search

choderalab / mtenn

Modular Training and Evaluation of Neural Networks
MIT License
5 stars 1 forks source link

CI fixes for conda forge #29

Closed hmacdope closed 9 months ago

hmacdope commented 9 months ago

Fixes #28

codecov-commenter commented 9 months ago

Codecov Report

Merging #29 (c6f3c04) into main (07ea82b) will increase coverage by 0.33%. Report is 9 commits behind head on main. The diff coverage is 32.14%.

Additional details and impacted files
hmacdope commented 9 months ago

@mikemhenry @kaminow I think the failure to solve for 3.9 and 3.11 is due to BLAS restrictions in our conda recipie for DGL, pushing a hopeful fix now https://github.com/conda-forge/dgl-feedstock/pull/20

hmacdope commented 9 months ago

@mikemhenry would you mind having a look here at the conflict graph, as far as I can tell there are 3.9 has a bunch of issues and 3.11 is blocked by some pytorch-cluster issues.

hmacdope commented 9 months ago

If we can't get to the bottom of this, then we may have to use the dglteam channel in the short term. @kaminow

ijpulidos commented 9 months ago

Interestingly enough, I can create an environment fine with mamba (not micromamba) if I use the following env file (removed the dglteam channel and added python=3.9).

name: test
channels:
  - conda-forge
dependencies:
  - python=3.9
  - pytorch
  - pytorch_geometric
  - pytorch_cluster
  - pytorch_scatter
  - pytorch_sparse
  - numpy
  - h5py
  - e3nn
  - dgllife
  - dgl
  - rdkit
  # testing dependencies
  - pytest
  - pytest-cov
  - codecov

and running mamba env create -f devtools/conda-envs/test_env.yaml -n mtenn-test

mikemhenry commented 9 months ago

If we can't get to the bottom of this, then we may have to use the dglteam channel in the short term. @kaminow

Looking into this, but IMHO, way better to just drop python 3.9 support (or not test it on CI since it sounds like it is only a problem on CI, since @ijpulidos can make an env and so can I, micromamba create -n mtenn-py39 --dry-run --file devtools/conda-envs/test_env.yaml python=3.9 works for me on on Linux)

mikemhenry commented 9 months ago

CONDA_OVERRIDE_CUDA="" micromamba create -n mtenn-py39 --dry-run --file devtools/conda-envs/test_env.yaml python=3.9 works for me as well, thought I should check that to make sure it wasn't a GPU thing

mikemhenry commented 9 months ago

So my vote is to not worry about this unless we need python 3.9 in a production environment and drop python 3.9 in our testing matrix.

kaminow commented 9 months ago

I think the main worry with ignoring the issues with 3.9 is that it may mess with the asapdiscovery CI for 3.9, so we'd have to drop it there as well

mikemhenry commented 9 months ago

It looks like we have mtenn in our CI already there and it is working https://github.com/choderalab/asapdiscovery/blob/main/devtools/conda-envs/asapdiscovery-ubuntu-latest.yml#L90 so we wouldn't have to drop 3.9 from our CI matrix for asapdiscovery