Flue: Fast Lightweight Unified Engine for Image Diffusion

Flue is a high-performance API server designed for efficiently loading and running Text-to-Image diffusion models. Built in Rust, Flue leverages the HuggingFace Candle, Transformers, and Hub libraries, delivering fast, scalable, and optimized model inference.

Features

• Unified API: Easily load and serve multiple diffusion models through a consistent interface.
• GPU Acceleration: Supports single and multi-GPU deployments across one or multiple nodes, leveraging CUDA and Metal.
• Optimized Pipelines:
  • FlashAttention
  • Quantization
  • Compiled computation graphs
  • Advanced parallelization techniques for maximum throughput and efficiency.
• Model Support: Currently supports Flux model family, with planned expansions to additional Candle-supported diffusion models.

Installation

Prerequisites

• Rust
• CUDA toolkit (if using NVIDIA GPUs)
• Metal (if deploying on Apple hardware)

Quickstart

Clone and build Flue:

git clone https://github.com/Apsu/flue.git
cd flue
cargo build --release --features cuda

Note: cudnn feature can be added if you have it installed for potential improved performance

Run the API server:

./target/release/flue-server

Configuration

Flue accepts command-line arguments to configure the engine.

--model "path or HF repo"   # Default: black-forest-labs/FLUX.1-schnell
--host 0.0.0.0              # Default 127.0.0.1
--port 1234                 # Default: 8080
--cpu                       # Force CPU offloading

Note: CPU offloading is currently all-or-nothing, and takes a lot of RAM and CPU for even a small request!

Usage

Send an HTTP POST request to generate images:

curl -X POST http://localhost:8080/v1/images/generations \
-H "Content-Type: application/json" \
-d '{
  "prompt": "A futuristic city skyline at sunset",
  "width": 512,
  "height": 512,
  "steps": 4,
  "guidance_scale": 7.5
}'

Note: FLUX.1-schnell ignores guidance_scale and requires less steps than FLUX.1-dev

Roadmap

• ✅ Flux models support
• ✅ FlashAttention integration
• ⬜ Broader model compatibility (Stable Diffusion, Hunyuan, etc.)
• ⬜ Distributed inference across multiple GPUs/nodes
• ⬜ Parallel optimization methods like PipeFusion, Ulysses, Ring, and DistVAE
• ⬜ Graph compilation and save/load
• ⬜ Advanced caching and resource management

Contributing

Contributions are welcome! Please open issues and pull requests directly on this repository.

License

Flue is licensed under MIT.