pub struct WorkerPhaseTimings {
pub forward_wall_ms: f64,
pub backward_wall_ms: f64,
pub fwd_setup_ms: f64,
pub bwd_setup_ms: f64,
}Expand description
Per-worker timing payload for TrainingEvent::WorkerTiming.
Names the four populated values explicitly. Replaces the previous
[f64; 16] payload where 12 of 16 slots were always zero on per-worker
events. WorkerPhaseTimings is 4 × f64 = 32 bytes; the previous
payload was 128 bytes (75% waste).
On Forward-phase events, forward_wall_ms and fwd_setup_ms are
populated; the backward fields are 0. On Backward-phase events,
backward_wall_ms and bwd_setup_ms are populated; the forward fields
are 0.
The writer adapter in training_output.rs maps these four fields into the
unchanged 16-wide cobre_io::WorkerTimingRecord via the
WORKER_TIMING_SLOT_* constants, preserving the Parquet output schema.
Fields§
§forward_wall_ms: f64Forward-pass wall time in ms (populated on Forward; 0 on Backward).
backward_wall_ms: f64Backward-pass wall time in ms (populated on Backward; 0 on Forward).
fwd_setup_ms: f64Forward setup time in ms (populated on Forward; 0 on Backward).
bwd_setup_ms: f64Backward setup time in ms (populated on Backward; 0 on Forward).
Trait Implementations§
Source§impl Clone for WorkerPhaseTimings
impl Clone for WorkerPhaseTimings
Source§fn clone(&self) -> WorkerPhaseTimings
fn clone(&self) -> WorkerPhaseTimings
1.0.0 · Source§fn clone_from(&mut self, source: &Self)
fn clone_from(&mut self, source: &Self)
source. Read more