Thread by @DeepSQUAREio

DeepSquare integrates the neural network #text2image model #StableDiffusion.

This thread presents how we #designed and implemented the #workflow to execute Stable Diffusion on the DeepSquare Grid.

We designed the workflow as follows:

The #prompt is sent over an environment variable.
Most of the parameters are sent via environment varibles.
Images are sent to transfer.deepsquare.run

The docker image is already compiled from github.com/Stability-AI/s… and is exported as registry-1.deepsquare.run/library/stable…

The steps are as follows:
1-Generate the images. Rows are distributed into task per #GPU.
2-Combine all the images into one grid.

Implementation​
We with the input, output and resources.
For the sake of readability, we use YAML. You can use an online YAML to #JSON converter if you wish to execute the workflow on the dev environment:
app.deepsquare.run/sandbox

Resource allocation, environment variable and input/output​. We will use 4 parallel tasks. Each task has 8 #CPU, 8 GB of #RAM, 1 #GPU

1. Inference loop: we need to execute a for directive which executes 4 steps in parallel.

DeepSquare already hosts its own models at /opt/models. If you plan to use your own model, you should import the model using the input directive.

2. Combine the rows​
Our docker image already includes #ImageMagick, meaning we can compose a new image from the generated images.

We send the "grid" image to transfer.deepsquare.run. The output directory contains all the images in their original resolution

Conclusion​
We have developed an #ML inference workflow. Most inference processes with an output will follow the same pattern:

-Launch the inference in parallel
-Reassemble the output

Learn more:
docs.deepsquare.run/docs/deploy-de…