UM2N Docker build

[jwallwork23]

Closes #58.

[jwallwork23]

Whenever I try to install pytorch3d it crashes my laptop, so I built a reduced version of the Docker image, pushed it to ghcr, enabled mesh-adaptation-docs to update it and have now triggered the update. Hopefully it works!

[ddundo]

I cancelled my workflow on #52 until you get this done since I see the error is that there is no space left. I guess it's too much to have 4 of them running at the same time :)

[jwallwork23]

Hmm "no space left on device" error https://github.com/mesh-adaptation/mesh-adaptation-docs/actions/runs/11642012121/job/32421127119?pr=59. I'll try without pytorch3d.

[jwallwork23]

I cancelled my workflow on #52 until you get this done since I see the error is that there is no space left. I guess it's too much to have 4 of them running at the same time :)

What, do the runners share memory? I thought they were separate. Do you have a link to docs on this?

[jwallwork23]

Free runners have 14GB and the UM2N image I created is already 10.4GB. https://docs.github.com/en/actions/using-github-hosted-runners/using-github-hosted-runners/about-github-hosted-runners#standard-github-hosted-runners-for--private-repositories

This makes cutting down the image sizes a priority: #13.

[ddundo]

I cancelled my workflow on #52 until you get this done since I see the error is that there is no space left. I guess it's too much to have 4 of them running at the same time :)

What, do the runners share memory? I thought they were separate. Do you have a link to docs on this?

No idea really! I just saw that in my last workflow the "Initialise container" step took over 15 minutes, when it's usually about a minute. So I assumed it's because you had these other 3 workflows running :)

Edit: yeah it's super slow still @jwallwork23. It seems to have got slow after the firedrake-um2n container was pushed to ghcr. Not sure if connected. Would you mind deleting that container on ghcr if you don't need it, just to test if it gets faster?

I am also confused what triggered this workflow to run https://github.com/mesh-adaptation/mesh-adaptation-docs/actions/runs/11642077765 and push a new image to ghcr. It's only supposed to push it if triggered on the main branch... do you know?

[jwallwork23]

Edit: yeah it's super slow still @jwallwork23. It seems to have got slow after the firedrake-um2n container was pushed to ghcr. Not sure if connected. Would you mind deleting that container on ghcr if you don't need it, just to test if it gets faster?

Okay, I deleted it for now.

I am also confused what triggered this workflow to run https://github.com/mesh-adaptation/mesh-adaptation-docs/actions/runs/11642077765 and push a new image to ghcr. It's only supposed to push it if triggered on the main branch... do you know?

No, sorry.

[ddundo]

Thanks, but that didn't help. Pulling locally is also very slow. Maybe it's a ghcr issue at the moment.

Hmm... I think we then need to add the branches filter in docker_firedrake-parmmg.yml, i.e.:

  workflow_run:
    workflows: ['Build bespoke PETSc Docker container']
    types: [completed]
    branches: [main]

Would you mind adding this in this PR before you build again please? It seems like the workflow_run trigger defaults to the main branch no matter where the triggering workflow was run.

Edit: added this in #61

[jwallwork23]

I realised that by dropping torchvision and torchaudio in 19a17cf we can cut down the size of the Docker image considerably.

[ddundo]

@jwallwork23 I think the size is so large because pip3 install torch torchvision torchaudio installs a lot of CUDA stuff that we can't make use of anyway. Could you try with pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu? As per torch installation instructions for CPU only https://pytorch.org/get-started/locally/

Edit: just saw the comment above... sorry. Might still be worth doing pip3 install torch --index-url https://download.pytorch.org/whl/cpu?

[jwallwork23]

Might still be worth doing pip3 install torch --index-url https://download.pytorch.org/whl/cpu?

Good point, addressed in ae5dd0e.

[ddundo]

This looks good! But when I pull the image and try to import UM2N, I get

ModuleNotFoundError                       Traceback (most recent call last)
Cell In[2], line 1
----> 1 import UM2N

File ~/firedrake/lib/python3.12/site-packages/UM2N/__init__.py:7
      3 os.environ["OMP_NUM_THREADS"] = "1"
      5 from pkg_resources import DistributionNotFound, get_distribution  # noqa
----> 7 from .processor import *  # noqa
      8 from .generator import *  # noqa
      9 from .model import *  # noqa

ModuleNotFoundError: No module named 'UM2N.processor'

This is what caused https://github.com/mesh-adaptation/mesh-adaptation-docs/actions/runs/11881234411/job/33105325669 to fail as well.

Edit: I was confused why test_import.py passes if I get this error. See mesh-adaptation/UM2N#57

[jwallwork23]

Okay, I think the combination of mesh-adaptation/movement#133, 3bf32d1 and 1817d15 does the trick. If I run the Docker image and activate the Firedrake venv inside it then I can run the tests successfully.

[ddundo]

Thanks @jwallwork23! Could you please push the UM2N container to ghcr so I can test it?

[jwallwork23]

Thanks @jwallwork23! Could you please push the UM2N container to ghcr so I can test it?

Sure, uploaded just now. (It took a while.)

[ddundo]

This looks good now! Thanks again :)