51fa\u8272\u6548\u80fd\u8868\u73fe\u3002<\/p>\n\n\n\n

\u4e0b\u5716\u53cd\u6620\u51fa\u4f7f\u7528 NVIDIA A100 \u548c NVIDIA H100 \u8207 CNN\/Daily Mail \u9032\u884c\u7684\u6587\u7ae0\u6458\u8981\u4f5c\u696d\uff0cCNN\/Daily Mail \u662f\u7528\u65bc\u8a55\u4f30\u6458\u8981\u8868\u73fe\u7684\u8457\u540d\u8cc7\u6599\u96c6\u3002<\/p>\n\n\n\n

\u5716 1 \u4e2d\u5149\u662f H100 \u5c31\u6bd4 A100 \u5feb\u4e0a\u56db\u500d\u3002\u52a0\u5165 TensorRT-LLM \u53ca\u5176\u5305\u62ec\u52d5\u614b\u6279\u6b21\u8655\u7406\u5728\u5167\u7684\u5404\u9805\u512a\u52e2\u5f8c\uff0c\u7e3d\u541e\u5410\u91cf\u63d0\u9ad8\u4e86\u516b\u500d\uff0c\u5be6\u73fe\u6700\u9ad8\u541e\u5410\u91cf\u3002<\/p>\n\n\n

\n
\"\"
\u25b2GPT-J-6B \u6a21\u578b\u5728\u4f7f\u7528 A100 \u8207 \u52a0\u4e0a\u548c\u672a\u52a0\u4e0a TensorRT-LLM \u7684H100\u6548\u80fd\u6bd4\u8f03<\/em>
Text summarization, variable I\/O length, CNN \/ DailyMail dataset | A100 FP16 PyTorch eager mode | H100 FP8 | H100 FP8, in-flight batching, TensorRT-LLM<\/em><\/figcaption><\/figure><\/div>\n\n\n

<\/p>\n\n\n\n

Llama 2 \u662f Meta \u8fd1\u671f\u767c\u5e03\u7684\u4e00\u6b3e\u71b1\u9580\u8a9e\u8a00\u6a21\u578b\uff0c\u53d7\u5230\u60f3\u8981\u63a1\u7528\u751f\u6210\u5f0f\u4eba\u5de5\u667a\u6167\u7684\u4f01\u696d\u5ee3\u6cdb\u4f7f\u7528\uff0c\u800c TensorRT-LLM \u5728 Llama 2 \u6a21\u578b\u4e0a\u9032\u884c\u63a8\u8ad6\u7684\u8868\u73fe\u662f A100 GPU \u7684 4.6 \u500d\u3002<\/p>\n\n\n

\n
\"\"
\u25b2700\u5104\u500b\u53c3\u6578\u7684Llama 2\u6a21\u578b\u5728\u4f7f\u7528A100\u8207\u52a0\u4e0a\u548c\u672a\u52a0\u4e0aTensorRT-LLM \u7684H100\u6548\u80fd\u6bd4\u8f03<\/em>
Text summarization, variable I\/O length, CNN \/ DailyMail dataset | A100 FP16 PyTorch eager mode| H100 FP8 | H100 FP8, in-flight batching, TensorRT-LLM<\/em><\/figcaption><\/figure><\/div>\n\n\n

<\/p>\n\n\n\n

\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u751f\u614b\u7cfb\u7d71\u5448\u73fe\u7206\u70b8\u6027\u767c\u5c55<\/strong><\/h3>\n\n\n\n

\u751f\u614b\u7cfb\u7d71\u6b63\u5728\u5feb\u901f\u5275\u65b0\uff0c\u767c\u5c55\u51fa\u5168\u65b0\u591a\u6a23\u5316\u7684\u6a21\u578b\u67b6\u69cb\u3002\u66f4\u5927\u7684\u6a21\u578b\u91cb\u653e\u51fa\u65b0\u7684\u529f\u80fd\u548c\u4f7f\u7528\u7bc4\u4f8b\u3002\u50cf\u662f Meta \u65d7\u4e0b\u64c1\u6709 700 \u5104\u500b\u53c3\u6578\u7684 Llama 2\uff0c\u9019\u4e00\u985e\u6700\u5927\u3001\u6700\u5148\u9032\u7684\u8a9e\u8a00\u6a21\u578b\u9700\u8981\u591a\u500b GPU \u5354\u540c\u5de5\u4f5c\uff0c\u624d\u80fd\u5373\u6642\u63d0\u4f9b\u56de\u61c9\u3002\u958b\u767c\u4eba\u54e1\u904e\u53bb\u5982\u679c\u60f3\u7372\u5f97\u6700\u4f73\u7684\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u63a8\u8ad6\u8868\u73fe\uff0c\u5c31\u5fc5\u9808\u91cd\u5beb\u548c\u624b\u52d5\u5c07\u4eba\u5de5\u667a\u6167\u6a21\u578b\u5206\u5272\u6210\u591a\u500b\u7247\u6bb5\uff0c\u7136\u5f8c\u8de8\u591a\u500b GPU \u5354\u8abf\u57f7\u884c\u3002<\/p>\n\n\n\n

TensorRT-LLM \u4f7f\u7528\u4e00\u7a2e\u6a21\u578b\u5e73\u884c\u5316\uff08model parallelism\uff09\u7684 tensor \u5e73\u884c<\/a>\uff0c\u5c07\u500b\u5225\u6b0a\u91cd\u77e9\u9663\u5206\u5272\u5230\u5404\u500b\u88dd\u7f6e\u4e0a\u3002\u5982\u6b64\u4e00\u4f86\u4fbf\u80fd\u4ee5\u5927\u898f\u6a21\u9ad8\u6548\u7387\u7684\u65b9\u5f0f\u9032\u884c\u63a8\u8ad6 – \u53ef\u4ee5\u5728\u900f\u904e NVLink \u9023\u63a5\u7684\u591a\u500b GPU \u548c\u591a\u500b\u4f3a\u670d\u5668\u4e0a\u5e73\u884c\u904b\u4f5c\u6bcf\u500b\u6a21\u578b\uff0c\u7121\u9700\u958b\u767c\u4eba\u54e1\u51fa\u624b\u5e72\u9810\u6216\u4fee\u6539\u6a21\u578b\u3002<\/p>\n\n\n\n

\u96a8\u8457\u63a8\u51fa\u65b0\u6a21\u578b\u548c\u6a21\u578b\u67b6\u69cb\uff0c\u958b\u767c\u4eba\u54e1\u53ef\u4ee5\u5229\u7528 TensorRT-LLM \u4e2d\u958b\u6e90\u578b\u614b\u7684\u6700\u65b0 NVIDIA \u4eba\u5de5\u667a\u6167\u5167\u6838\uff0c\u5c07\u5176\u6a21\u578b\u8abf\u6574\u81f3\u6700\u4f73\u72c0\u614b\u3002\u652f\u63f4\u7684\u5167\u6838\u878d\u5408\u5305\u62ec\u5c16\u7aef\u7684FlashAttention \u548c Masked Multi-Head Attention\u6280\u8853\uff0c\u7528\u65bcGPT\u6a21\u578b\u57f7\u884c\u7684\u4e0a\u4e0b\u6587\u548c\u751f\u6210\u968e\u6bb5\uff0c\u4ee5\u53ca\u8a31\u591a\u5176\u4ed6\u529f\u80fd\u3002<\/p>\n\n\n\n

\u6b64\u5916\uff0cTensorRT-LLM \u9084\u5305\u62ec\u7576\u4eca\u751f\u7522\u74b0\u5883\u4e2d\u5df2\u5ee3\u6cdb\u4f7f\u7528\u7684\u8a31\u591a\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u7684\u5b8c\u5168\u6700\u4f73\u5316\u3001\u53ef\u7acb\u5373\u904b\u884c\u7248\u672c\u3002\u5176\u4e2d\u5305\u62ec Meta Llama 2\u3001OpenAI GPT-2 \u548c GPT-3\u3001Falcon\u3001Mosaic MPT\u3001BLOOM \u53ca\u5176\u4ed6\u5341\u5e7e\u7a2e\uff0c\u90fd\u80fd\u900f\u904e\u7c21\u55ae\u6613\u7528\u7684 TensorRT-LLM Python API \u4f86\u64cd\u4f5c\u9019\u4e9b\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u3002<\/p>\n\n\n\n

\u9019\u4e9b\u529f\u80fd\u53ef\u5e6b\u52a9\u958b\u767c\u4eba\u54e1\u66f4\u5feb\u3001\u66f4\u6e96\u78ba\u5730\u958b\u767c\u81ea\u8a02\u7684\u5927\u578b\u8a9e\u8a00\u6a21\u578b\uff0c\u4ee5\u6eff\u8db3\u5e7e\u4e4e\u5404\u884c\u5404\u696d\u7684\u9700\u6c42\u3002<\/p>\n\n\n\n

<\/p>\n\n\n\n

\u52d5\u614b\u6279\u6b21\u8655\u7406<\/strong><\/h3>\n\n\n\n

\u7576\u4eca\u7684\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u6975\u70ba\u591a\u6a23\u5316\u3002\u4e00\u500b\u6a21\u578b\u53ef\u540c\u6642\u7528\u65bc\u8655\u7406\u591a\u7a2e\u5f7c\u6b64\u9593\u5dee\u7570\u6975\u5927\u7684\u4efb\u52d9\u3002\u5f9e\u804a\u5929\u6a5f\u5668\u4eba\u4e2d\u7c21\u55ae\u7684\u4e00\u554f\u4e00\u7b54\uff0c\u5230\u6587\u4ef6\u6458\u8981\u6216\u751f\u6210\u9577\u7bc7\u7a0b\u5f0f\u78bc\uff0c\u8981\u7528\u9ad8\u5ea6\u52d5\u614b\u7684\u65b9\u5f0f\u4f86\u8655\u7406\u9019\u4e9b\u5de5\u4f5c\uff0c\u800c\u8f38\u51fa\u7684\u5927\u5c0f\u5247\u662f\u5929\u5dee\u5730\u9060\u3002<\/p>\n\n\n\n

\u9019\u7a2e\u591a\u529f\u80fd\u6027\u53ef\u80fd\u6703\u5c0e\u81f4\u96e3\u4ee5\u6709\u6548\u5730\u6279\u91cf\u8655\u7406\u8acb\u6c42\u4e26\u884c\u57f7\u884c\u5b83\u5011\uff08\u800c\u9019\u662f\u7528\u65bc\u670d\u52d9\u795e\u7d93\u7db2\u8def\u6642\u5e38\u898b\u7684\u6700\u4f73\u5316\u65b9\u5f0f\uff09\uff0c\u9020\u6210\u67d0\u4e9b\u8acb\u6c42\u6bd4\u5176\u4ed6\u8acb\u6c42\u66f4\u65e9\u5b8c\u6210\u3002<\/p>\n\n\n\n

\u70ba\u4e86\u7ba1\u7406\u9019\u4e9b\u52d5\u614b\u8ca0\u8f09\uff0cTensorRT-LLM \u5305\u542b\u4e86\u4e00\u7a2e\u7a31\u70ba\u52d5\u614b\u6279\u6b21\u8655\u7406\uff08in-flight batching\uff09\u7684\u6700\u4f73\u8abf\u5ea6\u6280\u8853\u3002\u9019\u9805\u6280\u8853\u8b93\u4e00\u500b\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u7684\u6574\u9ad4\u6587\u5b57\u751f\u6210\u904e\u7a0b\u53ef\u62c6\u5206\u6210\u5728\u6a21\u578b\u4e0a\u57f7\u884c\u7684\u591a\u6b21\u8fed\u4ee3\u3002<\/p>\n\n\n\n

TensorRT-LLM\u00a0\u5229\u7528\u52d5\u614b\u6279\u6b21\u8655\u7406\u6280\u8853\uff0c\u4e0d\u6703\u7b49\u5230\u6574\u500b\u6279\u6b21\u8655\u7406\u5b8c\u6210\u5f8c\u518d\u8655\u7406\u4e0b\u4e00\u7d44\u8acb\u6c42\uff0c\u800c\u662f\u6703\u7acb\u5373\u5f9e\u6279\u6b21\u8655\u7406\u4e2d\u79fb\u9664\u5df2\u7d93\u5b8c\u6210\u7684\u5e8f\u5217\u3002\u5728\u5176\u4ed6\u8acb\u6c42\u4ecd\u88ab\u8655\u7406\u4e2d\uff0c\u5b83\u5c31\u958b\u59cb\u57f7\u884c\u65b0\u7684\u8acb\u6c42\u3002\u52d5\u614b\u6279\u6b21\u8655\u7406\u8207\u984d\u5916\u7684\u5167\u6838\u7d1a\u6700\u4f73\u5316\u6280\u8853\u63d0\u9ad8\u4e86\u00a0GPU\u00a0\u7684\u4f7f\u7528\u7387\uff0c\u5c07\u00a0H100 Tensor\u00a0\u6838\u5fc3\u00a0GPU\u00a0\u4e0a\u5be6\u969b\u00a0LLM\u00a0\u8acb\u6c42\u57fa\u6e96\u7684\u541e\u5410\u91cf\u6700\u5c11\u63d0\u9ad8\u4e86\u4e00\u500d\uff0c\u6709\u52a9\u65bc\u6f38\u5c11\u80fd\u8017\u4ee5\u53ca\u5c07\u7e3d\u6301\u6709\u6210\u672c\u964d\u81f3\u6700\u4f4e\u3002<\/p>\n\n\n\n

<\/p>\n\n\n\n

H100 Transformer<\/strong>\u5f15\u64ce\u52a0\u4e0aFP8<\/strong><\/h3>\n\n\n\n

\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u5167\u6709\u6578\u5341\u5104\u500b\u6a21\u578b\u6b0a\u91cd\u548c\u555f\u52d5\u9805\u76ee\uff0c\u901a\u5e38\u4f7f\u7528 16 \u4f4d\u5143\u6d6e\u9ede\uff08FP16 \u6216 BF16\uff09\u503c\u9032\u884c\u8a13\u7df4\u548c\u8868\u793a\uff0c\u6bcf\u500b\u503c\u4f54\u7528 16 \u4f4d\u5143\u8a18\u61b6\u9ad4\u3002\u7136\u800c\uff0c\u5728\u63a8\u8ad6\u904e\u7a0b\u4e2d\uff0c\u5927\u591a\u6578\u6a21\u578b\u53ef\u4ee5\u4f7f\u7528\u73fe\u4ee3\u91cf\u5316\u6280\u8853\u6709\u6548\u5730\u8868\u793a\u70ba\u8f03\u4f4e\u7cbe\u5ea6\uff0c\u59828\u4f4d\u751a\u81f34\u4f4d\u6574\u6578\uff08INT8\u6216INT4\uff09\u3002<\/p>\n\n\n\n

\u91cf\u5316\u662f\u5728\u4e0d\u5931\u6e96\u78ba\u5ea6\u7684\u524d\u63d0\u4e0b\uff0c\u964d\u4f4e\u6a21\u578b\u6b0a\u91cd\u548c\u555f\u52d5\u7cbe\u5ea6\u7684\u904e\u7a0b\u3002\u4f7f\u7528\u66f4\u4f4e\u7cbe\u5ea6\u4ee3\u8868\u6bcf\u500b\u53c3\u6578\u66f4\u5c0f\uff0c\u6a21\u578b\u4f54\u7528 GPU \u8a18\u61b6\u9ad4\u7684\u7a7a\u9593\u4e5f\u66f4\u5c0f\u3002\u9019\u6a23\u5c31\u80fd\u5728\u76f8\u540c\u786c\u9ad4\u689d\u4ef6\u4e0b\u5c0d\u66f4\u5927\u7684\u6a21\u578b\u9032\u884c\u63a8\u8ad6\uff0c\u540c\u6642\u5728\u57f7\u884c\u904e\u7a0b\u4e2d\u6e1b\u5c11\u8a18\u61b6\u9ad4\u904b\u4f5c\u6642\u9593\u3002<\/p>\n\n\n\n

NVIDIA H100<\/a> GPU \u52a0\u4e0a TensorRT-LLM\uff0c\u8b93\u4f7f\u7528\u8005\u80fd\u5920\u8f15\u9b06\u5c07\u6a21\u578b\u6b0a\u91cd\u8f49\u63db\u70ba\u65b0\u7684 FP8 \u683c\u5f0f\uff0c\u4e26\u81ea\u52d5\u7de8\u8b6f\u6a21\u578b\u4ee5\u5229\u7528\u6700\u4f73\u5316\u7684 FP8 \u5167\u6838\u3002\u5728 Hopper Transformer \u5f15\u64ce<\/a> \u6280\u8853\u5be6\u73fe\u7684\uff0c\u7121\u9700\u66f4\u52d5\u4efb\u4f55\u6a21\u578b\u7a0b\u5f0f\u78bc\u3002<\/p>\n\n\n\n

H100\u00a0\u5f15\u5165\u7684\u00a0FP8\u00a0\u8cc7\u6599\u683c\u5f0f\u4f7f\u5f97\u958b\u767c\u4eba\u54e1\u80fd\u5920\u91cf\u5316\u4ed6\u5011\u7684\u6a21\u578b\uff0c\u4e26\u5728\u4e0d\u964d\u4f4e\u6a21\u578b\u6e96\u78ba\u5ea6\u7684\u60c5\u6cc1\u4e0b\u5927\u5e45\u6e1b\u5c11\u8a18\u61b6\u9ad4\u6d88\u8017\u3002\u8207\u00a0INT8\u00a0\u6216\u00a0INT4\u00a0\u7b49\u5176\u4ed6\u8cc7\u6599\u683c\u5f0f\u76f8\u6bd4\uff0cFP8\u00a0\u91cf\u5316\u4fdd\u7559\u4e86\u66f4\u9ad8\u7684\u6e96\u78ba\u5ea6\uff0c\u540c\u6642\u505a\u5230\u6700\u5feb\u7684\u6548\u80fd\u8868\u73fe\u548c\u6700\u7c21\u55ae\u7684\u5be6\u73fe\u3002<\/p>\n\n\n\n

<\/p>\n\n\n\n

\u7d50\u8a9e<\/strong><\/h3>\n\n\n\n

\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u7684\u767c\u5c55\u65e5\u65b0\u6708\u7570\u3002\u6bcf\u5929\u90fd\u958b\u767c\u51fa\u4e0d\u540c\u7684\u6a21\u578b\u67b6\u69cb\uff0c\u4fc3\u9032\u751f\u614b\u7cfb\u7d71\u4e0d\u65b7\u767c\u5c55\u3002\u800c\u5927\u578b\u6a21\u578b\u4e5f\u6703\u91cb\u653e\u51fa\u65b0\u7684\u529f\u80fd\u548c\u4f7f\u7528\u7bc4\u4f8b\uff0c\u63a8\u52d5\u5404\u7522\u696d\u52a0\u4ee5\u63a1\u7528\u3002<\/p>\n\n\n\n

\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u63a8\u8ad6\u6280\u8853\u70ba\u8cc7\u6599\u4e2d\u5fc3\u6253\u958b\u65b0\u7684\u5c40\u9762\u3002\u66f4\u51fa\u8272\u7684\u6548\u80fd\u53ca\u66f4\u9ad8\u7684\u6e96\u78ba\u6027\u70ba\u4f01\u696d\u964d\u4f4e\u4e86\u7e3d\u6301\u6709\u6210\u672c\u3002\u6a21\u578b\u5275\u65b0\u80fd\u5920\u5275\u9020\u51fa\u66f4\u7f8e\u597d\u7684\u5ba2\u6236\u9ad4\u9a57\uff0c\u9032\u800c\u5e36\u4f86\u66f4\u9ad8\u7684\u6536\u5165\u548c\u6536\u76ca\u3002<\/p>\n\n\n\n

\u5728\u898f\u5283\u63a8\u8ad6\u90e8\u7f72\u9805\u76ee\u4e4b\u969b\uff0c\u8981\u5229\u7528\u6700\u5148\u9032\u7684 LLM \u9054\u5230\u6700\u4f73\u6548\u80fd\uff0c\u4ecd\u8981\u8003\u616e\u8a31\u591a\u5176\u4ed6\u56e0\u7d20\u3002\u6700\u4f73\u5316\u9019\u4ef6\u4e8b\u5f88\u5c11\u6703\u81ea\u52d5\u9032\u884c\u3002\u4f7f\u7528\u8005\u5fc5\u9808\u8003\u616e\u5e73\u884c\u3001\u7aef\u5230\u7aef\u5de5\u4f5c\u6d41\u7a0b\u548c\u5148\u9032\u8abf\u5ea6\u6280\u8853\u7b49\u5fae\u8abf\u6a21\u578b\u7684\u56e0\u7d20\u3002\u4ed6\u5011\u9084\u8981\u6709\u4e00\u500b\u80fd\u5920\u8655\u7406\u6df7\u5408\u7cbe\u5ea6\uff0c\u53c8\u4e0d\u6703\u964d\u4f4e\u6e96\u78ba\u5ea6\u7684\u904b\u7b97\u5e73\u53f0\u3002<\/p>\n\n\n\n

TensorRT-LLM \u5305\u62ec TensorRT \u7684\u6df1\u5ea6\u5b78\u7fd2\u7de8\u8b6f\u5668\u3001\u6700\u4f73\u5316\u7684\u5167\u6838\u3001\u524d\u8655\u7406\u548c\u5f8c\u8655\u7406\uff0c\u4ee5\u53ca\u5728\u4e00\u500b\u7c21\u55ae\u7684\u3001\u958b\u6e90\u7684Python API\u4e2d\u5be6\u73fe\u591aGPU\/\u591a\u7bc0\u9ede\u901a\u8a0a\uff0c\u7528\u65bc\u5b9a\u7fa9\u3001\u6700\u4f73\u5316\u548c\u57f7\u884c\u5927\u578b\u8a9e\u8a00\u6a21\u578b\uff0c\u4ee5\u9032\u884c\u751f\u7522\u74b0\u5883\u4e2d\u7684\u63a8\u8ad6\u3002<\/p>\n\n\n\n

<\/p>\n\n\n\n

\u958b\u59cb\u4f7f\u7528TensorRT-LLM<\/strong><\/h3>\n\n\n\n

NVIDIA TensorRT-LLM \u73fe\u5df2\u958b\u653e\u65e9\u671f\u8a66\u7528<\/a>\uff0c\u4e0d\u4e45\u5c07\u6574\u5408\u9032 NVIDIA NeMo \u6846\u67b6 – NeMo\u6846\u67b6\u662f NVIDIA AI Enterprise<\/a> \u7684\u4e00\u90e8\u5206\uff0c\u800c NVIDIA AI Enterprise \u662f\u4e00\u500b\u5177\u6709\u5b89\u5168\u6027\u3001\u7a69\u5b9a\u6027\u3001\u53ef\u7ba1\u7406\u6027\u548c\u652f\u63f4\u6027\u7684\u4f01\u696d\u7d1a\u4eba\u5de5\u667a\u6167\u8edf\u9ad4\u5e73\u53f0\u3002\u958b\u767c\u4eba\u54e1\u8207\u7814\u7a76\u4eba\u54e1\u53ef\u4ee5\u900f\u904e NGC<\/a> \u4e0a\u7684 NeMo \u6846\u67b6\u6216 GitHub \u4e0a\u7684\u6e90\u78bc\u5eab\u53d6\u5f97 TensorRT-LLM\u3002<\/p>\n\n\n\n

\u8acb\u6ce8\u610f\uff0c\u60a8\u5fc5\u9808\u5728 NVIDIA \u958b\u767c\u8005\u8a08\u756b<\/a>\u4e2d\u8a3b\u518a\u624d\u80fd\u7533\u8acb\u65e9\u671f\u8a66\u7528\u3002\u60a8\u9084\u5fc5\u9808\u4f7f\u7528\u8cb4\u7d44\u7e54\u7684\u96fb\u5b50\u90f5\u4ef6\u5730\u5740\u767b\u5165\u3002\u6211\u5011\u4e0d\u63a5\u53d7\u4f7f\u7528 Gmail\u3001Yahoo\u3001QQ \u6216\u5176\u4ed6\u500b\u4eba\u96fb\u5b50\u90f5\u4ef6\u5e33\u6236\u63d0\u51fa\u7684\u7533\u8acb\u3002<\/p>\n\n\n\n

\u5982\u8981\u53c3\u52a0\u65e9\u671f\u8a66\u7528\u6d3b\u52d5\uff0c\u8acb\u586b\u5beb\u7c21\u77ed\u7684\u7533\u8acb\u8868\uff0c\u4e26\u4e14\u8a73\u8ff0\u60a8\u7684\u4f7f\u7528\u6848\u4f8b\u3002<\/p>\n\n\n\n

<\/p>\n\n\n\n

\ud83d\udfe6ioioTIMES \u8981\u9001\u60a8SSD \u5feb\u4f86\u5b8c\u6210\u4efb\u52d9\u8a66\u624b\u6c23<\/a>
\ud83d\udfe6\u73fe\u5728\u5c31\u52a0\u5165 ioioTIMES \u81c9\u66f8\u7c89\u7d72\u5718<\/a> \u66f4\u591a\u4e92\u52d5\u3001\u66f4\u591a\u597d\u5eb7\u650f\u62b5\u52a0!!<\/strong>
\ud83d\udfe6\u6211\u5011\u6709LINE TODAY<\/a>\u983b\u9053\u4e86\uff0c\u5feb\u4f86\u8ffd\u8e2a\u6211\u5011\u5427!!–\u6700\u65b0\u79d1\u6280\u65b0\u805e \u76e1\u5728\u4f60\u624b<\/strong><\/h4>\n\n\n\n

<\/p>\n\n\n\n

<\/p>\n","protected":false},"excerpt":{"rendered":"

\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u63d0\u4f9b\u6975\u70ba\u51fa\u8272\u7684\u65b0\u529f\u80fd\uff0c\u64f4<\/p>\n","protected":false},"author":3,"featured_media":43492,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"rank_math_lock_modified_date":false,"footnotes":""},"categories":[13],"tags":[6359,71,7182,4692],"class_list":["post-43491","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-news","tag-h100-gpu","tag-nvidia","tag-tensorrt-llm","tag-4692"],"_links":{"self":[{"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=\/wp\/v2\/posts\/43491"}],"collection":[{"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=43491"}],"version-history":[{"count":2,"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=\/wp\/v2\/posts\/43491\/revisions"}],"predecessor-version":[{"id":43496,"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=\/wp\/v2\/posts\/43491\/revisions\/43496"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=\/wp\/v2\/media\/43492"}],"wp:attachment":[{"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=43491"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=43491"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=43491"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}