Forums

I have checked the existing posts and make sure all the kernels are killed. But I still keep having this problem when running the Pytorch tutorial.

which pytorch tutorial is this?

@conrad It is autograd_tutorial, I failed every time when trying to calculate the gradients at out.backward()

I see from the docs that PyTorch uses "acceleration libraries" such as Intel MKL and NVIDIA. If you run the 60 minute blitz -> autograd tutorial from a python shell/python script, you will see the error message:

Intel MKL FATAL ERROR: Cannot load <mkl-loader>.

I wonder if there is a way to switch this off. (eg: for numpy, it doesn't fail when it can't find Intel MKL, but I guess PyTorch expects it and fails hard when it doesn't find it)

Alternatively, you can try to compile MKL manually and link/set up the paths so that PyTorch can find it. (not 100% sure if this is possible) We will also raise a ticket to get it installed during one of our package builds.

Thanks a lot! @conrad

Hi

I am also receiving this error message, when trying to run a webapp that uses pytorch:

has there been any progress in making MKL available, and if available, how would I get access to it?

Thanks Don Cameron donrcameron@gmail.com

MKL is not available yet, but it's on the list of things we want to look closer at.

Mea culpa.

Not unexpectedly, it was my mistake. I had set up an environment that exactly matched my Windows 10 environment. Nuking the environment from orbit (pip uninstall all my packages), and then reinstalling all the latest Linux packages, fixed the problem. The mkl library gets loaded as part of pytorch, I assume.

Thanks for your time, Don Cameron