# encodings

Python API: `modelconverter.utils.encodings`

Parsing of externally supplied quantization encodings.

Quantization encodings describe the scale, offset and bit width used to quantize activations and parameters of a model. They can
be handed to the RVC4 conversion to override the values derived from calibration. The files come in several shapes, so this module
normalizes them into the `Encodings` model used by the conversion configuration.

## Functions

### parse_encodings

```python
def parse_encodings(value: ParamValue | Encodings) -> Encodings:
```

Parse quantization encodings into an `Encodings` model.

Accepts an `Encodings` instance, which is returned unchanged, a JSON string, or a dictionary with `activation_encodings` and
`param_encodings` keys. Each of the two groups may either be a mapping from tensor name to entries, or a list of entries carrying
a `name` field. Entries are normalized on the way: the legacy `bw` and `is_sym` keys are renamed to `bitwidth` and `is_symmetric`,
the `dtype` is lower-cased, unknown keys are dropped, single-element lists are unwrapped and per-channel entries are expanded into
one entry per channel.

Parameters

 * `value` (`ParamValue | Encodings`): Encodings to parse.

Returns

 * `Encodings`: Parsed encodings.

Raises

 * `TypeError`: If `value` is not a dictionary once deserialized, if a group is neither a list nor a dictionary, or if an
   individual entry is not a dictionary.
 * `ValueError`: If an entry in a list-shaped group has a missing or invalid tensor name, or if the per-channel fields of an entry
   have different lengths.

### validate_quantization_override_names

```python
def validate_quantization_override_names(encodings: Encodings, model_path: str | Path):
```

Reject override names that are absent or in the wrong encoding group.

## Attributes

### ALLOWED_ENCODING_KEYS