Hi! Hydra maintainer here. A heads-up about the upcoming Hydra 1.4 release.
Hydra 1.4 is not ready for production yet. It is a very large and long-delayed release, and it will contain several breaking changes accumulated during that extended development period. The current development release should only be used for compatibility testing.
One planned change is that Hydra 1.4 will stop accepting version_base values below "1.3". Omitting version_base will select current behavior, while version_base=None will be deprecated. The rationale and migration plan are tracked in Hydra #3237.
TorchRL currently:
- Declares
hydra-core>=1.1, which will allow Hydra 1.4 to be installed once released.
- Has approximately 50 entry points using
version_base="1.1", including the current PPO trainer.
- Has another four entry points using
version_base=None.
Under Hydra 1.4, the "1.1" entry points will fail during application startup.
Please migrate maintained entry points to version_base="1.3" and raise the Hydra dependency floor to hydra-core>=1.3. The version_base=None occurrences should normally remove the argument to select current behavior.
For compatibility testing, the current development release is hydra-core 1.4.0.dev6:
python -m pip install --upgrade "hydra-core==1.4.0.dev6"
Because 1.4 contains other breaking changes, I recommend running TorchRL's relevant test suite against this development release in addition to updating version_base.
TorchRL's structured configuration system is an extensive Hydra integration, so I wanted to flag this early enough to avoid surprising users when Hydra 1.4 becomes available.
Hi! Hydra maintainer here. A heads-up about the upcoming Hydra 1.4 release.
Hydra 1.4 is not ready for production yet. It is a very large and long-delayed release, and it will contain several breaking changes accumulated during that extended development period. The current development release should only be used for compatibility testing.
One planned change is that Hydra 1.4 will stop accepting
version_basevalues below"1.3". Omittingversion_basewill select current behavior, whileversion_base=Nonewill be deprecated. The rationale and migration plan are tracked in Hydra #3237.TorchRL currently:
hydra-core>=1.1, which will allow Hydra 1.4 to be installed once released.version_base="1.1", including the current PPO trainer.version_base=None.Under Hydra 1.4, the
"1.1"entry points will fail during application startup.Please migrate maintained entry points to
version_base="1.3"and raise the Hydra dependency floor tohydra-core>=1.3. Theversion_base=Noneoccurrences should normally remove the argument to select current behavior.For compatibility testing, the current development release is
hydra-core 1.4.0.dev6:python -m pip install --upgrade "hydra-core==1.4.0.dev6"Because 1.4 contains other breaking changes, I recommend running TorchRL's relevant test suite against this development release in addition to updating
version_base.TorchRL's structured configuration system is an extensive Hydra integration, so I wanted to flag this early enough to avoid surprising users when Hydra 1.4 becomes available.