Design to override individual config values from primary config

[addisonklinke]

Say I have a config like this (boilerplate _target_ fields removed)

trainer:
  dataset_info:
    prefix: s3://bucket
    size: full
  max_epochs: 10

Hydra's defaults list makes it easy for me to specify "config groups" like trainer/dataset_info=local in the primary config. However I haven't found a way to request "config value" overrides like trainer.max_epochs=1 with dot-syntax. It would be nice to codify these settings so the override doesn't have to be specified every time on the CLI

I'm aware something like this is possible

store(trainer, group="trainer", name="shorter", max_epochs=1)
hydra_defaults = [
    {"trainer": "shorter"},  # Use the newly stored config group
]

However this feels like a lot of boilerplate to create a whole config group just to override one primitive value

Here's an attempt that almost does what I want by introducing the concept of EntrypointConfig and using them to templatize_entrypoint()

from dataclasses import dataclass
from typing import Callable, Dict, List, Optional

from hydra import compose, initialize
from hydra.core.config_store import ConfigStore
from hydra_zen import zen, ZenStore


immediate_store = ZenStore(deferred_hydra_store=False, overwrite_ok=True)


# Application ----------------------------------------------------------------
@immediate_store(group="trainer/dataset_info", name="default")
@immediate_store(group="trainer/dataset_info", name="local", prefix="file://")
@dataclass
class DatasetInfo:
    prefix: str = "s3://bucket"
    size: str = "full"
    
    
@dataclass
class Trainer:
    dataset_info: DatasetInfo
    max_epochs: int = 10


def train(trainer: Trainer):
    """Inspect dataclass __str__ of the composed object"""
    print(trainer)


# Registration ---------------------------------------------------------------
@dataclass
class EntrypointConfig:
    name: str
    hydra_defaults: Optional[List[Dict[str, str]]]
    value_overrides: Optional[List[str]] = None


def templatize_entrypoint(
    func: Callable,
    configs: List[EntrypointConfig]
):
    """Store multiple root configs for a single entrypoint function"""
    for config in configs:
        immediate_store(
            func,
            name=config.name,
            hydra_defaults=config.hydra_defaults,
        )
        cs = ConfigStore.instance()
        original = cs.repo[config.name + ".yaml"].node  # For reference while debugging
        with initialize(config_path=None, version_base="1.3"):
            updated = compose(
                config_name=config.name,
                overrides=config.value_overrides,
            )
        immediate_store(updated, name=config.name)
    
    
if __name__ == "__main__":
    
    hydra_defaults = [
        "_self_",
        {"trainer/dataset_info": "default"},
    ]
    entrypoint_configs = [
        EntrypointConfig(
            name="config",
            hydra_defaults=hydra_defaults,
        ),
        EntrypointConfig(
            name="shorter",
            hydra_defaults=hydra_defaults,
            value_overrides=[
                "trainer.max_epochs=1",
            ],
        )
    ]
    templatize_entrypoint(
        func=train,
        configs=entrypoint_configs,
    )
    zen(train).hydra_main(
        config_name="config",
        config_path=None,
        version_base="1.3",
    )

Now you can run the default config

$ python app.py 
Trainer(dataset_info=DatasetInfo(prefix='s3://bucket', size='full'), max_epochs=10)

Or use one of the other config_names as an alias to avoid typing an entire override. Granted we're only overriding one config value here so characters saved are minimal, but in theory this supports unlimited config group and value settings under one flag

$ python app.py -cn shorter
Trainer(dataset_info=DatasetInfo(prefix='s3://bucket', size='full'), max_epochs=1)

Where things seem to break is using both -cn and a config group CLI override (local instead of default)

$ python app.py -cn shorter trainer/dataset_info=local
Trainer(dataset_info=DatasetInfo(prefix='file://', size='full'), max_epochs=1)  # Expected

# Actual
Could not override 'trainer/dataset_info'. No match in the defaults list.
To append to your default list use +trainer/dataset_info=local

Looking for feedback on

1. Is there a reason "config value" settings on the primary config aren't supported - perhaps proper "config groups" have enough advantages that they're worth the extra boilerplate?
2. Critiques of the proposed approach. One alternative I considered is dynamically creating the boilerplate config group so hydra_defaults can still be used to communicate the information vs. needing to compose() a new node
3. Explanations for why the proposed approach breaks when adding CLI config group args. While inspecting original vs. updated nodes in a debugger, I notice only the former includes defaults: ListConfig so I think the lack of this component in the Compose API return is suspicious