Add pytorch to ML image (or create separate image)

[rabernat]

For LEAP and M2LInES 2i2c Hubs, many researchers will want to use pytorch. Unfortunately, this image does not have pytorch.

When I try to add pytorch-gpu to ml-notebook/environment.yml and call

conda-lock lock --mamba -f environment.yml -f ../pangeo-notebook/environment.yml -p linux-64

I get the error

Encountered problems while solving.
Problem: nothing provides __glibc >=2.17,<3.0.a0 needed by pytorch-1.10.0-cuda102py37h689c94d_1

I found a similar problem in NVIDIA/MinkowskiEngine#354. That issue implies we may need to mix conda channels to get this to work. The workaround they recommend is to install openblas from anaconda and then switch to conda-forge for pytorch.

Any thoughts on the best path here?

cc @yuvipanda

Also related to #188

[ngam]

Encountered problems while solving.
Problem: nothing provides __glibc >=2.17,<3.0.a0 needed by pytorch-1.10.0-cuda102py37h689c94d_1

Could you try to force later pytorch and/or higher cudatoolkit? This error is likely related to the underlying image you're using (or the machine). __glibc <2.17 is pretty old :(

The openblas addition (at least for recent pytorch versions from conda-forge) likely makes no difference because pytorch uses mkl, not openblas, if I understand correctly.

You can also try to see the torch config: torch.__config__.show()

[ngam]

Or are you trying this on a personal machine. I will test it quickly to test why this is happening.

[ngam]

@rabernat the error is coming from this: https://github.com/conda-forge/pytorch-cpu-feedstock/blob/93333de760d702aee276dcafeedc890458a00417/recipe/meta.yaml#L34

Something is forcing __glibc <2.17. I will look more closely at this later. It is probably also holding up other packages from getting upgraded. Quite a few packages moved to __glibc >=2.17 on conda-forge.

[ngam]

Paging @hmaarrfk who may be able to advise on this situation. Odd because the successful image ends up being 1.10.2 and pulling glibc 2.17 if you look closely at the CI.

[hmaarrfk]

It seems that you are using ubuntu images with 20.04 though. So I'm really not sure.

[ngam]

Yep, I would guess it is a simpler conflict between packages somehow that may need to be sorted out manually. Thanks a lot @hmaarrfk and sorry for bothering you with this here!

[hmaarrfk]

Can you check the output of

conda info

before running any other conda commands?