Struct PendingLogits 

pub struct PendingLogits { /* private fields */ }

M2 increment 3: a step’s logits, still device-resident on the LAST stage. wait() orders the readback stream behind the step’s completion event, copies, and syncs — tokens enqueued after this step keep running on the stage streams while the caller drains token t. Dropping without waiting is safe (buffers free stream-ordered).

Implementations

impl PendingLogits

pub fn new(logits: CudaSlice<f32>, ev: CudaEvent, rb: Arc<CudaStream>) -> Self

pub fn wait(self) -> Result<Vec<f32>, Box<dyn Error>>

Blocks until this step’s logits are computed, returns them host-side. Only this step’s work is waited on (event-ordered) — NOT later tokens already enqueued on the stage streams.