Skip to main content

Devin post-training a FLUX LoRA on a Modal GPU, deploying the best checkpoint, and adding it to a chat app

Devin runs wherever the job needs to run. With Devin Outposts, a session can run on your own infrastructure or a partner platform; here it runs in a Modal GPU sandbox, so the same Devin that writes your code can train, evaluate, and deploy a model end to end. In this tutorial Devin fine-tunes a FLUX LoRA on 18 otter images so that “Devin the software engineer” renders as an otter, deploys it as a Modal endpoint, and adds it to the model picker of the example app, otterchat.

​
Before you start

  • Create a Modal outpost: run uvx modal-devin init and pick a GPU for the generated Modal app. See Run Devin on Modal and the Outposts quickstart.
  • Connect your repository: the example repo, otterchat, is a chat app with an image model picker. The outpost checks it out onto a persistent Modal Volume.
  • Prepare your training data: add captioned images to the repository or a Modal Volume the outpost mounts. The example uses 18 cartoon otter images with captions like “Devin the software engineer in a data center”.
  • Add secrets: store your Weights & Biases key under Settings > Secrets. The Modal sandbox already carries Modal credentials, so Devin can deploy endpoints without extra setup.
  • Set rules for long runs: add organization rules (or put them in the prompt) such as: post a progress update every 15 minutes, render validation images at every checkpoint, and include the preview link in every PR.
  • Connect Slack (optional): training runs take hours, so sync the session to a Slack thread to follow along.

​
Step 1: Start a session on the outpost

Open Devin, choose your Modal outpost from the Virtual environment menu below the prompt box, and turn on the Slack toggle in the composer. Check the hardware:
Devin reports an NVIDIA B200 with PyTorch and diffusers installed, and the project on a persistent Modal Volume.

​
Step 2: Check the training data

Devin finds the dataset in the project and posts the 18 images, with their captions, to the session and the Slack thread.

​
Step 3: Kick off the training run

Devin runs the job on its own, following the rules above:
  1. Trains the LoRA: about 1,000 steps, logged to Weights & Biases, with validation renders every 100 steps.
  2. Reports to Slack: progress every 15 minutes with the latest renders. Early renders are humans; the otter appears as training continues.
  3. Picks the best checkpoint: step 700, since steps 900 and 1,000 showed signs of overfitting.
  4. Deploys and ships: serves step 700 as a serverless Modal endpoint, adds otterimage 1.0 to the chat app’s model picker, records an end-to-end test, and opens a PR with a preview link.
Grid of validation images at each checkpoint

Validation renders from step 500 to step 1,000

​
Step 4: Try the model

Open the preview link from the PR and ask for an image:
The chat app returning a cartoon otter eating spaghetti
The same prompt on a general-purpose image model returns a human.

​
Step 5: Iterate with child sessions

Post-training is rarely one run. Hand Devin a paper and let it try the alternatives in parallel:
Devin adds the new methods to the training script in a PR and starts three child sessions on the outpost, one each for LoHa, LoKr, and native fine-tuning.
Devin's session showing three child sessions for LoHa, LoKr, and native fine-tuning

​
Next steps

Run Devin on Modal

Create a Modal-backed outpost with GPUs, Volumes, and Secrets

Devin Outposts

Run sessions on your own infrastructure or a partner platform

Slack integration

Follow long-running sessions from a Slack thread

Managed Devins

Spin up and monitor child sessions for parallel experiments