Struct Quantization 

pub struct Quantization {
    pub bits: Option<u8>,
    pub scheme: QuantScheme,
    pub group_size: Option<u32>,
    pub label: String,
}

Structured quantization descriptor.

This replaces a free-text string that mixed two vocabularies — MLX’s 4bit/6bit and GGUF’s Q4_K_M/Q8_0 — under one field, so nothing could tell a group-quantized integer checkpoint from a mixed k-quant one. It also threw away what the ingest paths already had in hand: MLX config.json declares {bits, group_size, mode} and only bits survived.

Deserializes from either the legacy bare string or the structured object, so catalogs, user models.json files, and models.register payloads written before this change keep loading unchanged. The object form requires label and rejects unknown fields: a misspelled key is a producer bug, and accepting it silently is how the field it replaced lost information in the first place.

Writes the bare label back whenever that is lossless, and the object only when it carries something parsing cannot recover — a group_size, or a scheme that disambiguates a label the parser reads as Unknown. Two reasons, both about blast radius rather than taste. This value is inside catalog_identity::row_digest, so an unconditional shape change would move the digest of every quantized row and hard-fail any client that pinned expected_catalog_revision. And registry::load_user_config fails a models.json whole, with its only production caller discarding the error — so an older daemon meeting an object it cannot parse boots with zero registered models and says nothing. Emitting the object only where it adds information keeps both costs proportional to what actually changed.

Fields

bits: Option<u8>

Weight bit width. None when the label names none.

scheme: QuantScheme

Which quantization family this is.

group_size: Option<u32>

Elements sharing one scale/zero point (MLX: 32, 64, 128). None when the scheme has no group concept, or the source declared none.

label: String

The label exactly as published. Never synthesized: it is what the user sees, what Hugging Face repos are named after, and the only thing that survives a scheme CAR does not recognize yet.

Implementations

impl Quantization

pub fn parse(label: &str) -> Self

Best-effort structure from a published label.

Never fails and never invents: an unrecognized label yields QuantScheme::Unknown with the label intact.

pub fn from_gguf_filename(filename: &str) -> Option<Self>

Recover the quantization a GGUF file names in its own filename (Qwen3-8B-Q4_K_M.gguf, ggml-large-v3-turbo-q5_0.gguf).

Scans the hyphen-separated segments from the right and takes the first that classifies, so a model whose name contains something quant-shaped does not outrank the real suffix. Returns None rather than guessing when nothing in the name is recognizable.

pub fn from_mlx_config( bits: Option<u8>, group_size: Option<u32>, mode: Option<&str>, ) -> Option<Self>

Descriptor for an MLX checkpoint, from the quantization block of its config.json. mode is MLX’s own name for the format (affine, mxfp4, mxfp8); absent means affine, which is MLX’s default.

Returns None when the block carried nothing usable — declaring AffineGroupInt on the strength of an empty object would assert exactly the thing this type exists to establish.

Trait Implementations

impl Clone for Quantization

fn clone(&self) -> Quantization

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for Quantization

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for Quantization

fn deserialize<D: Deserializer<'de>>(d: D) -> Result<Self, D::Error>

Deserialize this value from the given Serde deserializer.

impl Display for Quantization

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Eq for Quantization

impl PartialEq for Quantization

fn eq(&self, other: &Quantization) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl Serialize for Quantization

fn serialize<S: Serializer>(&self, serializer: S) -> Result<S::Ok, S::Error>

Serialize this value into the given Serde serializer.

impl StructuralPartialEq for Quantization