Open MDanialKhan opened 3 days ago
That's true, the "nightly" build is used in case of the image with CUDA, as you can see here: https://github.com/chaimeleon-eu/workstation-images/blob/bea80dbdd5cdfd3ceddb1a53b1da11f7b343fac1/ubuntu-python-pytorch/Dockerfile#L39
That's probably due to compatibility reasons with CUDA 11.8 when that dockerfile was created. I think this because on the other hand the image for CPU (without CUDA) is built with the stable version of pytorch (as seen in the line before). Maybe now (after a year) there is no problem to use the stable version also with CUDA libraries.
So a new version of the image with cuda and pytorch stable (2.5.1) has been uploaded and you can try with:
jobman -i ubuntu-python-pytorch:3.10cuda11 ...
or:
jobman -i ubuntu-python-pytorch:latest-cuda ...
I have a question regarding the pytorch version being used when we are running scripts on jobman. The current version is 2.6.0, of the dev branch when the current stable release is 2.5.1.
I have attached this image for your reference.
The updates break our code as we are using a previous stable version of pytorch. Is there a way to restrict the version ourselves? as we are not overriding the version that is used by the jobman images.