python-optimization-flags

Study: Python Optimization Flags (-O / -OO)

What the flags do

Flag	Effect
-O	Removes assert statements; sets __debug__ = False
-OO	All of -O, plus strips docstrings from compiled bytecode (.pyc)

Hypothesis

For ML-style workloads:

• Code with large docstrings should show little-to-no runtime difference (docstrings are not evaluated at call time), but .pyc files will be smaller.
• Pure compute (e.g., numpy matrix ops, PyTorch CNN forward/backward) should be unaffected, since the bottleneck is in native C++/CUDA, not Python bytecode.

Workloads

File	Description
workloads/docstring_heavy.py	Repeated instantiation of heavily-documented layer classes
workloads/matrix_compute.py	Numpy matmul loop (control — should be flat across all flags)
workloads/ppo_atari.py	CleanRL PPO update kernel on synthetic Atari observations (torch, CPU)

Running

uv sync
uv run python benchmark.py

Results

To be added