Had the same problem and got to the bottom of it. My configuration is a Tesla T4 with 410.92 driver and CUDA 9.2, Ubuntu 18.10.
The thing is, the torch installer tries for some reason to use the highest compute capability supported by the device (or the driver - not sure which), but ignores the compute capability supported by the CUDA toolkit.
So, in my case the device supports compute capability 7.5, but CUDA 9.2 supports only 7.0 or 7.2 (not sure which one). You can guess I got the same nvcc fatal : Unsupported gpu architecture 'compute_75'
The solution is to force the nvcc compile options to use a lower compute capability. This can be achieved by setting the following environment variable:
export TORCH_CUDA_ARCH_LIST="7.0"