[Bug]: Serialized QuantizationMixin subclasses fail to load

[brian-dellabetta]

⚙️ Your current environment

I introduced a bug in #1772

🐛 Describe the bug

When a recipe with modifiers extending the QuantizationMixin class is serialized, it fails to load back up from the yaml. If config_groups field is set, the targets field is set during validation to contain the full list of targets nested in config_groups. When saved, both targets and config_groups are saved, and when re-loaded this error is hit, because both targets and config_groups cannot be set.

🛠️ Steps to reproduce

from compressed_tensors.quantization import (
    QuantizationScheme,
    QuantizationStrategy,
    QuantizationType,
    QuantizationArgs,
)
from llmcompressor.modifiers.quantization import (
    QuantizationModifier,
)
from llmcompressor.recipe import Recipe

recipe = Recipe.from_modifiers(
    QuantizationModifier(
        config_groups={
            "group_0": QuantizationScheme(
                targets=["Linear"],
                weights=QuantizationArgs(
                    num_bits=4,
                    type=QuantizationType.INT,
                    strategy=QuantizationStrategy.GROUP,
                    group_size=128,
                    symmetric=True,
                    dynamic=False,
                ),
            )
        },
        ignore=["lm_head"],
    )
)

yaml_str = recipe.yaml()

new_recipe = Recipe.create_instance(path_or_modifiers=yaml_str)

results in error

pydantic_core._pydantic_core.ValidationError: 1 validation error for QuantizationModifier
  Value error, Please specify either `targets` or `config_groups` [type=value_error, input_value={'group': 'default', 'con..., 'ignore': ['lm_head']}, input_type=dict]

[brian-dellabetta]

Two potential solutions:

1. Update the validation to make sure targets and config_groups, if both are set, contain the same list of targets.
2. Update the validation to not set targets based on what is in config_groups, instead creating an in-memory @property def resolved_targets(self). This avoids updating the data model during validation, so the input recipe yaml will always match the output recipe yaml.

Option 2 would be cleaner but we'd have to make sure we're using self.resolved_targets everywhere instead of self.targets

[kylesayrs]

Shouldn't we not be setting default values for targets? I think that ideally targets is just a pass-through to config groups, and doesn't ever get set/serialized.

Said another way, targets should be a convenience field which makes the construction of config groups easier, and config groups is the ground truth.