auto-adpq 0.3.5

Implementing auto adpq

auto_adpq

Adaptive Post-Training Quantization tooling (replicating AdpQ)

This repository implements tools and reference code to reproduce the ideas from AdpQ: A Zero-shot Calibration Free Adaptive Post Training Quantization Method for LLMs.

This README explains how to install, run tests, build documentation (including multi-version docs), and contribute.

• auto_adpq
  • Installation
  • Quick usage
  • Running tests & linters
    • Debug mode
  • Documentation
    • Building the documentation
  • Tasklist
  • Quantized models
    • Performances
  • Contributing
  • Development notes
  • License

Installation

Install from PyPI (recommended):

python -m pip install auto_adpq

Install the latest development version directly from GitHub:

python -m pip install "git+https://github.com/Tfloow/auto_adpq.git"

To develop locally (editable install):

git clone https://github.com/Tfloow/auto_adpq.git
cd auto_adpq
python -m pip install -e .

Makefile helper:

# Run formatting, linting, coverage and docs targets as defined in Makefile
make

Quick usage

Import the package and use the public API. Example (replace with real API):

from auto_adpq import Auto_AdpQ

Add a short usage snippet here specific to the package functions you expect users to try first.

The most simple way to quantize a model is to follow a similar script as in examples/simple_quantization.py.

Running tests & linters

Coverage test: 91%

• Run tests with pytest:

pytest -q

• Run full coverage report (Makefile target):

make coverage

• Format & lint with ruff (Makefile target):

make ruff

Debug mode

To obtain logs of the package, it is possible to enable the logging module. To activate it please create the new environment variable AUTO_ADPQ_DEBUG by running:

# Linux
export AUTO_ADPQ_DEBUG=1

# Windows
$Env:AUTO_ADPQ_DEBUG = 1

Documentation

The documentation can be found here.

Building the documentation

This project uses Sphinx for documentation. There are two common workflows:

• Build a single-version site (useful for local writing and previews):

python -m pip install -r docs/requirements.txt
python -m sphinx -b html docs docs/_build/html

• Build a multi-version site using sphinx-multiversion (we configure this in docs/conf.py). This produces one static site containing each built branch and tag (useful for publishing versioned docs with a dropdown selector):

python -m pip install -r docs/requirements.txt
sphinx-multiversion docs docs/_build/html-mv

Notes about versions

• The project includes a small template docs/_templates/versions.html which renders a versions dropdown when the site is built with sphinx-multiversion.
• Adjust smv_tag_whitelist and smv_branch_whitelist in docs/conf.py to control which tags/branches are included in the build.

Tasklist

• Solve the datapacking issue #1
• Support efficient inference (maybe wrap in SpQR?)
• Optimize pydantic module AdpQQuantizedWeights
  • Currently, there is a major overhead when creating a new object to validate the field. Since it is used internally only, we could ditch the Pydantic module but would need to ensure proper dump and load function
• Support model and integrate with .safetensors

Quantized models

Pre-quantized models are available in this collection. They are simulated models meaning they are stored as bf16 values instead of the quantized versions. If I stored them in the custom format, I would either need an algorithm to reconstruct the weights in full at runtime or develop a custom CUDA kernel, which is quite tough.

Nonetheless, those models represent the quality and rounding errors that a typical quantized model can meet.

Performances

---

Model Variant	Quantization Method	PPL (Perplexity)
meta-llama/Llama-3.1-8B	Baseline	4.8693
	BNB	5.0733
	AdpQ	5.3671
meta-llama/Llama-3.1-8B-Instruct	Baseline	4.9080
	BNB	4.9993
	AdpQ	5.0069
	AWQ	5.0440
	GPTQ	nan
meta-llama/Llama-3.2-1B	Baseline	6.5546
	AdpQ 9%	6.9491
	BNB	6.9971
	AdpQ 2%	7.0380
meta-llama/Llama-3.2-3B-Instruct	Baseline	5.7864
	AWQ	5.8339
	AdpQ	5.9040

Contributing

Contributions are welcome. A suggested workflow:

1. Fork the repository and create a feature branch.
2. Add tests for new functionality.
3. Run ruff to format and lint.
4. Open a pull request describing the change.

Please include unit tests and keep the public API stable when possible.

Development notes

• Docs templates: docs/_templates/versions.html — version switcher used by sphinx-multiversion.
• Makefile targets: make ruff, make coverage, make docs (runs single and multiversion builds).

License

This work is under Apache 2.0 License.