Struct Config 

pub struct Config {

    pub hidden_size: usize,
    pub intermediate_size: usize,
    pub vocab_size: usize,
    pub num_hidden_layers: usize,
    pub num_attention_heads: usize,
    pub num_key_value_heads: usize,
    pub rms_norm_eps: f64,
    pub rope_theta: f32,
    pub bos_token_id: Option<u32>,
    pub eos_token_id: Option<EosTokenId>,
    pub rope_scaling: Option<RopeScaling>,
    pub tie_word_embeddings: bool,
    pub max_seq_len: usize,
    pub use_qkv_bias: bool,
    pub model_prefix: String,
    pub head_dim: Option<usize>,
    pub partial_rotary_factor: f32,
    pub linear_attn: Option<LinearAttnConfig>,
    pub residual_rms_norm: bool,
}

Generalized LLM configuration shared by all decoder-only text models.

Fields

hidden_size: usize

intermediate_size: usize

vocab_size: usize

num_hidden_layers: usize

num_attention_heads: usize

num_key_value_heads: usize

rms_norm_eps: f64

rope_theta: f32

bos_token_id: Option<u32>

eos_token_id: Option<EosTokenId>

rope_scaling: Option<RopeScaling>

tie_word_embeddings: bool

max_seq_len: usize

use_qkv_bias: bool

Whether Q/K/V projections use bias (true for Qwen2, false for LLaMA).

model_prefix: String

Weight tensor prefix for the transformer stack (e.g. “model” or “model.language_model”).

head_dim: Option<usize>

Explicit head dimension when it differs from hidden_size / num_attention_heads.

partial_rotary_factor: f32

Fraction of head dims to apply rotary embeddings to (1.0 = all, 0.25 = first quarter).

linear_attn: Option<LinearAttnConfig>

Linear attention configuration (None for pure softmax-attention models).

residual_rms_norm: bool

Whether RMS norm uses residual weight: (1 + weight) * norm(x) instead of weight * norm(x). True for Qwen3.5 whose norm weights are initialized to zero with +1 applied at runtime.

Trait Implementations

impl Clone for Config

fn clone(&self) -> Config

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Config

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

Formats the value using the given formatter.