[PyTroch] torch_cuda.set_device() and torch.manual_seed() hang about javacpp-presets HOT 5 CLOSED

In fact, these functions hang on multi-GPU box. They work fine on a single GPU box. I tried official libtorch build from pytorch.org and saw the behaviors. I suspect that the issue is not with PyTorch but with cuda-platform package. cudart.cudaDeviceProp method hangs on multi-GPU box too.

from javacpp-presets.

Python wrapper of these methods work fine. I don't think that C++ functions would hang. I guess that we may have to do some extra initialization work on multi-GPU systems?

from javacpp-presets.

cuInit does the magic. Thanks a lot! It is interesting why we don't need to call it in single GPU system.

from javacpp-presets.

That basically just calls the CUDA functions through JNI, so I'm guessing that calling them directly from C++ also hangs, right?

from javacpp-presets.

I guess not calling cuInit() could do that?
https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__INITIALIZE.html

from javacpp-presets.