Specifying CUDA Location when Building from Source

I am trying to build the pytorch/master branch from source on an AWS instance. The default CUDA nvcc version is 10.0, found in /usr/local/cuda/bin/nvcc, and is incompatible with PyTorch version 1.9+ (since version 10.2 or above is required).

On the instance, I also have newer versions of CUDA (e.g. in /usr/local/cuda-11.0/). I cannot figure out how to reconfigure the PyTorch build to use that version instead of the default. Any help?

Try to use CUDA_HOME=/usr/local/cuda-11.0/ python setup.py install .

1 Like