Struct ModelMetadata 

pub struct ModelMetadata {

    pub architecture: String,
    pub hidden_size: usize,
    pub intermediate_size: usize,
    pub num_hidden_layers: usize,
    pub num_attention_heads: usize,
    pub num_key_value_heads: usize,
    pub vocab_size: usize,
    pub max_position_embeddings: usize,
    pub rms_norm_eps: f64,
    pub rope_theta: f64,
    pub tie_word_embeddings: bool,
    pub head_dim: usize,
    pub attention_bias: bool,
    pub bos_token_id: Option<u32>,
    pub eos_token_ids: Vec<u32>,
    pub vision: Option<VisionConfig>,
    pub attention_pattern: AttentionPattern,
    pub activation: Activation,
    pub rope_scaling: RopeScaling,
}

Architecture + hyperparameter description of a model, format-agnostic.

Field names follow the HuggingFace Llama config convention; other architecture families remap their config onto this struct in their adapter.

Fields

architecture: String

Architecture identifier, e.g. "llama", "smollm2" (from config.json::model_type). The model registry keys on this string.

hidden_size: usize

Hidden size (model dimension).

intermediate_size: usize

MLP intermediate size.

num_hidden_layers: usize

Number of transformer layers.

num_attention_heads: usize

Number of attention query heads.

num_key_value_heads: usize

Number of key/value heads (GQA). Equal to num_attention_heads for MHA.

vocab_size: usize

Vocabulary size.

max_position_embeddings: usize

Maximum positional embeddings the model was built for.

rms_norm_eps: f64

RMSNorm epsilon.

rope_theta: f64

RoPE base frequency (theta).

tie_word_embeddings: bool

Whether lm_head is tied to the embedding matrix.

head_dim: usize

Per-head dimension, derived: hidden_size / num_attention_heads.

attention_bias: bool

Whether attention projections carry biases.

bos_token_id: Option<u32>

Beginning-of-sequence token id, if defined.

eos_token_ids: Vec<u32>

End-of-sequence token ids (merged from config + generation_config).

vision: Option<VisionConfig>

Vision-tower hyperparameters for multimodal models (Idefics3/SmolVLM today); None for text-only models.

attention_pattern: AttentionPattern

Layer-type attention pattern (Gemma sliding-window interleave); defaults to all-global (Llama-family behavior).

activation: Activation

MLP activation (hidden_act / hidden_activation).

rope_scaling: RopeScaling

RoPE frequency scaling (rope_scaling); None variant when absent.

Implementations

impl ModelMetadata

pub fn diffusion_placeholder(architecture: &str) -> Self

Minimal placeholder metadata for diffusion components that do not carry a language-model config.json (UNet, VAE, etc.).

pub fn from_hf_config( config: &Value, generation_config: Option<&Value>, ) -> Result<Self>

Parses metadata from a HuggingFace config.json value, optionally merged with a generation_config.json value (which can override/add bos/eos ids).

Trait Implementations

impl Clone for ModelMetadata

fn clone(&self) -> ModelMetadata

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for ModelMetadata

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

Formats the value using the given formatter.