Since the last release a few things were changed:
The backend was changed to metatomic from metatonsor-atomistic) and last layer readouts were added.
For the next release a convenience function should be added that avoids confusion with the users:
If users update shiftml-versions, old ShiftML3 model files (metatensor-atomistic models) may persist in the user cache.
Ideally we would either catch an import error, tryingn a forced download or compar ie MD5 hashes and again force a download of new checkpoints.
Since the last release a few things were changed:
The backend was changed to metatomic from metatonsor-atomistic) and last layer readouts were added.
For the next release a convenience function should be added that avoids confusion with the users:
If users update shiftml-versions, old ShiftML3 model files (metatensor-atomistic models) may persist in the user cache.
Ideally we would either catch an import error, tryingn a forced download or compar ie MD5 hashes and again force a download of new checkpoints.