Correspondence between pytoch version, CUDA version and graphics card driver version

Hello everyone , I meet you again , I'm your friend, Quan Jun .

Reference link ：INSTALLING PREVIOUS VERSIONS OF PYTORCH solve PyTorch And CUDA Version mismatch

1.CUDA Drive and CUDAToolkit Corresponding version

notes ： The driver is downward compatible , It determines what can be installed CUDA and CUDAToolkit Highest version .

2.CUDA And its availability PyTorch Corresponding version （ Refer to the official website , Welcome to the comments area to add ）

notes ： Although some cards CUDA The version can be updated to the new version , And PyTorch You can also update to the new version . But some corresponding installation packages cannot be used , It may be because the card is too old .

3. Installation instructions

（1） Specify installation PyTorch edition When known CUDA version , According to table 2 Directly query the corresponding version PyTorch, function conda install pytorch=X.X.X -c pytorch You can install the specified version PyTorch. This order is made by conda Decision and PyTorch Corresponding CUDAToolkit. But there is no guarantee PyTorch Normal use ,CUDAToolkit The version is not suitable for the graphics card driver , Can lead to CUDAToolkit Version higher than CUDA drive . （2） Appoint CUDAToolkit edition The first run nvidia-smi Inquire about CUDA Driver version , According to 1 Query the corresponding CUDAToolkit edition , Run again conda install pytorch cudatoolkit=X.X -c pytorch You can install the specified CUDAToolkit Version of PyTorch. （3） At the same time specified PyTorch and CUDAToolkit edition If you are quite sure CUDA Version and corresponding PyTorch and CUDAToolkit The corresponding version can run conda install pytorch=X.X.X cudatoolkit=X.X -c pytorch

After installation, you can use python see

import torch
print(torch.__version__)
print(torch.cuda.is_available())

Copy

Publisher ： Full stack programmer stack length , Reprint please indicate the source ：https://javaforall.cn/148561.html Link to the original text ：https://javaforall.cn