diffusion-rs-accelerate

diffusion-rs

Blazingly fast inference of diffusion models.

| Rust Documentation | Python Documentation | Discord |

Features

• Quantization
  • bitsandbytes format (fp4, nf4, and int8)
  • GGUF (2-8 bit quantization)
• Easy: Strong support for running 🤗 DDUF models.
• Strong Apple Silicon support: support for the Metal, Accelerate, and ARM NEON frameworks
• Support for NVIDIA GPUs with CUDA
• AVX support for x86 CPUs
• Allow acceleration of models larger than the total VRAM size with offloading

Please do not hesitate to contact us with feature requests via Github issues!

Upcoming features

• 🚧 LoRA support
• 🚧 CPU + GPU inference with automatic offloading to allow partial acceleration of models larger than the total VRAM

Installation

Check out the installation guide for details about installation.

Examples

After installing, you can try out these examples!

Download the DDUF file here: wget https://huggingface.co/DDUF/FLUX.1-dev-DDUF/resolve/main/FLUX.1-dev-Q4-bnb.dduf

CLI:

diffusion_rs_cli --scale 3.5 --num-steps 50 dduf -f FLUX.1-dev-Q4-bnb.dduf

More CLI examples here.

Python:

More Python examples here.

from diffusion_rs import DiffusionGenerationParams, ModelSource, Pipeline
from PIL import Image
import io

pipeline = Pipeline(source=ModelSource.DdufFile("FLUX.1-dev-Q4-bnb.dduf"))

image_bytes = pipeline.forward(
    prompts=["Draw a picture of a sunrise."],
    params=DiffusionGenerationParams(
        height=720, width=1280, num_steps=50, guidance_scale=3.5
    ),
)

image = Image.open(io.BytesIO(image_bytes[0]))
image.show()

Rust crate:

Examples with the Rust crate: here.

use std::time::Instant;

use diffusion_rs_core::{DiffusionGenerationParams, ModelSource, Offloading, Pipeline, TokenSource};
use tracing::level_filters::LevelFilter;
use tracing_subscriber::EnvFilter;

let filter = EnvFilter::builder()
    .with_default_directive(LevelFilter::INFO.into())
    .from_env_lossy();
tracing_subscriber::fmt().with_env_filter(filter).init();

let pipeline = Pipeline::load(
    ModelSource::dduf("FLUX.1-dev-Q4-bnb.dduf")?,
    false,
    TokenSource::CacheToken,
    None,
    None,
)?;

let start = Instant::now();

let images = pipeline.forward(
    vec!["Draw a picture of a sunrise.".to_string()],
    DiffusionGenerationParams {
        height: 720,
        width: 1280,
        num_steps: 50,
        guidance_scale: 3.5,
    },
)?;

let end = Instant::now();
println!("Took: {:.2}s", end.duration_since(start).as_secs_f32());

images[0].save("image.png")?;

Support matrix

Model	Supports DDUF	Supports quantized DDUF
FLUX.1 Dev/Schnell	✅	✅

Contributing

• Anyone is welcome to contribute by opening PRs
  • See good first issues for a starting point!
• Collaborators will be invited based on past contributions