StandardPipeline.show_details uses Python builtin max()/min() on the fitness tensor — O(P) interpreter time per generation (~3 s/gen at pop 1M)

[xavierhillroy]

Summary

StandardPipeline.show_details (default is_show_details=True) computes fitness statistics with
Python's builtin max()/min() over the fitness tensor
(src/evogp/pipeline/standard.py lines 96–97, commit 5538d75):

max_f, min_f, mean_f, std_f = (
    max(valid_fitness),      # Python builtin -> element-by-element tensor iteration
    min(valid_fitness),      # Python builtin
    torch.mean(valid_fitness),
    torch.std(valid_fitness),
)

The builtins iterate the tensor element-by-element in the Python interpreter (one __getitem__
plus comparison per element), so this logging step costs O(pop_size) interpreter time every
generation, and at large populations it dominates the entire evolutionary loop.

Measurements (RTX 5070 Ti, torch 2.9.0+cu130, evogp @ 5538d75)

Microbenchmark on a 1M-element fitness tensor:

import torch, time
t = -torch.rand(1_000_000) * 100
t0 = time.perf_counter(); m = max(t);       print(time.perf_counter() - t0)  # 1.48 s
t0 = time.perf_counter(); m = torch.max(t); print(time.perf_counter() - t0)  # 0.0003 s

Pipeline-level impact (SR, pop 1M, 100 generations, D=64, TournamentSelection + DefaultCrossover

• DefaultMutation):

	time / generation
is_show_details=True (default)	~2.5 s
is_show_details=False	~0.09 s

i.e. with the default setting, ~96% of wall-clock at this population size is the two builtin
calls in the logging path — not the (genuinely fast) evolutionary pipeline. At pop 10k the
logging is still ~90% of run time (~2.6 s vs ~0.28 s per 100 generations).

This matters beyond aesthetics: anyone timing StandardPipeline.run() at large populations with
the default settings (e.g. for benchmarking or framework comparisons) will measure the logging,
not EvoGP — we hit exactly this while running a comparison study and initially mis-measured
EvoGP's whole-run performance by ~25× before tracing it here.

Suggested fix (one line each)

max_f, min_f, mean_f, std_f = (
    torch.max(valid_fitness),
    torch.min(valid_fitness),
    torch.mean(valid_fitness),
    torch.std(valid_fitness),
)

(len(valid_fitness) in the print is fine; optionally the whole stats block could stay on GPU
and transfer scalars only.)

Environment

• evogp installed from source at commit 5538d75 (pip install --no-build-isolation .)
• PyTorch 2.9.0+cu130, CUDA 13.1, NVIDIA RTX 5070 Ti (sm_120), Linux

I will open a PR :) — and thanks for the framework; the pipeline itself
benchmarks impressively once the logging is bypassed.