Closed kfir4444 closed 1 year ago
Thanks! Can you explain the rationale?
Sure @alongd!
The old tani_env was indeterministic in package version. When specifying:
- pytorch::pytorch
without a version, most likely mamba/conda takes a version that "works" with all package requirements and is the latest.
The issue I solve here stems from the fact that pytorch deprecated a function: torch.symeig
in favor of torch.linalg.eigh
. However, TorchANI uses torch.symeig
, so we no longer have the privilege of not specifying which exact version of pytorch. Yet, when specifying an exact pytorch version, there are conflicts with other packages, that seem to last a small infinity.
So, to solve this issue, @Laxzal and I used an existing tani_env
with the latest pytorch version that allowed usage of torchani, and froze it with all the specific packages, so we always know exactly which packages to install for torchani, and we know that it always works.
Codecov Report
see 1 file with indirect coverage changes
:mega: We’re building smart automated test selection to slash your CI/CD build times. Learn more