pyserq.Stage

One stage's report row. Obtain it from Report.stage, Report.stages_named or Report.stages. An array has a row per member. Stage() is not a public constructor; attributes are read-only.

Attributes

Attribute	Type	Description
name	str	Stage declaration name.
index	int or None	Array member index; None for a scalar stage.
mean_number	float	Time-average jobs present, waiting or in service.
utilization	float	Fraction of measured time with at least one job; for shared stages, average fraction of capacity carried by flows.
completed	int	Jobs completed after warm-up.
throughput	float	Completions per measured clock unit.
mean_wait	float	Mean wait before service.
mean_service	float	Mean time in service.

Step-engine statistics

Attribute	Type	Description
iterations	int	Iterations over the whole run, including warm-up; 0 on other stage kinds.
prefill_only	float	Fraction of measured time running prefill-only iterations.
decode_only	float	Fraction running decode-only iterations.
mixed	float	Fraction running mixed prefill/decode iterations.
mean_decodes	float	Time-average number of decoders in the running iteration, 0 while idle.
mean_decode_batch	float	Mean decoder count over decode-carrying iterations started after warm-up.
mean_decode_step	float	Mean duration of those decode-carrying iterations.
mean_itl	float	Mean gap between successive tokens of a turn ending on this stage after warm-up.
itl_p50	float	Median inter-token gap, within 0.5%.
itl_p99	float	99th percentile inter-token gap, within 0.5%.
idle_with_work	bool	The run ended with work waiting and the last iteration attempt scheduling nothing.

The batch-time fractions and mean_decodes are 0 for other stage kinds. Decode-batch and inter-token statistics are NaN when there are no matching iterations or gaps. A token gap can include transfer, queueing or preemption time. See stage statistics.

Examples

import pyserq

report = pyserq.run(pyserq.compile(source="""
fn main() {
  stage svc[2] : fifo;
  workload {
    arrive batch(1);
  }
  server {
    run svc[1] (cost(svc, 2));
  }
}
""", horizon=10))
rows = report.stages_named("svc")
assert [row.index for row in rows] == [0, 1]
assert rows[1].completed == 1

See also

Report, Pool, stage declarations.