Scalene: CPU, GPU, and Memory Profiler

Warnings

• gotcha Scalene explicitly requires Python versions `!=3.11.0` and `>=3.8`. This means Python 3.11.0 is not supported, though 3.11.1+ should be fine.
  Fix: Ensure your Python environment is not 3.11.0. Upgrade to 3.11.1+ or use another compatible version (3.8, 3.9, 3.10, 3.12, etc.).
• breaking Earlier versions of Scalene (before 2.2.1) could crash with `SIGSEGV` when used with PyTorch Lightning on Python 3.12+ due to `sys.monitoring` tool ID conflicts.
  Fix: Upgrade to Scalene v2.2.1 or newer. This version includes a fix that handles conflicts by falling back to alternative tool IDs or the legacy `PyEval_SetTrace` tracer.
• breaking TensorFlow 2.21+ changed `trace.enabled` from a callable to a boolean, causing compatibility issues with Scalene versions prior to 2.2.1.
  Fix: Upgrade to Scalene v2.2.1 or newer to ensure compatibility with TensorFlow 2.21 and later.
• gotcha On Windows, Scalene v2.1.x could experience extreme slowness and memory explosion during CPU profiling due to a hardcoded 1ms sampling interval, overriding the configured rate.
  Fix: Upgrade to Scalene v2.1.2 or newer to resolve this issue. Version 2.1.2 fixed the Windows timer loop to use the actual configured sampling rate.
• gotcha The primary and recommended way to use Scalene for profiling is via the command-line interface (`python -m scalene your_script.py`) or Jupyter magics (`%%scalene`). While a programmatic API exists, directly importing and calling functions to *run* the profiler in a general context is less common and might not capture the full execution scope as effectively.
  Fix: For general profiling, always prefer `python -m scalene <your_script.py> [args]` or use the Jupyter `%%scalene` magic. Use the programmatic API (`from scalene import scalene_profiler; scalene_profiler.start()`) for fine-grained control within specific code blocks.

Install

• pip install scalene Install Scalene

Imports

• scalene_profiler

  from scalene import scalene_profiler

  For programmatic control (start/stop) of the profiler. Most users will use the CLI directly.

Quickstart

Create a Python script (e.g., `my_script.py`) and then run Scalene from the command line using `python -m scalene my_script.py`. Scalene will generate a detailed profile in your terminal or an HTML report.

# my_script.py
import time
import math

def calculate_heavy_stuff(n):
    result = 0
    for _ in range(n):
        result += math.sin(math.sqrt(time.time()))
    return result

def main():
    print("Starting heavy calculations...")
    calculate_heavy_stuff(1000000) # Perform 1 million iterations
    print("Calculations finished.")

if __name__ == "__main__":
    main()

# To run the profiler from your terminal:
# python -m scalene my_script.py