morel
Docs/morel — Benchmarks

morel — Benchmarks

Benchmarks use pytest-benchmark. Run with:

pytest benchmarks/ --benchmark-only --no-cov --benchmark-min-rounds=20 --benchmark-autosave

Suites

  • benchmarks/forward.py — forward-pass latency for the encoder + codebook
    • decoder at scales [1k, 10k, 100k] items.
  • benchmarks/dataset.py — full k-core + graph construction end-to-end.
  • benchmarks/retrieve.py — anchor + ACS + MAGE throughput.
  • benchmarks/train.py — completion training epoch on a 50k-item synthetic dataset.

Threshold

The benchmark suite is an opt-in local gate. Record the machine, dependency lock, dataset scale, and command alongside results; do not compare timings across unlike hardware. CI validates the benchmark code and package gates, but does not claim a universal latency budget.