class

Profiler

Profiler(with_memory: bool = True)

source

Record op-level timing and memory traces from the C++ engine.

Drives the engine's profiling hook so that every dispatched op becomes an OpEvent. Suitable for both interactive use (with block) and programmatic control via explicit start / stop. Recorded events can be exported as a Chrome / Perfetto trace for visual inspection.

Parameters

with_memorybool= True

If True (default), also capture allocator stats via MemoryStats.

Notes

Total recorded time can be summarised per op as

T_{\text{op}} = \sum_{i : \text{name}(e_i) = \text{op}} t_i,

via key_averages. Only one Profiler may be active at a time; entering a nested instance overwrites the engine's current profiler pointer until the outer context exits.

Examples

>>> import lucid
>>> with lucid.profiler.profile() as prof:
...     y = lucid.randn(64, 64) @ lucid.randn(64, 64)
>>> top = prof.key_averages()[:5]

Methods (10)

dunder

init

→None

__init__(with_memory: bool = True)

source

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

start

→None

start()

source

Start recording operations.

stop

→None

stop()

source

Stop recording operations.

events

→list[OpEvent]

events()

source

Return all recorded OpEvents.

key_averages

→list[ProfileSummary]

key_averages()

source

Return per-operation-name summaries sorted by total time.

memory_stats

→MemoryStats | None

memory_stats()

source

Return memory stats if available.

export_chrome_trace

→None

export_chrome_trace(path: str)

source

Export events as a Chrome trace JSON file.

Open with chrome://tracing or https://ui.perfetto.dev.

clear

→None

clear()

source

Clear all recorded events.

dunder

enter

→Profiler

__enter__()

source

Enter the context. Returns self so the value can be bound via with ... as.

dunder

exit

→None

__exit__(args: object = ())

source

Exit the context, restoring any state that was modified on entry.