lmsys/qwen-image-2512-modelopt-nvfp4-sglang

Qwen Image 2512 ModelOpt NVFP4 for SGLang

This repository contains a full Hugging Face-style SGLang Diffusion checkpoint exported from Qwen/Qwen-Image-2512 with NVIDIA ModelOpt NVFP4 quantization support. The checkpoint can be loaded directly with --model-path in SGLang; no separate --transformer-path override is required.

Export provenance

The checkpoint was generated on an NVIDIA B200 machine by applying NVIDIA ModelOpt PR #1706 locally (qwen-image-svdquant-nvfp4-pr) and exporting a full pipeline directory for SGLang validation.

Export settings:

• —Base model: Qwen/Qwen-Image-2512
• —Pipeline class: QwenImagePipeline
• —Quantization recipe: SVDQuant NVFP4 / FP4 export through ModelOpt
• —Main export flag: --format fp4
• —Calibration: max calibration with 8 representative prompts
• —Block size: 16
• —Quantized scope: diffusion transformer blocks
• —BF16 fallbacks: first two and last two transformer blocks, plus attention quantizers disabled during export
• —Validation target: NVIDIA B200 / Blackwell using SGLang Diffusion

Validated as a full SGLang Diffusion text-to-image checkpoint on B200 and covered by the SGLang B200 diffusion NVFP4 test.

SGLang usage

sglang generate \
  --model-path lmsys/qwen-image-2512-modelopt-nvfp4-sglang \
  --prompt "A tiny astronaut reading a book under a glass greenhouse" \
  --save-output

For edit pipelines, pass the normal SGLang Diffusion image-edit inputs in addition to --model-path lmsys/qwen-image-2512-modelopt-nvfp4-sglang.

Notes

This is a performance and integration validation checkpoint for SGLang Diffusion NVFP4 support. Please compare generated images with the corresponding BF16 base model for quality-sensitive use cases.