torchmatch
module

bench

module torchmatch.bench

Benchmark collection CLI.

Helpers to capture a pytest-benchmark sweep on the local machine, scrub host-identifying fields, and write the result under ``benchmarks/results//`` with a canonical filename that encodes the torchmatch / Python / PyTorch versions and the CUDA wheel variant. The full contributor flow is :func:`init_machine` (once) + :func:`collect_run` (once per release). The aggregator under ``scripts/benchmark_aggregate.py`` then turns the per-machine files into the JSON datasets that the docs site consumes.

Members

function

  • build_slugCompose the flat slug used as the per-machine directory name.
  • collect_runRun the benchmark sweep and write a scrubbed JSON to the slug dir.
  • current_run_idFilename (sans .json) for a run captured right now on this interpreter.
  • detect_machineFull machine.json payload for the local box.
  • detect_variantMap the running PyTorch build to a wheel-variant tag.
  • init_machineWrite benchmarks/results/<slug>/machine.json and return its path.
  • make_filename
  • parse_filenameReverse a canonical filename into its component dict, or None.
  • scrub_machine_infoDrop hostname/kernel fields in-place and redact stray hostname strings.
  • short_nameNormalize a noisy CPU/GPU brand string to a filesystem-safe short slug.

Source: torchmatch/bench/__init__.py:1