This repository provides an optimized FP8 (float8_e4m3fn) weight-only quantized version of the newly released Krea 2 OSS (Turbo) transformer.
This optimization reduces the model size from the original 24.76 GiB (BF16) down to 12.01 GiB, making it highly accessible and runnable on standard consumer hardware (such as 16GB and 24GB GPUs) without sacrificing output quality.
⚠️ Licensing & Disclaimer
Original Model Creators: All credit goes to KREA.ai for the original research, architecture, and weights.
License: This model is subject to the KREA 2 License Agreement. Please read and comply with the official license terms before using these weights: KREA 2 Licensing Terms.
Purpose: This repository is a community-contributed utility. It does not claim ownership of the original model or architecture. Its sole purpose is to provide optimized, consumer-hardware-friendly weights for the open-source community.
🛠️ Quantization Details (Quality-First FP8)
Unlike generic global quantization scripts that aggressively convert every parameter (which often degrades generation details or introduces NaN/promotion calculation errors in neural networks), this model was quantized using a selective weight-only strategy:
Targeted Quantization: Only 2D floating-point weight matrices (.weight keys with ndim >= 2 and element count > 1024) were quantized to torch.float8_e4m3fn.
Preserved Precision:
All 1D vectors, biases, and normalization scales are kept in their native high-precision (float32 / bfloat16).
Highly sensitive projection/modulation layers (such as LastLayer.modulation.lin vectors) are completely preserved in high-precision. This prevents typical mathematical promotion bugs (such as BFloat16 and Float8 promotion issues in PyTorch) and retains original output fidelity.
Special thanks to the KREA.ai team for releasing Krea 2 to the open-source community. For any commercial licensing inquiries or details about the model, please visit krea.ai.
