Use structured configs to override Hydra config

[addisonklinke]

Background

Hydra launchers only get activated in multirun mode. I want to submit to a remote instance (SageMaker, Ray, k8s, etc) while avoiding the need to remember to specify --multirun each time (because end-users of the script aren't very familiar with Hydra)

Example

Attempt to make a new hydra config group called "multirun" (based on the configuring experiments guide)

from hydra.conf import HydraConf
from hydra.types import RunMode
from hydra_zen import store, zen


@store(name="config")
def add(x: int = 1, y: int = 1) -> None:
    print(x + y)


if __name__ == '__main__':

    store(
        HydraConf(mode=RunMode.MULTIRUN),  # Inherit all the other default Hydra settings
        group="hydra",
        name="multirun",
        package="_global_",  # Needed to avoid conflict with default hydra node in defaults list
    )

    store.add_to_hydra_store()
    zen(add).hydra_main(config_path=None, config_name="config", version_base="1.1")

Actual Behavior

$ python add.py 
2  # Works

$ python add.py +hydra=multirun --cfg hydra
In 'hydra/launcher/basic': ConfigKeyError raised while composing config:
Key 'launcher' not in 'Builds_add'
    full_key: launcher
    object_type=Builds_add

Expected Behavior

$ python add.py +hydra=multirun --cfg hydra
hydra:
  mode: MULTIRUN
  ...

Notes

I realize +hydra=multirun is just substituting another CLI override instead of --multirun. However, my long-term goal would be to encode this in Python with something like the following

@store(
    name="config",
    hydra_defaults=[
        "_self_",
        {"override hydra": "multirun"},
    ],
)

Then python add.py is multirun by default (without any CLI flags)

There are other use cases for @store(HydraConf(...)) besides setting multirun (i.e. job logging, output folder name, etc), so it seems like there should be a supported path

[rsokl]

Hrmm.. I think being able to swap out the full hydra config from the CLI is tricky here. Afaict Hydra looks specifically for the store entry: group=hydra, name=config, which is how the simple customize hydra config how-to works.

I have been finding a lot of success for my own use cases with scaffolding above the Hydra CLI. Specifically I overwrite zen.hydra_main and transform the sys.argv list before it gets sent to the underlying hydra_main call. I also do things like "if you specify a launcher, I will automatically add --multirun for you", and it is just so much easier to work with this logic than wrestle with complexities like the above. For this reason, I also don't have much insight into how to achieve ends like yours via Hydra itself. I just figured I'd share my approach, especially since I am not much help here.

[addisonklinke]

I overwrite zen.hydra_main and transform the sys.argv list before it gets sent to the underlying hydra_main call

Ah this is a great insight. I've actually done similar around the @hydra.main() decorator for re-running an old config (based on approach 2 in this comment). However the syntax of overwriting the decorator directly is a bit odd, and working with zen.hydra_main() would be much cleaner. I'll have to give it a try!