How to specify pytorch / cuda version in pipenv

Question

I am trying to install a specific version of pytorch that is compatible with a specific cuda driver version with pipenv. The pytorch website shows how to to this with pip:

pip3 install torch==1.3.1+cu92 torchvision==0.4.2+cu92 -f https://download.pytorch.org/whl/torch_stable.html

I tried to convert this into an entry in my Pipfile like this:

[[source]]
name = "pytorch"
url = "https://download.pytorch.org/whl/torch_stable.html"
verify_ssl = false

pytorch = {version="==1.3.1+cu92", index="pytorch"}
torchvision = {version="==0.4.2+cu92", index="pytorch"}

However, this does not work. The dependency with this version can not be resolved. I am not sure if the url that is listed with the -f parameter in the pip3 command is even a valid source for pipenv.

I could install both libraries by just passing the command through to pip like this:

pipenv run pip install torch==1.3.1+cu92 torchvision==0.4.2+cu92 -f https://download.pytorch.org/whl/torch_stable.html

but I am not really satisfied with that solution since the dependencies are not in the Pipfile and I have to manually document the usage of this command.

Answer 1

The problem with the approach above lies in the structure of https://download.pytorch.org/whl/torch_stable.html. Pipenv can only find torch versions 0.1 to 0.4.1 because all others have the cuda (or cpu) version as a prefix e.g. cu92/torch-0.4.1-cp27-cp27m-linux_x86_64.whl.

But the cuda version is a subdirectory. So if you change the url of the source to the cuda version and only specify the torch version in the dependencies it works.

[[source]]
name = "pytorch"
url = "https://download.pytorch.org/whl/cu92"
verify_ssl = false

[packages]
torch = {index = "pytorch",version = "==1.4.0"}

The only problem I encountered is that numpy is not recognized as a dependency of pytoch 1.4.0. But this seems to be a problem of the specific pytorch wheel. With version 1.3.1 or 1.5.1 and a recent pipenv version it works.

So if after the installation with pipenv install, the command pipenv run python -c "import torch" throws an error, numpy must be added manually.