WiktorMatuszek/MOSS-Image-SFT-NVFP4-Core

MOSS-Image-SFT — NVFP4-Core

Release status

Experimental checkpoint — further testing required. Packed NVFP4 weights and saved-shard integrity were verified; native NVFP4 full 50-step inference, performance and image fidelity have not been validated. Short fallback / weight-reconstruction tests do not demonstrate native FP4 speed.

Benchmark interpretation: The image gallery and measurements are reference examples for visual inspection, not a required reproducibility package or guarantee of quality, speed, or VRAM usage on another setup.

Original MOSS-Image-SFT DiT with genuinely packed NVFP4 storage weights. The model is quantized from the original SFT checkpoint; there is no extra training, LoRA, distillation or architecture change.

Property	This checkpoint
Quantization	Packed E2M1 4-bit weights, FP8 block scales and FP32 global scale
Coverage	228 / 269 transformer Linear projections (precision-sensitive edge layers retained in BF16)
Stored DiT weights	3.706 GiB, 60.9% smaller than original 9.467 GiB BF16 DiT
Checkpoint files	Three TorchAO/PyTorch .bin weight shards; not generic safetensors
Upstream revision	OpenMOSS-Team/MOSS-Image-SFT, fc5de0d56fc910cbca6f6625a70f5f582b26deb1
Integrity	All 463 original state keys matched; shards reloaded and SHA-256 verified

This repo contains the transformer only. The original Qwen text encoder, tokenizer, FP32 VAE, scheduler and pipeline code are loaded from the pinned upstream revision as needed. The stored DiT size is not total RAM/VRAM or whole-pipeline model size.

Evidence and measured limits

Weights: all 228 quantized Linear weight matrices compared with the original BF16 tensors: parameter-weighted MSE 1.98e-05, SQNR 20.47 dB, cosine 0.995515. These are weight-space errors, not generated-image correctness. Detailed per-layer measurements, quantization receipt, checkpoint integrity.

Original SFT50 image study: Not measured with the original 50-step pipeline and native runtime in the recorded SM80 benchmark. Checkpoint weight verification does not imply that image quality or GPU execution passed.

Runtime qualification: native NVFP4 execution did not pass the original SFT50 kernel check in the recorded SM80 setup. A distinct short fallback reconstructed weights for higher-precision computation; it is not native FP4 speed evidence, nor a quality-accepted 50-step run. See kernel and image study.

Use with original SFT inference

Install a CUDA-compatible PyTorch for your GPU, followed by the repository requirements. The conversion environment was PyTorch 2.7.0+cu128 / TorchAO 0.12.0+cu128; compatibility with packed tensor subclasses and native kernels is device dependent.

hf download WiktorMatuszek/MOSS-Image-SFT-NVFP4-Core --local-dir ./moss-quant
cd moss-quant
python -m pip install -r requirements.txt

# Check all real saved shards before using the model
python generate.py --verify-only

# Original MOSS-SFT: 50 denoising steps, CFG 3, 1024px
python generate.py --prompt-file example_layout.json --output image.png

The launcher downloads the unchanged missing upstream components, not duplicate original BF16 DiT weights. Optional Qwen FP8 encoder / Flux2 VAE FP16 can be selected with --encoder-repo and --vae-repo; that combination needs its own end-to-end quality validation.

Files and limitations

transformer/ contains the quantized checkpoint; generate.py, run_quantized_sft.py, compact_components.py contain the strict pinned loader; QUANTIZATION_RECEIPT.json and COMPLETE.json record checksums. Frozen prompt suites are archived once in the research dataset, not copied into this model repo.

Pixel similarity does not prove correct Polish spelling, photorealism, aesthetic quality or native-kernel support. Test speed and GPU memory on your intended stack. The full cross-format gallery, including explicitly missing measurements, is maintained once in the benchmark dataset.

Attribution: independently quantized from OpenMOSS-Team/MOSS-Image-SFT at fc5de0d56fc910cbca6f6625a70f5f582b26deb1; original Apache-2.0 license in LICENSE. Not an official OpenMOSS-Team release.