Pinning exact versions of dependencies in a package published on PyPI can cause problems for people with specific requirements. Please see this StackOverflow answer for background.
In my case, my GPU version is newer than what is supported by the version of torch listed in requirements.txt / setup.py.
Also I want to use one of the PyTorch packages (python-pytorch-opt-cuda) provided by my distro (Arch Linux), even in virtualenvs, because of specific compilation options and disk space concerns.
My distro package for PyTorch also depends on a version of numpy which is incompatible with the version in requirements.txt / setup.py.
These are just some of the issues that people like me may face.
Pinning exact versions of dependencies in a package published on PyPI can cause problems for people with specific requirements. Please see this StackOverflow answer for background.
In my case, my GPU version is newer than what is supported by the version of
torchlisted inrequirements.txt/setup.py. Also I want to use one of the PyTorch packages (python-pytorch-opt-cuda) provided by my distro (Arch Linux), even in virtualenvs, because of specific compilation options and disk space concerns. My distro package for PyTorch also depends on a version ofnumpywhich is incompatible with the version inrequirements.txt/setup.py.These are just some of the issues that people like me may face.