haste_pytorch does not install properly with conda cudatoolkit? about haste HOT 3 CLOSED

My best guess is that your installed PyTorch isn't using the same CUDA toolkit as Haste. When a prebuilt PyTorch is installed (e.g. via pip or conda), it uses its own copy of CUDA Toolkit instead of what's installed on your system. Haste, on the other hand, uses the CUDA toolkit found in /usr/local/cuda. I've seen segfaults with Haste + PyTorch when I had CUDA version mismatches.

Conda doesn't install the CUDA toolkit to /usr/local/cuda so you'll need to edit Haste's Makefile to point to the right place.

from haste.

Hi Thanks for the answer!
As for the Segfault message, When I changed the Cuda from 11.4 to 11.1 (in PATH ) and downloaded and built it from the source, rather than with pip, it worked. (It did not work with pip nor source with 11.4. Maybe I should've collected the build error msg. my bad)

As for the other question, thank you for the clarification. seems like the repo2docker uses old (<10.1 ) cuda toolkit by default and installing it by changing the base docker image to devel version of nvidia/ubuntu that contains all the necessary build tools is not possible, so I'll go with another option on that one.

Thanks again!

from haste.

No problem, glad you got up and running at least locally. Let me know if you run into any other issues that I can help with.

from haste.