实验环境为RTX 3060的Windows 11。我在一开始没有安装合适的CUDA版本，导致最终Pytorch训练时报错，最终只能推倒重来，浪费了不少时间。

为了提醒更多的同学在一开始就安装正确的版本，在此记录。

经实验，完全没问题的版本是：

1. CUDA Toolkit 11.1.1

CUDA Toolkit 11.1 Update 1 Downloads | NVIDIA DeveloperPlease Note: We advise customers updating to Linux Kernel 5.9+ to use the latest NVIDIA Linux GPU driver R455 that will be available for download from NVIDIA website and repositories, starting today. Select Target Platform Click on the green buttons that describe your target platform. Only supported platforms will be shown. By downloading and using the software, you agree tohttps://developer.nvidia.cn/cuda-11.1.1-download-archive依次选择：Windows-x86_64-10-exe(local)

2. ​cuDNN v8.1.1 (Feburary 26th, 2021), for CUDA 11.0,11.1 and 11.2​

cuDNN Archive | NVIDIA Developerhttps://developer.nvidia.cn/rdp/cudnn-archive#a-collapse811-1113. Pytorch V1.8.0

安装指令为：

conda install pytorch==1.8.0 torchvision==0.9.0 torchaudio==0.8.0 cudatoolkit=11.1 -c pytorch -c conda-forge

如果这篇博客对你有帮助，请给博主点个赞鼓励一下吧！