[Windows] Cuda shows disabled using TORCH_CUDA_VERSION=cu101

[asilvas]

Install works, and verified in the build output of torch-sys that it's building dummy_cuda_dependency.cpp which indicates it detected Cuda. But for some reason that I've not yet determined both tch::Cuda::is_available() and tch::Cuda::cudnn_is_available() return false. Everything functions otherwise (in CPU mode), just need it to get this working with GPU.

To be clear, this isn't an issue with pytorch. I've verified it's working correctly.

>>> import torch
>>> torch.cuda.current_device()
0
>>> torch.cuda.device(0)
<torch.cuda.device object at 0x0000018746452390>
>>> torch.cuda.device_count()
1
>>> torch.cuda.get_device_name(0)
'GeForce GTX 1080'
>>> torch.cuda.is_available()
True

[LaurentMazare]

I don't have a windows box at hand to try so won't be of much help here. What I would suggest doing is a manual install of libtorch to see if it works better (details here). On linux, I would use ldd to check the shared library that the executable uses and check if it gets the cuda version, seems that the windows equivalent is dumpbin /dependents.

[asilvas]

I've used pip, libtorch prebuilt, and the torch-sys install methods. I can install fine, but as indicated above the bindings are reporting Cuda unavailable despite manually verifying via python Cuda is available. Hoping someone else had a similar issue and found a workaround.. I don't get how bindings would report different information despite verifying that python is loading the same version of pytorch.

[rookboom]

I see the same on windows using TORCH_CUDA_VERSION=cu110, but only for --release, CUDA works fine for debug builds.

[LaurentMazare]

If it's only for the release version, this may be closer to #291. This one is likely to get away once some cargo feature has landed (I think it will be part of the 1.50 release although I'm not sure whether this will be in stable at this point).

[LaurentMazare]

Closing old issue, feel free to re-open if it's still a problem to you (I just tested it on linux and it worked fine, also sadly the new cargo feature from 1.50 is not general enough to handle this very use case).