Struct ModelSpec 

pub struct ModelSpec {

    pub arch: String,
    pub d: usize,
    pub vocab: usize,
    pub ctx_len: usize,
    pub rms_eps: f64,
    pub embedding_scale: Option<f64>,
    pub tied_output: bool,
    pub final_logit_softcap: Option<f64>,
    pub layers: Vec<Layer>,
    pub mtp_blocks: usize,
    pub ffn_norm: &'static str,
    pub hyper: Option<HyperConnections>,
    pub ple: Option<Ple>,
    pub indexer: Option<Indexer>,
}

What a checkpoint computes.

Fields

arch: String

general.architecture.

d: usize

Hidden width.

vocab: usize

Vocabulary size (rows of the token embedding).

ctx_len: usize

The trained context window.

rms_eps: f64

The RMSNorm epsilon every norm of the model uses.

embedding_scale: Option<f64>

A scale applied to the token embedding as it is read (Gemma: √d).

tied_output: bool

The output projection is the token embedding (no output.weight in the file).

final_logit_softcap: Option<f64>

Soft-capping of the final logits (cap · tanh(x / cap)).

layers: Vec<Layer>

mtp_blocks: usize

Draft blocks for multi-token prediction appended after the model’s layers.

ffn_norm: &'static str

The tensor stem of the norm on each block’s FFN input: ffn_norm, or post_attention_norm in Qwen3.5, whose converter names the pre-FFN norm after the attention it follows.

hyper: Option<HyperConnections>

The residual stream’s hyper-connections, in place of every norm (Qwen3.8-Flash-Next); None for one residual.

ple: Option<Ple>

The hashed n-gram embedding one layer’s input takes (Qwen3.8-Flash-Next’s per-layer embedding).

indexer: Option<Indexer>

The sparse selection of the positions every attention layer’s queries attend (Qwen’s QSA).

Implementations

impl ModelSpec

pub fn from_gguf(g: &Gguf) -> Result<ModelSpec, SpecError>

Read the architecture of a GGUF checkpoint.

pub fn from_source<S: SpecSource + ?Sized>( g: &S, ) -> Result<ModelSpec, SpecError>

Read the architecture of a checkpoint from anything that describes it (SpecSource).

pub fn attention_layers(&self) -> Vec<usize>

The indices of the attention layers.

Trait Implementations

impl Clone for ModelSpec

fn clone(&self) -> ModelSpec

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for ModelSpec

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

Formats the value using the given formatter.

impl PartialEq for ModelSpec

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

Tests for self and other values to be equal, and is used by ==.

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

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl StructuralPartialEq for ModelSpec