class

ProfileSummary

ProfileSummary(name: str, events: list[OpEvent])

source edit

Aggregated timing summary for a single op name across many calls.

Produced by Profiler.key_averages, which buckets the recorded OpEvent list by op name and exposes total time, average time, call count, and cumulative FLOP estimate. This is the natural object for a "top-k by self-time" performance report.

Parameters

namestr

Op name shared by all events in the summary.

eventslist of OpEvent

Recorded events bucketed under name.

Notes

The average is computed as

\bar{t} = \frac{1}{N} \sum_{i=1}^{N} t_i,

where $N$ is the call count and $t_i$ is the per-call duration in microseconds.

Examples

>>> import lucid
>>> with lucid.profiler.profile() as prof:
...     for _ in range(10):
...         y = lucid.randn(64, 64) @ lucid.randn(64, 64)
>>> for summary in prof.key_averages()[:3]:
...     print(summary)

Used by 1

lucid.autograd.profiler

Constructors

dunder

init

→None

__init__(name: str, events: list[OpEvent])

source edit

Initialise the instance. See the class docstring for parameter semantics.

Properties

prop

avg_time_us

→float

avg_time_us: float

source edit

Average time in microseconds per call.

prop

count

→int

count: int

source edit

Number of times this operation was called.

prop

total_flops

→int

total_flops: int

source edit

Total FLOPs across all calls.

prop

total_time_us

→float

total_time_us: float

source edit

Total time in microseconds across all calls.

Dunder methods

dunder

repr

→str

__repr__()

source edit

Return a developer-facing string representation of the instance.

>>> import lucid >>> with lucid.profiler.profile() as prof: ... for _ in range(10): ... y = lucid.randn(64, 64) @ lucid.randn(64, 64) >>> for summary in prof.key_averages()[:3]: ... print(summary)