dynquant 0.1.0

Training-dynamics-driven mixed-precision LLM quantization, with prebuilt CUDA kernels.

dynquant

Mixed-precision LLM quantization that decides bit-widths from your fine-tune's own training dynamics.

pip install dynquant
dynquant doctor

This distribution contains no code. It is the one name to install, and it pulls in:

• dynquant-core — the Python half: signal collection hook, role classification, scoring, allocation, packing, CLI. Installs anywhere, no compiler required.
• dynquant-kernels — prebuilt CUDA kernels, where a wheel exists for your platform. Without them everything still works on the reference backend; you lose inference speed and the VRAM saving, not correctness.

Usage

Collect signals during the fine-tune you were going to run anyway:

from transformers import Trainer
from dynquant import DynQuantCallback

trainer = Trainer(model=model, ..., callbacks=[DynQuantCallback("stats/")])
trainer.train()

Allocate and pack:

dynquant quantize ./merged --stats stats/dynquant_stats.json --target 3.0 -o ./q3

Load through plain transformers:

from transformers import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("./q3")

Full documentation: https://github.com/kambojvikram/dynquant

Extras

pip install 'dynquant[train]'     # transformers, peft, trl, datasets
pip install 'dynquant[eval]'      # lm-eval-harness
pip install 'dynquant[triton]'    # portability fallback for ROCm / newer GPUs
pip install 'dynquant[kernels]'   # force the compiled kernels (builds from sdist)

License

Apache-2.0.