|
|
--- |
|
|
base_model: |
|
|
- ByteDance-Seed/BAGEL-7B-MoT |
|
|
base_model_relation: quantized |
|
|
pipeline_tag: any-to-any |
|
|
tags: |
|
|
- dfloat11 |
|
|
- df11 |
|
|
- lossless compression |
|
|
- 70% size, 100% accuracy |
|
|
--- |
|
|
|
|
|
# DFloat11 Compressed Model: `ByteDance-Seed/BAGEL-7B-MoT` |
|
|
|
|
|
This model uses **DFloat11** lossless compression. It's 32% smaller than the original BFloat16 model, yet produces bit-identical outputs and runs efficiently on GPUs. |
|
|
|
|
|
### 📊 Performance Comparison |
|
|
|
|
|
| Metric | BAGEL-7B-MoT (BFloat16) | BAGEL-7B-MoT (DFloat11) | |
|
|
| ---------------------------------- | ------------------------- | ------------------------- | |
|
|
| Model Size | 29.21 GB | 19.89 GB | |
|
|
| Peak GPU Memory<br>(1024x1024 image generation) | 30.07 GB | 21.76 GB | |
|
|
| Generation Time<br>(on an A100 GPU) | 54 seconds | 58 seconds | |
|
|
|
|
|
### 🔍 How It Works |
|
|
|
|
|
We apply Huffman coding to the exponent bits of BFloat16 model weights, which are highly compressible. We leverage hardware-aware algorithmic designs to enable highly efficient, on-the-fly weight decompression directly on the GPU. Find out more in our [research paper](https://arxiv.org/abs/2504.11651). |
|
|
|
|
|
### 🔧 How to Use |
|
|
|
|
|
A complete usage guide is available in our GitHub repository (forked from the official Bagel repository): [https://github.com/LeanModels/Bagel-DFloat11](github.com/LeanModels/Bagel-DFloat11). |
|
|
|
|
|
### 📄 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) |
|
|
|
