Diffusers / Stable Diffusion in docker with a REST API, supporting various models, pipelines & schedulers. Used by kiri.art, perfect for banana.dev.

Copyright (c) Gadi Cohen, 2022. MIT Licensed. Please give credit and link back to this repo if you use it in a public project.

• Models: stable-diffusion, waifu-diffusion, and easy to add others (e.g. jp-sd)

• Pipelines: txt2img, img2img and inpainting in a single container

  ( all diffusers official and community pipelines are wrapped, but untested)

• All model inputs supported, including setting nsfw filter per request

• Permute base config to multiple forks based on yaml config with vars

• Optionally send signed event logs / performance data to a REST endpoint

• Can automatically download a checkpoint file and convert to diffusers.

• S3 support, dreambooth training.

Note: This image was created for kiri.art. Everything is open source but there may be certain request / response assumptions. If anything is unclear, please open an issue.

• Official docker-diffusers-api Forum: help, updates, discussion.
• Subscribe ("watch") these forum topics for:
  • notable main branch updates
  • notable dev branch updates
• Always check the CHANGELOG for important updates when upgrading.

Official help in our dedicated forum https://banana-forums.dev/c/open-source/docker-diffusers-api/16.

See the dev branch for the latest features. Pull Requests must be submitted against the dev branch.

Firstly, fork and clone this repo.

Most of the configuration happens via docker build variables. You can see all the options in the Dockerfile, and edit them there directly, or set via docker command line or e.g. Banana's dashboard UI once support for build variables land (any day now).

If you're only deploying one container, that's all you need! If you intend to deploy multiple containers each with different variables (e.g. a few different models), you can edit the example scripts/permutations.yaml] file and run scripts/permute.sh to create a number of sub-repos in the permutations directory.

Lastly, there's an option to set MODEL_ID=ALL, and all models will be downloaded, and switched at request time (great for dev, useless for serverless).

Deploying to banana? That's it! You're done. Commit your changes and push.

Building

1. docker build -t diffusers-api --build-arg HF_AUTH_TOKEN=$HF_AUTH_TOKEN .
2. See CONTRIBUTING.md for more helpful hints.
3. Note: your first build can take a really long time, depending on your PC & network speed, and especially when using the CHECKPOINT_URL feature. Great time to grab a coffee or take a walk.

Running

1. docker run -it --gpus all -p 8000:8000 diffusers-api
2. Note: the -it is optional but makes it alot quicker/easier to stop the container using Ctrl-C.
3. If you get a CUDA initialization: CUDA unknown error after suspend, just stop the container, rmmod nvidia_uvm, and restart.

The container expects an HTTP POST request with the following JSON body:

{
  "modelInputs": {
    "prompt": "Super dog",
    "num_inference_steps": 50,
    "guidance_scale": 7.5,
    "width": 512,
    "height": 512,
    "seed": 3239022079
  },
  "callInputs": {
    "MODEL_ID": "runwayml/stable-diffusion-v1-5",
    "PIPELINE": "StableDiffusionPipeline",
    "SCHEDULER": "LMSDiscreteScheduler",
    "safety_checker": true,
  },
}

If you're using banana's SDK, it looks something like this:

const out = await banana.run(apiKey, modelKey, { "modelInputs": modelInputs, "callInputs": callInputs });

NB: if you're coming from another banana starter repo, note that we explicitly name modelInputs above, and send a bigger object (with modelInputs and callInputs keys) for the banana-sdk's "modelInputs" argument.

If provided, init_image and mask_image should be base64 encoded.

Schedulers: docker-diffusers-api is simply a wrapper around diffusers, literally any scheduler included in diffusers will work out of the box, provided it can loaded with its default config and without requiring any other explicit arguments at init time. In any event, the following schedulers are the most common and most well tested: DPMSolverMultistepScheduler (fast! only needs 20 steps!), LMSDiscreteScheduler, DDIMScheduler, PNDMScheduler, EulerAncestralDiscreteScheduler, EulerDiscreteScheduler.

There are also very basic examples in test.py, which you can view and call python test.py if the container is already running on port 8000. You can also specify a specific test, change some options, and run against a deployed banana image:

$ python test.py
Usage: python3 test.py [--banana] [--xmfe=1/0] [--scheduler=SomeScheduler] [all / test1] [test2] [etc]

# Run against http://localhost:8000/ (Nvidia Quadro RTX 5000)
$ python test.py txt2img
Running test: txt2img
Request took 5.9s (init: 3.2s, inference: 5.9s)
Saved /home/dragon/www/banana/banana-sd-base/tests/output/txt2img.png

# Run against deployed banana image (Nvidia A100)
$ export BANANA_API_KEY=XXX
$ BANANA_MODEL_KEY=XXX python3 test.py --banana txt2img
Running test: txt2img
Request took 19.4s (init: 2.5s, inference: 3.5s)
Saved /home/dragon/www/banana/banana-sd-base/tests/output/txt2img.png

# Note that 2nd runs are much faster (ignore init, that isn't run again)
Request took 3.0s (init: 2.4s, inference: 2.1s)

The best example of course is https://kiri.art/ and it's source code.

• 403 Client Error: Forbidden for url

  Make sure you've accepted the license on the model card of the HuggingFace model specified in MODEL_ID, and that you correctly passed HF_AUTH_TOKEN to the container.

You have two options.

1. For a diffusers model, simply set the MODEL_ID docker build variable to the name of the model hosted on HuggingFace, and it will be downloaded automatically at build time.

2. For a non-diffusers model, simply set the CHECKPOINT_URL docker build variable to the URL of a .ckpt file, which will be downloaded and converted to the diffusers format automatically at build time.

Per your personal preferences, rebase or merge, e.g.

1. git fetch upstream
2. git merge upstream/main
3. git push

Or, if you're confident, do it in one step with no confirmations:

git fetch upstream && git merge upstream/main --no-edit && git push

Check scripts/permute.sh and your git remotes, some URLs are hardcoded, I'll make this easier in a future release.

Set CALL_URL and SIGN_KEY environment variables to send timing data on init and inference start and end data. You'll need to check the source code of here and sd-mui as the format is in flux.

This info is now logged regardless, and init() and inference() times are sent back via { $timings: { init: timeInMs, inference: timeInMs } }.

Originally based on https://github.com/bananaml/serverless-template-stable-diffusion.