Darkstar-Qwen-Image-2.1-Base-ModelOpt-W4A4-NVFP4

Darkstar-Qwen-Image-2.1-Base-ModelOpt-W4A4-NVFP4

Non-commercial research artifact. NVIDIA ModelOpt NVFP4 (W4A4) quantization of Qwen/Qwen-Image-2.1, built by the Model Forge pipeline from the pinned upstream BF16 — no third-party quantized checkpoints were used as build inputs.

Published revision: 16f1b7fd748aa07e4f19c532991b0defe047dce0

License — Qwen Research License Agreement (read this first)

This is a derivative of Qwen-Image-2.1 and inherits the Qwen Research License Agreement (September 20, 2026). The full agreement text is shipped in this repository as LICENSE.

Qwen is licensed under the Qwen RESEARCH LICENSE AGREEMENT, Copyright (c) 2026 Hangzhou Tongyi Laboratory Technology Co., Ltd. All Rights Reserved.

Commercial use requires a separate license from the Qwen team ([email protected]).

Summary

NVFP4 (W4A4, group-16) quantization of the 7B single-stream DiT (QwenImage21Transformer2DModel, 32 blocks) with the Qwen3-VL 8B text encoder and VAE kept in BF16. The model does text-to-image generation and image-conditioned editing in one pipeline, with native RGBA output support.

Precision map

Provenance

• Upstream model: Qwen/Qwen-Image-2.1
• Upstream revision: b3179ad355be050328e483a9dfdd9e60cd62adfa
• Weight edit: ModelOpt NVFP4 W4A4 quantization (own pipeline; calibration = Gustavosta Stable-Diffusion-Prompts, max algo)
• Toolchain: nvidia-modelopt 0.46.0rc2 @ 43fd41a… + diffusers main @ 80c7ed26… (bridge-fork overlay registering the 2.1 ModelType; sha-pinned)
• Recipe: darkstar-qwen-image-2-1-base-nvfp4.yaml (in the engineering repository)
• Engineering repository: HangGlidersRule/model-forge

Serving (vLLM-Omni — validated)

docker run --gpus all -p 8000:8000 \
  -e NCCL_CUMEM_ENABLE=0 \
  -v <hf-cache>:/root/.cache/huggingface \
  vllm/vllm-omni:qwen-image21 \
  bash -c 'vllm serve HangGlidersRule/Darkstar-Qwen-Image-2.1-Base-ModelOpt-W4A4-NVFP4 --omni --port 8000'

OpenAI-compatible endpoints: POST /v1/images/generations, POST /v1/images/edits. Seed determinism is byte-identical for repeated same-seed requests (max_num_seqs=1 profile).

Note: the checkpoint config carries component-prefixed ignore patterns (*transformer_blocks.N.*). The vLLM-Omni runtime matches quantization exclusions against component-prefixed module names; unprefixed patterns silently fail to exclude (known upstream issue, documented by Model Forge).

Consumption without a server (diffusers / ComfyUI-style)

diffusers from_pretrained on this repo will attempt to parse the ModelOpt quantization config and fails on the current diffusers main schema (known upstream gap — diffusers' ModelOpt quantizer expects a different config shape and does not support NVFP4). Working consumption paths:

1. vLLM-Omni (above) — the validated serving path.
2. Dequant-on-load (documented reference implementation in the Model Forge repo): strip quantization_config from a local copy's transformer/config.json, load the pipeline, then dequantize the packed NVFP4 tensors (e2m1 nibble LUT × per-16-block fp8 scale × global weight_scale_2) into the BF16 model. Validated end-to-end below.
3. ComfyUI native NVFP4 loader (PR #11635+) — expected to consume this format natively on Blackwell.

Evaluation (frozen protocol)

Quality manifest image-eval-v1 (sha 09daf654e4015a49d1f54961687d4f6df41da9910e4c0c25d1981582190aeee9), 25 frozen cases, full-denominator scoring (blanks/errors count as failures):

• Typography spot-check: legible text preserved (no quantization degradation visible).
• Dequant verification vs upstream BF16 weights: max err 0.083, correlation 0.996.
• Quant validators: 196 quantized modules, 0 misclassified vs policy, 0 degenerate scales.
• Performance: 4.78 s/img dequant-on-load (parity with BF16 control's 4.73; native fp4 kernels are the serving speedup path). RTX PRO 6000 Blackwell, 20 steps, 1024².

Family context

The Darkstar Qwen-Image-2.1 family ships a four-cell matrix plus FP8 variants (Base BF16 control profile, this NVFP4 cell, Base FP8, and Ablit cells pending the DiT-side behavior-transform research — tracked in the Model Forge private ledger). Ablit cells are not part of this release.

Non-commercial. Research and evaluation use only. Built with Qwen.