LeanModels/DFloat11

DFloat11 [NeurIPS '25]: Lossless Compression of LLMs and DiTs for Efficient GPU Inference

45
/ 100
Emerging

Compresses model weights via Huffman coding of BFloat16 exponents, with on-the-fly GPU decompression during matrix multiplications that discards data immediately after use. Integrates with HuggingFace Transformers and Diffusers, supporting LLMs (Qwen, Llama, Gemma, Phi) and diffusion models (FLUX.1, Wan2.1) with optional CPU offloading to reduce peak GPU memory by 40-80%.

610 stars.

No Package No Dependents
Maintenance 6 / 25
Adoption 10 / 25
Maturity 15 / 25
Community 14 / 25

How are scores calculated?

Stars

610

Forks

37

Language

Python

License

Apache-2.0

Last pushed

Nov 24, 2025

Commits (30d)

0

Get this data via API

curl "https://pt-edge.onrender.com/api/v1/quality/diffusion/LeanModels/DFloat11"

Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.