gcbt-quant

GCBT checkpoint and experiment handoff — 2026-09-20

Public migration snapshot for Duke-CEI-SVD. Upload scope is restricted to the requested model weights, generated evaluation media/records, and minimal loading/restoring/scoring helpers; full repository and execution-source snapshots are excluded. This bundle contains one unchanged, genuinely integer-packed checkpoint: the authors' Qwen-Image SVDQuant INT4 / rank-32 Native model (11,521,979,944 bytes; 600 integer qweight tensors). Its provenance, full SHA256, tensor dtypes and finite-value checks are in models/qwen-image-int4-r32/provenance.json.

Runtime limitation: weights and model loading passed local checks, but end-to-end inference failed on the available H200 because official Nunchaku 1.2.1 does not include Hopper sm_90 kernels. The provided script targets Ampere/Ada SM80/86/89. It has not been validated end-to-end on those GPUs in this handoff. H100/H200 need a separately validated kernel build; B200/RTX 5090 need a separately validated format/runtime. Details: validation/qwen_int4_runtime.json and the official build configuration.

This is the upstream Native checkpoint, not a fused GCBT INT4 checkpoint. Archived experimental images made through dequantized BF16 reference hosts are not claimed to be bit-identical to Nunchaku output. No new GCBT kernel support or packed deployment quality result is implied.

SD3.5-Large PTQ completed and its five original files remain on the source machine. All five passed size, SHA256 and safe deserialization checks; all floating tensors were finite. These files store dequantized BF16 PTQ states and are not included as deployable INT4 weights. See validation/sd35_ptq_audit.json and models/eligibility.json. PixArt, Wan, LTX, SANA and SDXL reference-runtime checkpoints are likewise not relabeled as packed deployments.

Download

After the upload is confirmed complete, download this public bundle:

python -m pip install 'huggingface_hub==0.36.2'
hf download Duke-CEI-SVD/gcbt-quant --local-dir gcbt-handoff
cd gcbt-handoff

release_manifest.json records the exact files and SHA256 values in this snapshot. artifact_catalog.json and each run's index.json enumerate saved artifacts. Snapshot dates and missing samples must be respected; a complete archive does not mean the experiment finished.

Load the real INT4 Qwen model on a supported GPU

Use a fresh Python 3.12 environment, Linux x86-64, an appropriate NVIDIA driver, and an Ampere/Ada GPU. The version combination below imported and loaded successfully here; generation was blocked by H200 architecture support as described above.

python -m pip install torch==2.8.0 torchvision==0.23.0 --index-url https://download.pytorch.org/whl/cu128
python -m pip install -r runtime/requirements-qwen.txt
python -m pip install --no-deps 'https://github.com/nunchux-ai/nunchaku/releases/download/v1.2.1/nunchaku-1.2.1%2Bcu12.8torch2.8-cp312-cp312-linux_x86_64.whl'
python runtime/run_qwen_int4.py --offload --output qwen-native.png

The script loads the local packed transformer and downloads the text encoder, tokenizer, VAE and other pipeline components from Qwen/Qwen-Image, pinned to revision 75e0b4be04f60ec59a75f475837eced720f823b6. These auxiliary base-model weights are not duplicated in this repository. For an offline machine, download the base components first and pass --base-model /path/to/Qwen-Image. An entirely offline fresh-machine deployment has not been validated.

The PNG sidecar records the prompt, seed, steps, versions, GPU and image checksum. Defaults use 50 steps and true CFG 4.0. An optional --steps 2 --height 512 --width 512 is only a loading/runtime smoke test, not an image-quality evaluation. Do not use these new images as replacement pairs for archived results without matching the original generation protocol.

Restore saved images, videos and experiment records

Archives preserve original completed PNGs, lossless MKV videos, MP4 previews, per-sample metadata, prompts, split manifests, available calibration summaries and scores. They exclude PTQ .pt states, embedding caches, base-model weights and datasets. Active runs are snapshots, so later outputs on the source machine are not automatically added. Historical partial/failed runs are retained as such.

python runtime/restore_artifacts.py --output restored
# Or restore only the SD3.5 run:
python runtime/restore_artifacts.py --output restored --run sd35_svdquant_paper256
# Check archive SHA256 without extracting:
python runtime/restore_artifacts.py --verify-only

The restore command verifies archive SHA256, uses safe tar extraction, then checks every extracted file. Avoid mixing run folders. Full source trees remain in the local backup and are not part of this upload. Existing source manifests retain their hashes for provenance, but the corresponding source files are not bundled. The original source machine's absolute paths are provenance and may require editing before resuming generation elsewhere. This bundle is not a claim that every historical experiment is a turnkey resumable deployment.

Recompute paired fidelity

Install the metric dependencies and system ffmpeg/ffprobe. LPIPS also needs its torchvision backbone weights, downloaded on first use unless already cached.

python -m pip install -r runtime/requirements-metrics.txt
python runtime/score_pairs.py \
  --reference restored/runs/sd35_svdquant_paper256/sd35_svdquant/pilot/development/fp \
  --candidate restored/runs/sd35_svdquant_paper256/sd35_svdquant/pilot/development/native \
  --expected-count 32 --allow-partial --output sd35-pilot-partial.json

The restore/metric path was tested on all 16 available SD3.5 pilot pairs: LPIPS, SSIM and PSNR reproduced every original score exactly on this machine (validation/sd35_metric_reproduction.json).

This SD3.5 snapshot has 32 FP pilot images and 16 Native pilot images; the command explicitly produces a partial diagnostic, not a completed 32-prompt result. Its formal FP set contains 256 MJHQ + 256 sDCI images; no completed formal quantized comparison is claimed.

For current paper256 videos, pass --video and the matching main/vbench/fp and main/vbench/native or gcbt directories under runs/wan14b_orbit_paper256/wan14b_orbit. Use lossless .mkv, not .mp4, to recompute fidelity. Require --expected-count 256; use --allow-partial only when intentionally reporting an incomplete diagnostic.

The helper checks sample IDs, prompts, seeds, protocol hashes and available file hashes before scoring. It applies the current paper256 definition: LPIPS v0.1 AlexNet for images / VGG for videos, Gaussian SSIM, PSNR, with frame averages for each video. It reports per-sample scores and means. Formal paired confidence intervals and category-stratified reports require the original project scoring code, frozen manifest and both complete arms; only the minimal per-sample scorer is included here. Historical smoke protocols may use different definitions: consult their recorded settings and scores before recomputing them. FID, CLIP or VBench suite evaluation also requires its original datasets/dependencies and is not performed by this helper.

Attribution

The packed Qwen checkpoint is unchanged from nunchaku-ai/nunchaku-qwen-image, based on Qwen/Qwen-Image. Both upstream cards declare Apache-2.0; recorded cards and available license text are in attribution/. Nunchaku is installed from its official release, not redistributed here. Third-party datasets, reference model families retain their original license terms. The repository's Apache-2.0 tag does not relicense those materials.