dynquant-core 0.1.0

Training-dynamics-driven mixed-precision LLM quantization: signals, scoring, allocation, packing. Pure Python.

dynquant-core

The pure-Python half of DynQuant: training-dynamics-driven mixed-precision LLM quantization.

Most users should pip install dynquant, which adds the prebuilt CUDA kernels. Install dynquant-core directly when you want the quantization pipeline without any binary dependency — CPU-only machines, Windows, macOS, ARM, and CI.

What is in here

• dynquant.signals — the training-time hook and the stats file format
• dynquant.graph — architecture-generic module role classification
• dynquant.quant — group-wise n-bit packing and the quantizer driver
• dynquant.runtime — backend selection and the packed Linear
• dynquant.cli — the dynquant command, including dynquant doctor

Everything except inference speed works with no compiler and no GPU. The reference (torch) backend dequantizes to compute, so it saves no memory and is not fast — but it is the oracle the CUDA kernels are tested against, and it is what makes quantizing a checkpoint possible on a laptop.

Install

pip install dynquant-core            # quantize anywhere
pip install 'dynquant-core[hf]'      # + transformers, accelerate
pip install 'dynquant-core[train]'   # + peft, trl, datasets
pip install 'dynquant-core[dev]'     # + pytest, ruff, mypy, docs

License

Apache-2.0.