File size: 6,725 Bytes
729204d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4f8f857
729204d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
---
base_model:
- Qwen/Qwen-Image-Edit
base_model_relation: quantized
tags:
- dfloat11
- df11
- lossless compression
- 70% size, 100% accuracy
pipeline_tag: image-to-image
---

# DFloat11 Compressed Model: `Qwen/Qwen-Image-Edit`

This is a **DFloat11 losslessly compressed** version of the original `Qwen/Qwen-Image-Edit` model. It reduces model size by **32%** compared to the original BFloat16 model, while maintaining **bit-identical outputs** and supporting **efficient GPU inference**.

🔥🔥🔥 Thanks to DFloat11 compression, Qwen-Image-Edit can now run on **a single 32GB GPU**, or on **a single 24GB GPU with CPU offloading**, while maintaining full model quality. 🔥🔥🔥

### 📊 Performance Comparison

| Model                                          | Model Size | Peak GPU Memory                              | Generation Time (A100 GPU) |
|------------------------------------------------|------------|----------------------------------------------|----------------------------|
| Qwen-Image-Edit (BFloat16)                     | ~41 GB     | OOM                                          | -                          |
| Qwen-Image-Edit (DFloat11)                     | 28.43 GB   | 30.11 GB                                     | 280 seconds                |
| Qwen-Image-Edit (DFloat11 + CPU Offloading)    | 28.43 GB   | 22.71 GB                                     | 570 seconds                |

### 🔧 How to Use

1. Install or upgrade the DFloat11 pip package *(installs the CUDA kernel automatically; requires a CUDA-compatible GPU and PyTorch installed)*:

    ```bash
    pip install -U dfloat11[cuda12]
    ```

2. Install or upgrade diffusers:

    ```bash
    pip install git+https://github.com/huggingface/diffusers
    ```

3. Save the following code to a Python file `qwen_image_edit.py`:

    ```python
    import argparse
    import torch
    from diffusers.utils import load_image
    from diffusers import QwenImageTransformer2DModel, QwenImageEditPipeline
    from transformers.modeling_utils import no_init_weights
    from dfloat11 import DFloat11Model

    def parse_args():
        parser = argparse.ArgumentParser(description='Edit images using Qwen-Image-Edit model')
        parser.add_argument('--cpu_offload', action='store_true', help='Enable CPU offloading')
        parser.add_argument('--cpu_offload_blocks', type=int, default=30, help='Number of transformer blocks to offload to CPU')
        parser.add_argument('--no_pin_memory', action='store_true', help='Disable memory pinning')
        parser.add_argument('--image', type=str, default="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cat.png",
                            help='Path to input image or URL')
        parser.add_argument('--prompt', type=str, default='Add a hat to the cat.',
                            help='Text prompt for image editing')
        parser.add_argument('--negative_prompt', type=str, default=' ',
                            help='Negative prompt for image editing')
        parser.add_argument('--num_inference_steps', type=int, default=50,
                            help='Number of denoising steps')
        parser.add_argument('--true_cfg_scale', type=float, default=4.0,
                            help='Classifier free guidance scale')
        parser.add_argument('--seed', type=int, default=42,
                            help='Random seed for generation')
        parser.add_argument('--output', type=str, default='qwen_image_edit.png',
                            help='Output image path')
        return parser.parse_args()

    args = parse_args()
    model_id = "Qwen/Qwen-Image-Edit"

    with no_init_weights():
        transformer = QwenImageTransformer2DModel.from_config(
            QwenImageTransformer2DModel.load_config(
                model_id, subfolder="transformer",
            ),
        ).to(torch.bfloat16)

    DFloat11Model.from_pretrained(
        "DFloat11/Qwen-Image-Edit-DF11",
        device="cpu",
        cpu_offload=args.cpu_offload,
        cpu_offload_blocks=args.cpu_offload_blocks,
        pin_memory=not args.no_pin_memory,
        bfloat16_model=transformer,
    )

    pipeline = QwenImageEditPipeline.from_pretrained(
        model_id, transformer=transformer, torch_dtype=torch.bfloat16,
    )
    pipeline.enable_model_cpu_offload()
    pipeline.set_progress_bar_config(disable=None)

    image = load_image(args.image)
    inputs = {
        "image": image,
        "prompt": args.prompt,
        "generator": torch.manual_seed(args.seed),
        "true_cfg_scale": args.true_cfg_scale,
        "negative_prompt": args.negative_prompt,
        "num_inference_steps": args.num_inference_steps,
    }

    with torch.inference_mode():
        output = pipeline(**inputs)
        output_image = output.images[0]
        output_image.save(args.output)

    max_gpu_memory = torch.cuda.max_memory_allocated()
    print(f"Max GPU memory allocated: {max_gpu_memory / 1000 ** 3:.2f} GB")
    ```

4. To run without CPU offloading (32GB VRAM required):
    ```bash
    python qwen_image_edit.py
    ```

    To run with CPU offloading (24GB VRAM required, 50GB CPU RAM required):
    ```bash
    python qwen_image_edit.py --cpu_offload
    ```

    If you are getting out of (CPU or GPU) memory errors, try limiting the number of offloaded blocks or disabling memory-pinning:
    ```bash
    # Offload only 12 blocks (offloading more blocks uses less GPU memory and more CPU memory; offloading less blocks is faster):
    python qwen_image_edit.py --cpu_offload --cpu_offload_blocks 12

    # Disable memory-pinning (the most memory efficient way, but could be slower):
    python qwen_image_edit.py --cpu_offload --cpu_offload_blocks 60 --no_pin_memory
    ```

### 🔍 How It Works

We apply **Huffman coding** to losslessly compress the exponent bits of BFloat16 model weights, which are highly compressible (their 8 bits carry only ~2.6 bits of actual information). To enable fast inference, we implement a highly efficient CUDA kernel that performs on-the-fly weight decompression directly on the GPU.

The result is a model that is **~32% smaller**, delivers **bit-identical outputs**, and achieves performance **comparable to the original** BFloat16 model.

Learn more in our [research paper](https://arxiv.org/abs/2504.11651).

### 📄 Learn More

* **Paper**: [70% Size, 100% Accuracy: Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float](https://arxiv.org/abs/2504.11651)
* **GitHub**: [https://github.com/LeanModels/DFloat11](https://github.com/LeanModels/DFloat11)
* **HuggingFace**: [https://huggingface.co/DFloat11](https://huggingface.co/DFloat11)