⚡️ DFloat11: Lossless LLM Compression for Efficient GPU Inference

DFloat11 is a lossless compression framework that reduces the size of LLMs and Diffusion Models by approximately 30% while preserving bit-for-bit identical outputs to the original model. It enables efficient GPU inference on resource-constrained hardware without sacrificing accuracy.

🚀 Key Features

• Lossless Compression: Achieves ~30% model size reduction with outputs identical to the original BFloat16 models.
• GPU-Efficient: All decompression is handled on-GPU, eliminating CPU overhead and host-device data transfers.
• Scalable Performance: Decompression overhead remains constant per forward pass and is independent of batch size.
• Broad Model Support: Compatible with various models, including Qwen3, Gemma3, Llama3, Phi4, Wan2.1, FLUX.1, and BAGEL.

🛠 Installation

Ensure you have a CUDA-compatible GPU and PyTorch installed.

# For CUDA 12
pip install -U dfloat11[cuda12]

# For CUDA 11
pip install -U dfloat11[cuda11]

🧪 Quick Start

For example usage, refer to the examples directory in the GitHub repository.

📄 Learn More

• Paper: 70% Size, 100% Accuracy: Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float
• GitHub Repository: LeanModels/DFloat11
• Hugging Face Models: DFloat11 on Hugging Face