AbsorbQuant — W4A4 NVFP4 diffusion-transformer checkpoints

Pre-quantized 4-bit (weights and activations) checkpoints produced by AbsorbQuant: decompose-first H-metric low-rank quantization with a rank-32 fp16 branch on the raw input and a GPTQ-quantized NVFP4 residual, running on the real nunchaku gemm_w4a4 kernel. Code, method, and theory: https://github.com/chenjiaj109550158/AbsorbQuant.

Release lam0.1-v4 — search-free recipe (2026-09-23)

Six additional checkpoints built with protocol v4 and the search-free recipe of docs/ALPHA_SELECTION_SSE.md: H-SVD damping λ fixed at 0.1 and the smoothing strength α chosen per layer type from the qdiff-128 layer-SSE tables (config smooth: {family: amax, rule: sse-per-type, tau: 0.0}), i.e. no Algorithm-1 image generation at all. Calibration inputs: the clean-room v4 qdiff-128 caches and covariances (machine #3, 2026-09-21). Each folder also carries PROVENANCE.md, the per-hook alpha_map_per_type.json, the selection record select_sse.json and the builder sidecars.

Final 5000-image evaluation (MJHQ-5000 and sDCI-5000, one library each; baselines = the archived SVDQuant+GPTQ / official SVDQuant libraries on the same 16-bit references; S2 minus strongest baseline, PSNR, and wins on PSNR/LPIPS/SSIM/FID-ref/FID-GT):

Full tables, the development-set study (MJHQ-500 / sDCI-500, which also chose this recipe) and caveats are in docs/ALPHA_SELECTION_SSE.md. Download and verify:

python scripts/download.py --model pixart-sigma-lam0.1-v4
python scripts/verify.py --digest workdir/pixart-sigma-lam0.1-v4/checkpoints/absorbquant-pixart-sigma-nvfp4-lam0.1-v4.pt
# rebuild from scratch (metric-level reproduction; bit-exact from the same calibration inputs):
bash scripts/from_scratch/sse.sh pixart-sigma

Release v2 (2026-09-09). The selection algorithm's smoothing gate was corrected (per-hook gains measured at the built strength α with both activation and weight NVFP4-quantized; amax closed-form family; gate threshold 0) and every recipe was re-derived from scratch. The four v2 files above replace their v1 versions; SANA-1.6B's v2 file followed on 2026-09-10 (re-derived from scratch on an independent RTX 5090, bit-identical when rebuilt on the release machine), and FLUX.1-schnell's on 2026-09-10 (re-derived from scratch on that independent machine, Algorithm-1 included: same recipe and same content digest as the release machine's product). v1 recipes are preserved at the repository's git tag v1; v1 files are not reproducible from the current main code.

Usage

git clone https://github.com/chenjiaj109550158/AbsorbQuant && cd AbsorbQuant
pip install -e ".[models]"   # plus the nunchaku wheel for your torch/CUDA
python scripts/download.py --model pixart-sigma
python scripts/generate.py --model pixart-sigma --prompt "a corgi in space"

Provenance / verification

Every checkpoint here is the byte-for-byte output of the repository's from-scratch pipeline (scripts/calibrate.py + scripts/build.py) under the pinned environment in the repository README: calibration uses only the 128 fixed COCO-derived prompts shipped in the repo — no SVDQuant calibration artifacts. scripts/verify.py re-checks tensor-level bit-identity between a rebuild and these files and prints a canonical content digest; the digests for these files are recorded in the repository's configs/*.yaml.

Licenses

The quantization code is Apache-2.0. Quantized weights are derivatives of their base models and inherit each base model's license; this repository does not grant a blanket license for the weights. In particular FLUX.1-dev weights fall under the FLUX.1 [dev] Non-Commercial License and SDXL-Turbo under the Stability AI Non-Commercial Research Community License, so neither may be used commercially. Check the base model licenses before commercial use.