Nunchaku Lite quantization in place - #14608

Draft

rootonchair wants to merge 13 commits into

huggingface:mainfrom

rootonchair:feature/data-free-quantization

Commits on Aug 25, 2026

Add data-free SVDQuant quantize-on-load to Nunchaku Lite

rootonchair
and
claude
committed
fc27874
Infer data-free quantization targets automatically

rootonchair
and
claude
committed
ebfc2c9
Infer data-free exclusions structurally

rootonchair
and
claude
committed
26bf67e
Rename modules_to_not_convert to exclude_targets

rootonchair
and
claude
committed
ae79c1b
Silence false missing/unexpected key warnings in data-free mode

rootonchair
and
claude
committed
67da88e

Commits on Aug 26, 2026

Rename data_free.py to svdquant.py

rootonchair
and
claude
committed
6cb4b2e
Rename mixin test to test_nunchaku_lite_quantize_on_load

rootonchair
and
claude
committed
cae6967

Commits on Aug 31, 2026

Materialize quantize-on-load params for checkpoint keys absent from the model

rootonchair
and
claude
committed
c4886e3
Expose data-free smoothing strength as svdq_w4a4.smooth_exponent

rootonchair
and
claude
committed
aaa0d21
Document rank and smooth_exponent tuning for data-free quantization

rootonchair
and
claude
committed
aa4428d

Commits on Sep 1, 2026

Allow CPU quantize-on-load and save_pretrained for Nunchaku Lite

rootonchair
and
claude
committed
93f6cc2
Stop coercing exclude_targets=None into () before target inference

rootonchair
and
claude
committed
8ffd5ef
Show pipeline-level save/load in the CPU-quantize doc example

rootonchair
and
claude
committed
f0ed907