FLUX.1 [dev] -- Flumina Server App

This repository contains an implementation of FLUX.1 [dev] inference on Fireworks AI's new Flumina Server App toolkit.

Deploying FLUX.1 [dev] to Fireworks On-Demand

FLUX.1 [dev] (bfloat16) is available on Fireworks via on-demand deployments. It can be deployed in a few simple steps:

Prerequisite: Clone the Weights into this Repository

Authenticate with HuggingFace with huggingface-cli login
Go to the model repository and accept the license agreement
From the root of this repository, clone the model into the data/ folder by running clone_weights.sh

Prerequisite: Install the Flumina CLI

The Flumina CLI is included with the fireworks-ai Python package. It can be installed with pip like so:

pip install 'fireworks-ai[flumina]>=0.15.7'

Also get an API key from the Fireworks site and set it in the Flumina CLI:

flumina set-api-key YOURAPIKEYHERE

Creating an On-Demand Deployment

flumina deploy can be used to create an on-demand deployment. When invoked with a model name that exists already, it will create a new deployment in your account which has that model:

flumina deploy accounts/fireworks/models/flux-1-dev

When successful, the CLI will print out example commands to call your new deployment, for example:

curl -X POST 'https://api.fireworks.ai/inference/v1/workflows/accounts/fireworks/models/flux-1-dev/control_net?deployment=accounts/u-6jamesr6-63834f/deployments/2a7d8da9' \
    -H 'Authorization: Bearer API_KEY' \
    -F "prompt=<value>" \
    -F "control_image=<value>" \
    -F "control_mode=<value>" \
    -F "aspect_ratio=16:9" \
    -F "guidance_scale=3.5" \
    -F "num_inference_steps=30" \
    -F "seed=0" \
    -F "controlnet_conditioning_scale=<value>"

curl -X POST 'https://api.fireworks.ai/inference/v1/workflows/accounts/fireworks/models/flux-1-dev/text_to_image?deployment=accounts/u-6jamesr6-63834f/deployments/2a7d8da9' \
    -H 'Authorization: Bearer API_KEY' \
    -H "Content-Type: application/json" \
    -d '{
        "prompt": "<value>",
        "aspect_ratio": "16:9",
        "guidance_scale": 3.5,
        "num_inference_steps": 30,
        "seed": 0
    }'

Your deployment can also be administered using the Flumina CLI. Useful commands include:

flumina list deployments to show all of your deployments
flumina get deployment to get details about a specific deployment
flumina delete deployment to delete a deployment

Add-ons

Add-ons are packages you can upload to Fireworks and load into your deployed model. This Flumina app implements two types of add-ons: "controlnet_union" and "lora".

Deploying an existing add-on to your FLUX on-demand deployment

Existing addons can be deployed to Fireworks via the flumina create deployed_addon command. The command takes two arguments:

The resource name of the addon to be deployed
The resource name of the deployment on which to deploy the add-on

Both resource names can either be an ID (e.g. my-addon or 012345 for addons and deployments respectively) in which case it is assumed that the resource is in your Fireworks account or they can be fully-qualified names like accounts/my-account/models/my-addon or accounts/my-account/deployments/my-deployment.

For example:

flumina create deployed_addon accounts/fireworks/models/flux-1-dev-controlnet-union 0123456

This will deploy the flux-1-dev-controlnet-union add-on in the fireworks account to your deployment named 0123456.

Creating custom addons (ControlNet and LoRA adapters)

This model supports addons (specifically ControlNet Union and LoRA adapters) in the HuggingFace diffusers format. Here is an example of taking an existing LoRA addon from HuggingFace and deploying it to Fireworks:

git clone https://huggingface.co/brushpenbob/flux-midjourney-anime
cd flux-midjourney-anime
flumina init addon lora --allow-non-empty
flumina create model flux-1-dev-midjourney-anime --base-model accounts/fireworks/models/flux-1-dev

Then you can deploy it to your on-demand deployment like so:

flumina create deployed_addon flux-1-dev-midjourney-anime 0123456

What is Flumina?

Flumina is Fireworks.ai’s new system for hosting Server Apps that allows users to deploy deep learning inference to production in minutes, not weeks.

What does Flumina offer for FLUX models?

Flumina offers the following benefits:

Clear, precise definition of the server-side workload by looking at the server app implementation (you are here)
Extensibility interface, which allows for dynamic loading/dispatching of add-ons server-side. For FLUX:
- ControlNet (Union) adapters
- LoRA adapters
Off-the-shelf support for standing up on-demand capacity for the Server App on Fireworks
- Further, customization of the logic of the deployment by modifying the Server App and deploying the modified version.

Deploying Custom FLUX.1 [dev] Apps to Fireworks On-demand

Flumina base app upload is coming soon!