So as far as I understand CUDA and cuDNN are automatically installed when running pip/pip3 install torch
. But what about when using conda install pytorch torchvision -c pytorch
? Does the PyTorch Conda package also come with it’s own copies of CUDA and cuDNN?
As far as I know, yes, the conda package also ships with its CUDA and cuDNN binaries.
This build script also suggests the same.
I can confirm ptrblck’s answer. pytorch 0.4 statically links to cudnn, nccl and most cuda libraries within itself.
Almost pytorch conda package still requires conda cuda toolkit package, which will bring the following unnecessary goodies with it:
/opt/conda/pkgs/cudatoolkit-9.0-h13b8566_0/lib/libcusolver.so.9.0.176
/opt/conda/pkgs/cudatoolkit-9.0-h13b8566_0/lib/libcusparse.so.9.0.176
/opt/conda/pkgs/cudatoolkit-9.0-h13b8566_0/lib/libcurand.so.9.0.176
/opt/conda/pkgs/cudatoolkit-9.0-h13b8566_0/lib/libnppif.so.9.0.176
/opt/conda/pkgs/cudatoolkit-9.0-h13b8566_0/lib/libcufft.so.9.0.176
/opt/conda/pkgs/cudatoolkit-9.0-h13b8566_0/lib/libcublas.so.9.0.176
I installed pytorch 1.0 on anaconda after installing only the graphic driver.
And I confirmed that the torch.cuda.current_device () function works well.
That means that if I install only a graphic driver,
can I use pytorch with every functions and full speed without installing cuda and cudnn?
If yes, where is it mentioned?