Struct Qwen3Config 

pub struct Qwen3Config {

    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 head_dim: 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 qk_norm: bool,
    pub sliding_window: Option<usize>,
    pub max_window_layers: usize,
    pub use_sliding_window: bool,
    pub num_experts: usize,
    pub num_experts_used: usize,
    pub expert_ffn_size: usize,
    pub shared_expert_ffn_size: usize,
    pub expert_weights_scale: f32,
}

Fields

vocab_size: usize

hidden_size: usize

intermediate_size: usize

num_hidden_layers: usize

num_attention_heads: usize

num_key_value_heads: usize

head_dim: usize

max_position_embeddings: usize

rms_norm_eps: f64

rope_theta: f64

hidden_act: String

tie_word_embeddings: bool

attention_bias: bool

qk_norm: bool

Whether the model uses per-head RMS-norm on Q/K before RoPE (a.k.a. “QK-norm”). Qwen 3 has it; Qwen 2 does NOT. Defaults to true to match the historical Qwen 3 build path.

sliding_window: Option<usize>

Sliding-window size; None (or absent) means full causal.

max_window_layers: usize

Number of leading layers that use full causal attention; layers [max_window_layers, num_hidden_layers) use sliding window when use_sliding_window is true. HF default: all layers full.

use_sliding_window: bool

num_experts: usize

Total number of routed experts per MoE layer (0 = dense model). HF key: num_experts / n_routed_experts. GGUF key: qwen3.expert_count.

num_experts_used: usize

Number of experts activated per token (top-k routing). HF key: num_experts_per_tok. GGUF key: qwen3.expert_used_count.

expert_ffn_size: usize

FFN inner width for each routed expert. When 0 falls back to intermediate_size / num_experts_used to match upstream defaults. GGUF key: qwen3.expert_feed_forward_length.

shared_expert_ffn_size: usize

FFN inner width for the always-on shared expert (0 = no shared expert). GGUF key: qwen3.expert_shared_feed_forward_length.

expert_weights_scale: f32

Multiplier applied to routed-expert logits before softmax (default 1.0). GGUF key: qwen3.expert_weights_scale.

Implementations

impl Qwen3Config

pub fn from_file(path: &Path) -> Result<Qwen3Config, Error>

pub fn kv_group_size(&self) -> usize

Repetition factor for GQA: how many Q heads share each KV head.

pub fn q_proj_dim(&self) -> usize

Q projection output width (num_attention_heads * head_dim).

pub fn kv_proj_dim(&self) -> usize

K/V projection output width (num_key_value_heads * head_dim).

pub fn is_moe(&self) -> bool

True when the config carries MoE routing (num_experts > 0). qwen3-30b-a3b-instruct and qwen3-coder-next MoE variants will return true; dense Qwen3 returns false.

pub fn expert_ffn_dim(&self) -> usize

Routed-expert SwiGLU inner width. Falls back to intermediate_size / num_experts_used when the explicit expert_ffn_size is absent (matches upstream defaults).

pub fn shared_expert_ffn_dim(&self) -> usize

Shared-expert SwiGLU inner width (0 when there’s no shared expert).

pub fn layer_uses_swa(&self, idx: usize) -> bool

Does layer idx use sliding-window attention?

Trait Implementations

impl Clone for Qwen3Config

fn clone(&self) -> Qwen3Config

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Qwen3Config

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

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for Qwen3Config

fn deserialize<__D>( __deserializer: __D, ) -> Result<Qwen3Config, <__D as Deserializer<'de>>::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.