Loosen the dependency to avoid conflict with other package

[chiang-yuan]

@ilyes319 could we loosen the dependency of e3nn to e3nn>=0.4.4

mace/setup.cfg

Line 18 in 22a2e3e

e3nn==0.4.4

[ilyes319]

Can you try locally updating to the latest e3nn and see if everything training works.

[chiang-yuan]

I am trying to wrap a pyproject.toml dependency. e3nn is the only one causing the problem. Locally updating is possible for user but not package wrapping unfortunately. :( Any reason that later version of e3nn is not working / unsafe?

[ilyes319]

I meant can you confirm that locally e3nn latest version works fine?

[chiang-yuan]

Not sure at this moment. I could update locally and try training for the next model. Another way is test workflow? Is there anything missing in the test that we should probably add?

[ilyes319]

big missing part in the testing workflow atm are training with h5 files and multigpu (that one we can't not really test).

[chiang-yuan]

I see. I will try adding offline testing script and small demo data. And pick up #452 and solve it altogether

[chiang-yuan]

related to #555

[ilyes319]

I figured something changed in the Clebsch Gordan convention between 0.4.4 and 0.5.1 that breaks backward compatibility with foundation models so I will keep the pin for e3nn for now.