Struct PlacementInputs 

pub struct PlacementInputs {
    pub budget: u64,
    pub non_expert: u64,
    pub per_expert: u64,
    pub moe_layers: usize,
    pub n_experts: usize,
    pub top_k: usize,
    pub attn_per_layer: u64,
    pub graph_prefix: bool,
}

Everything the placement decision reads — plain data, so the policy is testable without a device.

Fields

budget: u64

Device weight budget in bytes (CMF_GPU_VRAM_MB or VRAM − reserve).

non_expert: u64

Device bytes of every non-expert weight (attention, norms, routers, dense layers, lm_head) — they must stay resident in every mode.

per_expert: u64

Bytes of one expert (gate + up + down).

moe_layers: usize

n_experts: usize

top_k: usize

attn_per_layer: u64

Mean non-expert device bytes of one layer (attention, router, norms).

graph_prefix: bool

A whole-token device graph can run whole resident layers of this model (then a whole-layer prefix costs one submit, not per-layer fences).

Trait Implementations

impl Clone for PlacementInputs

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for PlacementInputs

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

Formats the value using the given formatter.