Precision AI Pace

Pace follows semantic versioning. Before 1.0, a minor version may change the Python API, but the [pace ...] line format only grows: new keys, never a changed one. (0.1 replaced the 0.0.x [pace:1 ...] format, which it does not read.)

Pace is a logger with progress reporting built in. Its output reads like any good logger in a terminal, and every progress line also carries a short, versioned trailer that a scheduler can read to estimate when the job will finish.

• One way to log across projects: the same levels, format and progress bars everywhere.
• Progress for real ML work: loops, manual counters, timed steps, phases, PyTorch Lightning and Hugging Face training.
• Completion estimates: the terminal shows time remaining, and a scheduler such as OrchAI gets the same estimate from the job's logs.
• No dependencies in the core, and Python 3.11 or later, so it fits into existing training images.

Installation

pip install precisionai-pace

Usage

from precisionai.pace import capture_logging, logger, progress, track

capture_logging()  # route PyTorch's and Transformers' standard logging through Pace

logger.info("Loading model")
logger.success("Model ready")

for batch in track(loader, description="Inference", unit="batch"):
    infer(batch)

with progress(total=len(requests), description="Evaluation", unit="request") as run:
    for future in as_completed(futures):
        consume(future.result())
        run.advance()

Also available: timed, phase and plan for multi-stage jobs, root=True to mark the scope that stands for the whole job, and completed= with resumed=True to resume from a checkpoint. Settings come from logger.configure(...) or PACE_* environment variables, so a scheduler can change them without code changes.

In a terminal, progress is one bar that redraws in place. In a container or a log file it is a rate-limited line like this one:

2026-11-03T17:36:42Z PROGRESS Train  1840/5000  36.8%  2.4 step/s · ~21m57s  [pace train c=1840 t=5000 e=763.1 q=57 r=6bfc3dd5 root]

The part before [pace ...] is for people. The bracketed trailer is a stable, versioned format described in docs/design.md; a program that reads it should use parse_line, LineSplitter and JobForecast (see Reading Pace output).

Documentation

The HTML build of these pages (and the module reference) is produced with Sphinx:

cd docs
make html       # HTML docs → docs/_build/html/index.html
make latexpdf   # PDF      → docs/_build/latex/documentation.pdf

Dependencies: pip install -r docs/requirements.txt

Development

python -m venv .venv
source .venv/bin/activate

pip install -e ".[dev]"
pre-commit install

python -m pytest

Coverage must remain at or above 90%, enforced by pytest and the pre-commit hook. CI runs the tests on Python 3.11 through 3.13.

Project layout

precisionai/pace/   # the package (core: standard library only)
docs/               # Sphinx (HTML + LaTeX/PDF), including design.md
skills/adopt-pace/  # Claude Code skill that adds Pace to a repository
examples/           # runnable examples
tests/              # mirrors the package structure

See CLAUDE.md for the coding standard.

Releasing

Push a tag matching v*.*.* (e.g. v0.1.0) to trigger .github/workflows/release.yml, which runs the test suite, builds the sdist and wheel, publishes to PyPI via trusted publishing, and creates a GitHub Release. The version comes from the git tag via setuptools-scm. Update CHANGELOG.md before tagging.

Security

To report a security vulnerability, see SECURITY.md. Do not open a public issue.

Contributing

See CONTRIBUTING.md for setup and PR guidelines, and CODE_OF_CONDUCT.md for community expectations.

License

Apache 2.0 © Precision AI