Enum KimiExpertBacking 

pub enum KimiExpertBacking {
    Resident(Vec<KimiExpertWeights>),
    Stored {
        store: Arc<ExpertStore<KimiExpertSource>>,
        layout: KimiStoredExpertLayout,
        n_experts: usize,
        layer: u32,
    },
}

How a Kimi layer’s routed experts are held: Resident is the original always-constructed form (zero-copy MXFP4 mmap views); Stored holds only the per-layer byte layout and materializes one expert at a time from a bounded, lease-protected store shared by every layer – same design (and same bit-equivalence argument) as the GGUF path’s ferrox_models::decoder::ExpertBacking. Every expert in a Kimi layer has identical dims, so one layout serves the whole layer.

Variants

Resident(Vec<KimiExpertWeights>)

Stored

Fields

store: Arc<ExpertStore<KimiExpertSource>>

layout: KimiStoredExpertLayout

n_experts: usize

layer: u32

Implementations

impl KimiExpertBacking

pub fn n_experts(&self) -> usize

pub fn with_expert<R>( &self, e: usize, f: impl FnOnce(&KimiExpertWeights) -> R, ) -> R

Runs f against expert e, materializing it from the store first when store-backed (the lease pins the cache entry for exactly f’s borrow).