Enum GraphPhase 

pub enum GraphPhase {
    Prefill,
    Decode,
}

Which half of the run is asking.

The phase exists because the graph is plausibly two decisions, not one — but on the hardware measured so far it is only ever a decode decision. On an Adreno 642L with bonsai-1.7b, from identical clean starts and two repeats each: decode 11.6 tok/s without it and 0.72 with, while prefill is 4.2 either way. A first reading of 3.4 -> 18.0 for prefill did not survive a controlled re-run — it was a dirty probe cache between configurations, not the graph, and the prefill route through the graph is GDN-only in the first place, which this dense model never takes.

Variants

Prefill

Decode

Trait Implementations

impl Clone for GraphPhase

fn clone(&self) -> GraphPhase

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for GraphPhase

impl Debug for GraphPhase

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

Formats the value using the given formatter.

impl Eq for GraphPhase

impl PartialEq for GraphPhase

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

Equality operator ==.

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

Inequality operator !=.

impl StructuralPartialEq for GraphPhase