tch::Cuda::is_available() returns False

[pramodith]

Hi, thanks for the work put into the library it's awesome.

Using Windows 10.

I've been able to run the transformers models on CPU, but not on GPU.

I see that
tch::Cuda::is_available()
returns False.

This is despite the fact that I have Cuda installed on my system.
**nvcc --version **
returns

I've also donwloaded LibTorch and added it to my system path. Any guidance to get the model working on GPU would be appreciated.

[guillaume-be]

Hello @pramodith and thank you for the interest in the library.
This is an issue with the set-up of the tch-rs crate dependency. A user experienced similar issues with Windows a few days ago (see #24).

I am also running the library on Windows with CUDA 10.2 and am not experiencing issues so far. Please try the following and let me know if any of these help:

• try turning off your environment variables for LIBTORCH, on windows this was not helping in my case
• set the environment variable for your account TORCH_CUDA_VERSION to 10.2 (or try 10.1
• Delete your compilation target folder to force a clean re-compilation of tch-sys

If these do not help I suggest you open an issue on the tch-rs crate