Struct StepEpExps 

pub struct StepEpExps {
    pub runtime: Arc<TpE4m3HostBounce>,
    pub experts: StepEpExpertBank,
    pub devices: Vec<usize>,
    pub configured_by_tp: bool,
    pub activation_limit: Option<f32>,
    pub grouped_decode: Option<Mutex<StepEpGroupedDecode>>,
}

Fields

runtime: Arc<TpE4m3HostBounce>

experts: StepEpExpertBank

devices: Vec<usize>

configured_by_tp: bool

activation_limit: Option<f32>

grouped_decode: Option<Mutex<StepEpGroupedDecode>>

Persistent one-token grouped projection/combine state for eager decode. Opt-in prefill uses the model-scoped executor instead of multiplying capacity workspaces per layer.