Struct Moe 

pub struct Moe {
    pub experts: usize,
    pub top_k: usize,
    pub expert_dff: usize,
    pub act: Act,
    pub gating: Gating,
    pub norm_top_k: bool,
    pub weights_scale: f64,
    pub shared_experts: usize,
    pub shared_dff: usize,
    pub shared_gate: bool,
    pub biases: bool,
    pub groups: Option<(usize, usize)>,
}

A mixture of experts.

Fields

experts: usize

top_k: usize

expert_dff: usize

act: Act

gating: Gating

norm_top_k: bool

The chosen experts’ weights are renormalized to sum to one.

weights_scale: f64

A scale applied to the routing weights.

shared_experts: usize

shared_dff: usize

shared_gate: bool

The shared experts’ output scaled by σ(x·g), g the checkpoint’s ffn_gate_inp_shexp (Qwen’s mixtures).

biases: bool

Biases on the router and on every expert’s projections (gpt-oss).

groups: Option<(usize, usize)>

Experts chosen within groups (DeepSeek V3): the experts split into groups of equal size, each group scored by its two best experts, only the best used groups’ experts eligible for the top-k.

Trait Implementations

impl Clone for Moe

fn clone(&self) -> Moe

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Moe

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

Formats the value using the given formatter.

impl PartialEq for Moe

fn eq(&self, other: &Moe) -> 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 Moe