curobo.profiling module

Profiling utilities for cuRobo.

Lightweight timing helpers for measuring GPU work. Timer behavior is gated by curobo.runtime.cuda_event_timers; when disabled, timers are no-ops and return 0.0 with minimal overhead.

Example

from curobo.profiling import CudaEventTimer

timer = CudaEventTimer().start()
# ... GPU work ...
elapsed_seconds = timer.stop()
class CudaEventTimer

Bases: object

Lightweight CUDA event-based timer for accurate GPU timing.

Uses CUDA events instead of CPU time + synchronize for more accurate GPU timing without blocking the entire device.

Example

>>> timer = CudaEventTimer().start()
>>> # ... GPU work ...
>>> elapsed_seconds = timer.stop()

The timer respects runtime.cuda_event_timers. When disabled, stop() returns 0.0 with minimal overhead.

Initialize the timer.

__init__()

Initialize the timer.

start(stream=None)

Start the timer by recording a CUDA event.

Parameters:

stream (torch.cuda.streams.Stream | None) – Optional CUDA stream on which to record the event.

Returns:

Self for method chaining.

Return type:

curobo._src.util.cuda_event_timer.CudaEventTimer

checkpoint(
name,
stream=None,
)

Record a named, non-synchronizing interval boundary.

The elapsed interval ends at this checkpoint and starts at the previous checkpoint (or start). Reusing a name accumulates repeated intervals, which is useful for iterative algorithms. Values become available through intervals after stop.

Parameters:

  • name (str) – Stable name for the interval ending at this checkpoint.

  • stream (torch.cuda.streams.Stream | None) – Optional CUDA stream on which to record the event.

Returns:

Self for method chaining.

Return type:

curobo._src.util.cuda_event_timer.CudaEventTimer

stop(
final_interval=None,
stream=None,
)

Stop the timer and return elapsed time in seconds.

This synchronizes only on the end event, not the entire device.

Parameters:

  • final_interval (str | None) – Optional name for the interval from the last checkpoint (or start) through the stop event.

  • stream (torch.cuda.streams.Stream | None) – Optional CUDA stream on which to record the event.

Returns:

Elapsed time in seconds, or 0.0 if timing is disabled.

Return type:

float

property intervals: Dict[str, float]

Return named interval durations in seconds after stop.