Struct LayerRoutingSkew 

pub struct LayerRoutingSkew {
    pub layer: u32,
    pub routed: u64,
    pub working_set: usize,
    pub experts_for_90pct: usize,
    pub norm_entropy: f64,
    pub oracle_hit_at_slots: f64,
}

One MoE layer’s routing concentration over the observed decode histogram.

Every field is derived from that layer’s histogram row alone, so a layer that was never routed to has no entry at all rather than a row of zeros – a zero working set and a zero oracle hit rate would read as “this layer is unservable” when the truth is “no data”.

Fields

layer: u32

routed: u64

Total routes observed for this layer, counting duplicates.

working_set: usize

Distinct experts that were routed to at least once.

experts_for_90pct: usize

How many of the hottest experts it takes to cover 90% of this layer’s routing mass.

norm_entropy: f64

Routing entropy over ln(num_experts). 1.0 is a perfectly flat router, 0.0 a layer that always picks the same expert.

oracle_hit_at_slots: f64

The top-oracle_slots share of this layer’s routing mass: the best hit rate any per-layer policy could reach on the observed distribution, with no LRU/LFU dynamics in it at all.

This is the number that separates the two diagnoses. A high oracle hit rate next to a low realized one (ExpertCacheStats::miss_rate) means the policy is losing residency it could have kept; a low oracle hit rate means the layer’s routing is flat and no cache size will help it.

Trait Implementations

impl Clone for LayerRoutingSkew

fn clone(&self) -> LayerRoutingSkew

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for LayerRoutingSkew

impl Debug for LayerRoutingSkew

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

Formats the value using the given formatter.

impl PartialEq for LayerRoutingSkew

fn eq(&self, other: &LayerRoutingSkew) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for LayerRoutingSkew