load_config(path, config_cls) reads a YAML document and constructs the requested configuration class. It is an intentionally strict field-name gate, not a generic nested-schema parser or automatic type coercer. Consumers include model and training CLI entry points that pass their dataclass definitions.
Read and parse the mapping
The signature and file read appear at source L11–L18.
with path.open("r", encoding="utf-8") as f:
data = yaml.safe_load(f)Here the snippet is whitespace-normalized for display; the original uses tabs for indentation. Path(path) admits string or Path inputs, safe_load parses standard YAML values, and non-mapping top-level data raises a ValueError. File and YAML syntax errors are not silently caught or rewritten.
Demand an exact dataclass field set
The decisive set operations are source L20–L34.
expected = {field.name for field in fields(config_cls)}
provided = set(data)
missing = expected - provided
unknown = provided - expectedfields(config_cls) enumerates the declared dataclass fields. If even one field is absent, the function raises with the sorted missing names. If extra keys are present, it separately reports the sorted unknown names. Dataclass defaults are not treated as permission to omit keys: the comparison requires exact equality of the two name sets. Python raises the underlying TypeError if config_cls is not a supported dataclass type.
Construction is the final validation stage
Only after key checking does source L36–L36 execute return config_cls(**data). Values are passed through as parsed, without manual annotations inspection or coercion. Further range, membership and cross-field validation belongs to the dataclass’s constructor or __post_init__ (see KimiK3Config in model.py and TrainConfig in trainer.py). A provided YAML list can be normalized later by its target configuration class; load_config itself does not normalize it.
Verification and limitations
The integration checks in test test_moe_grouped_vs_naive.py L208–L223 assert that a valid saved configuration round-trips and that unknown and missing keys are rejected. The configuration unit tests test test_config.py L17–L38 exercise model-level constraints after construction, not YAML parsing per se. These assertions were inspected but not run in this documentation session. There is no recursive schema validation or protection against a caller passing an unsuitable class.