DFloat11 Compressed Model: Qwen/Qwen-Image-Edit-2511

Qwen-Image-Edit-2511

📊 Performance Comparison

🔧 How to Use

1. Install or upgrade the DFloat11 pip package (installs the CUDA kernel automatically; requires a CUDA-compatible GPU and PyTorch installed):
   pip install -U dfloat11[cuda12]
2. Install or upgrade diffusers:
   pip install git+https://github.com/huggingface/diffusers
3. Save the following code to a Python file qwen_image_edit.py:
   python

   1import os 2import torch 3import argparse 4from diffusers import QwenImageEditPlusPipeline 5from diffusers.utils import load_image 6from dfloat11 import DFloat11Model 7 8parser = argparse.ArgumentParser(description="Qwen Image Edit with DFloat11") 9parser.add_argument("--image", default="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cat.png", help="Image URL or path") 10parser.add_argument("--prompt", default="Make this cat an astronaut gazing at planet earth from space", help="Edit prompt") 11parser.add_argument("--output", default="qwen_image_edit_output.png", help="Output image path") 12parser.add_argument("--steps", type=int, default=40, help="Number of inference steps") 13parser.add_argument("--seed", type=int, default=42, help="Random seed") 14parser.add_argument("--true_cfg_scale", type=float, default=4.0, help="True CFG scale") 15parser.add_argument("--negative_prompt", default=" ", help="Negative prompt") 16parser.add_argument("--guidance_scale", type=float, default=1.0, help="Guidance scale") 17parser.add_argument("--cpu_offload", action="store_true", help="Enable CPU offloading") 18parser.add_argument("--cpu_offload_blocks", type=int, default=20, help="Number of blocks to offload to CPU for block swapping") 19parser.add_argument("--cpu_offload_no_pin_memory", action="store_true", help="Disable memory pinning for CPU offloading") 20args = parser.parse_args() 21 22pipeline = QwenImageEditPlusPipeline.from_pretrained("Qwen/Qwen-Image-Edit-2511", torch_dtype=torch.bfloat16) 23DFloat11Model.from_pretrained( 24 "DFloat11/Qwen-Image-Edit-2511-DF11", 25 bfloat16_model=pipeline.transformer, 26 device="cpu", 27 cpu_offload=args.cpu_offload, 28 cpu_offload_blocks=args.cpu_offload_blocks, 29 pin_memory=not args.cpu_offload_no_pin_memory, 30) 31pipeline.enable_model_cpu_offload() 32 33image = load_image(args.image) 34inputs = { 35 "image": [image], 36 "prompt": args.prompt, 37 "generator": torch.manual_seed(args.seed), 38 "true_cfg_scale": args.true_cfg_scale, 39 "negative_prompt": args.negative_prompt, 40 "num_inference_steps": args.steps, 41 "guidance_scale": args.guidance_scale, 42 "num_images_per_prompt": 1, 43} 44with torch.inference_mode(): 45 output = pipeline(**inputs) 46 output_image = output.images[0] 47 output_image.save(args.output) 48 print("Image saved at", os.path.abspath(args.output)) 49 50max_memory = torch.cuda.max_memory_allocated() 51print(f"Max memory: {max_memory / (1000 ** 3):.2f} GB")
4. To run without CPU offloading (32GB VRAM required):
   python qwen_image_edit.py
   To run with CPU offloading (24GB VRAM required):
   python qwen_image_edit.py --cpu_offload
   If you are getting out-of-CPU-memory errors, try limiting the number of offloaded blocks or disabling memory-pinning:
   bash

   1# Offload only 16 blocks (offloading more blocks uses less GPU memory and more CPU memory; offloading less blocks is faster): 2python qwen_image_edit.py --cpu_offload --cpu_offload_blocks 16 3 4# Disable memory-pinning (the most memory efficient way, but could be slower): 5python qwen_image_edit.py --cpu_offload --no_pin_memory

🔍 How It Works

📄 Learn More

• Paper: 70% Size, 100% Accuracy: Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float
• GitHub: https://github.com/LeanModels/DFloat11
• HuggingFace: https://huggingface.co/DFloat11