Struct Llama32Config 

pub struct Llama32Config {

    pub vocab_size: usize,
    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 max_position_embeddings: usize,
    pub rms_norm_eps: f64,
    pub rope_theta: f64,
    pub hidden_act: String,
    pub tie_word_embeddings: bool,
    pub attention_bias: bool,
    pub head_dim: Option<usize>,
    pub rope_scaling: Option<Llama32RopeScaling>,
    pub rope_style: RopeStyle,
    pub gguf_arch: Option<String>,
    pub rope_dim: Option<usize>,
}

Fields

vocab_size: usize

hidden_size: usize

intermediate_size: usize

num_hidden_layers: usize

num_attention_heads: usize

num_key_value_heads: usize

max_position_embeddings: usize

rms_norm_eps: f64

rope_theta: f64

hidden_act: String

tie_word_embeddings: bool

attention_bias: bool

head_dim: Option<usize>

Explicit head dim (Llama 3.x); when absent, derived from hidden/heads.

rope_scaling: Option<Llama32RopeScaling>

rope_style: RopeStyle

RoPE pairing flavor. GGUF Llama weights are permuted by the HF→GGUF converter for llama.cpp’s interleaved (NORM) RoPE, so GGUF-backed inference must rotate with rlx_ir::RopeStyle::GptJ; HF-safetensors checkpoints use rlx_ir::RopeStyle::NeoX (default). Not present in HF config.json, so skipped during deserialization.

gguf_arch: Option<String>

GGUF general.architecture tag when loaded from GGUF (llama, phi3, …).

rope_dim: Option<usize>

Rotary dimension when it differs from [head_dim] (Phi-3 partial RoPE).

Implementations

impl Llama32Config

pub fn from_file(path: &Path) -> Result<Llama32Config, Error>

pub fn from_gguf(raw: &GgufFile) -> Result<Llama32Config, Error>

pub fn head_dim(&self) -> usize

pub fn kv_group_size(&self) -> usize

pub fn q_proj_dim(&self) -> usize

pub fn kv_proj_dim(&self) -> usize

pub fn n_rot(&self) -> usize

Leading per-head dims that receive RoPE (equals [head_dim] for Llama; may be smaller for Phi-3 partial RoPE).

pub fn uses_partial_rope(&self) -> bool

pub fn is_phi_arch(&self) -> bool

Trait Implementations

impl Clone for Llama32Config

fn clone(&self) -> Llama32Config

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Llama32Config

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

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for Llama32Config

fn deserialize<__D>( __deserializer: __D, ) -> Result<Llama32Config, <__D as Deserializer<'de>>::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.