Build Pytorch with Openblas and Cuda

Hello, I have a AMD Threadripper CPU and want to try to avoid installing Pytorch with Intels MLK library. I downloaded and installed OpenBLAS 0.3.7 on my Linux (Ubuntu) machine and installed Cuda drivers 10.1 for my 2080ti. When I clone Pytorch from git and install it via “python setup.py install”, it failed to auto detect Openblas and my Cuda drivers and installs it with MLK. Do I need to install it specifying anything?

From the setup.py file, you should be able to set BLAS=OpenBLAS to force it to use open blas.
The cuda issue should be independent as the blas libraries above are cpu only. If your cuda is not detected, again from the setup.py file, you can set the CUDA_HOME=/path/to/your/cuda.

1 Like