flux fp8

Introduction

The FLUX-FP8 model developed by Kijai on Hugging Face features specific float8 tensor formats, particularly float8_e4m3fn and float8_e5m2. These formats are applied to various model weights sourced from different labs.

Architecture

The primary focus of the FLUX-FP8 model is on utilizing the float8 tensor formats: float8_e4m3fn and float8_e5m2. This architecture supports efficient model performance and compatibility by maintaining original upload names for backward compatibility.

Training

The model utilizes float8_e4m3fn weights sourced from multiple Hugging Face repositories, including Black Forest Labs and Shakker Labs. These weights are specifically designed to optimize neural network computations by leveraging the efficiency of float8 tensor formats.

Guide: Running Locally

1. Clone the Repository: Access the model repository on Hugging Face and clone it to your local environment.
2. Install Dependencies: Ensure you have the necessary packages installed, potentially using pip for Python environments.
3. Load the Model: Utilize frameworks such as PyTorch or TensorFlow to load the model weights, ensuring compatibility with float8 formats.
4. Run Inference: Execute inference tasks locally, taking advantage of the model's efficient architecture.

For enhanced performance, consider using cloud GPUs from providers like AWS or Google Cloud.

License

• The file flux1-dev-fp8.safetensors and its variants fall under the FLUX.1 [dev] Non-Commercial License.
• The file flux1-schnell-fp8.safetensors is licensed under Apache-2.0 License.