chore(config): move shared config primitives into runtime package

[samet-akcay]

Summary

• add a runtime-owned physicalai.config package for shared configuration utilities
• move ComponentSpec out of physicalai.inference.manifest while keeping the old import path as a compatibility re-export
• update runtime inference code to use the shared config primitive directly

Why

Config, ComponentSpecand the generic config helpers need to live in the torch-free runtime package so both the runtime repo and the training library can use the same primitive. This keeps configuration ownership in one place and removes the long-term duplicatephysicalai.config` implementation.

Compatibility

This PR keeps the existing import path working:

from physicalai.inference.manifest import ComponentSpec

New code can use:

from physicalai.config import ComponentSpec

⚠️ Follow-up

This is the first half of the migration. The paired training-library PR removes the duplicate library/src/physicalai/config package and switches library imports to the runtime-owned physicalai.config module.

[Copilot]

Pull request overview

This PR introduces a runtime-owned physicalai.config package to centralize shared configuration primitives (notably ComponentSpec) so both runtime and training code can depend on a torch-free implementation, while preserving backward compatibility via re-exports from physicalai.inference.manifest.

Changes:

• Added new physicalai.config package (Config base, ComponentSpec, instantiation + serialization helpers).
• Moved ComponentSpec out of physicalai.inference.manifest and updated inference code to import it from physicalai.config.
• Added a unit test asserting the legacy physicalai.inference.manifest.ComponentSpec import remains compatible.

Reviewed changes

Copilot reviewed 11 out of 11 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
tests/unit/inference/test_manifest.py	Updates imports and adds a test asserting ComponentSpec is re-exported from the legacy path.
src/physicalai/inference/runners/factory.py	Switches runner-spec validation to use physicalai.config.ComponentSpec.
src/physicalai/inference/model.py	Removes runtime import of ComponentSpec from manifest; keeps type-only reference to the shared primitive.
src/physicalai/inference/manifest.py	Removes local ComponentSpec definition and imports ComponentSpec from physicalai.config to maintain compatibility.
src/physicalai/inference/component_factory.py	Updates doc/type references to point to physicalai.config.ComponentSpec.
src/physicalai/config/init.py	Exposes the new shared config primitives as a public package API.
src/physicalai/config/base.py	Adds dataclass-backed Config helper for dict/YAML roundtrips.
src/physicalai/config/component.py	Adds the shared ComponentSpec Pydantic model.
src/physicalai/config/instantiate.py	Adds generic instantiation utilities for {class_path, init_args}-style configs.
src/physicalai/config/mixin.py	Adds FromConfig mixin + decorator for config-based constructors.
src/physicalai/config/serializable.py	Adds dataclass (de)serialization helpers with basic type-hint reconstruction.

Comments suppressed due to low confidence (2)

src/physicalai/config/instantiate.py:58

• After config = config[key], config is no longer guaranteed to be a dict. The subsequent if "class_path" in config: and later config.keys() usages will raise TypeError/AttributeError if the selected value is not a mapping. Add an isinstance(config, dict) check (and a clear ValueError) after extracting key to enforce the expected shape.

    if key is not None:
        if key not in config:
            msg = f"Configuration must contain '{key}' key. Got keys: {list(config.keys())}"
            raise ValueError(msg)
        config = config[key]

    if "class_path" in config:
        cls = _import_class(config["class_path"])

src/physicalai/config/instantiate.py:115

• yaml.safe_load() can return None (empty file) or a non-dict top-level value; passing that directly into instantiate_obj_from_dict() will crash with a non-obvious TypeError (e.g. 'NoneType' is not iterable). Consider validating that the loaded config is a dict (or defaulting None to {}) and raising a helpful error if not.

    """Instantiate an object from a YAML/JSON configuration file."""
    with Path(file_path).open("r", encoding="utf-8") as f:
        config = yaml.safe_load(f)
    return instantiate_obj_from_dict(config, key=key, target_cls=target_cls)

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[samet-akcay]

Just reading through https://jsonargparse.readthedocs.io/en/stable/ and feel like we could re-write all of this stuff using just jsonargparse