Cookbook — Profiling

1. Time a region with the Profiler

Intent: Measure wall-clock time of a code block.

from SneppX_ALG import Profiler, Timer

prof = Profiler(enabled=True)
with Timer(prof, "forward"):
    out = model(x)              # recorded under "forward"
with Timer(prof, "backward"):
    loss.backward()

prof.print_summary()
# Operation                          Calls    Total(s)   Avg(s)   Min(s)   Max(s)
# backward                                1     0.2456   0.2456   0.2456   0.2456
# forward                                 1     0.1123   0.1123   0.1123   0.1123

Notes: Timer is a context manager around Profiler.record. CPU-safe.

2. Decorate a function with @timeit

Intent: Profile a function call by name.

from SneppX_ALG import get_profiler, timeit

prof = get_profiler(); prof.enabled = True

@timeit(prof)
def train_step(x, y):
    return (model(x) - y).pow(2).mean()

for batch in loader:
    train_step(*batch)
prof.print_summary()

Notes: timeit is the global-profiler equivalent of torch.profiler's record_section. The decorator appends to the global _GLOBAL_PROFILER.

3. Track memory usage

Intent: Detect leaks and peak consumption.

from SneppX_ALG import MemoryTracker

mt = MemoryTracker()
mt.start()
# ... allocate tensors / run forward ...
peak = mt.peak()              # bytes
mt.checkpoint("after-forward")  # named snapshot
mt.stop()
print(mt.report())            # text summary + growth between checkpoints

Notes: On CUDA, MemoryTracker queries cudaMemGetInfo; on CPU it reads /proc/self/status (Linux) or GetProcessMemoryInfo (Windows). CPU-safe fallback reads RSS.