Skip to content

Create a new Flux fine-tune #1

Create a new Flux fine-tune

Create a new Flux fine-tune #1

Workflow file for this run

name: Train a model
# This workflow is triggered manually from the GitHub website.
# To run it, click the "Actions" tab on the repo page.
on:
workflow_dispatch:
inputs:
destination:
required: true
description: The name of the Replicate model to publish to, in the format `username/model`. The model must already exist.
trigger_word:
required: true
description: A short and unique non-word string representing your custom trained style or concept. This should be an uncommon string of letters like `ZXZX`, rather than a common word or name like `sarah` of `dog`. You'll use this string in prompts when running the model like "a pencil sketch of ZXZX sitting in a meadow".
steps:
required: true
description: Total number of training steps to perform. Higher numbers produce more accurate results, but take longer to run. Set this to something low like 100 to test that your workflow is working, then run again at a higher number like 1000 to train more accurately.
type: number
default: 1000
trainer_model:
required: true
description: The name of the trainer model to use, in the format `owner/model`.
type: string
default: "ostris/flux-dev-lora-trainer"
trainer_version:
required: true
description: The version of the trainer model to use.
type: string
default: "7f53f82066bcdfb1c549245a624019c26ca6e3c8034235cd4826425b61e77bec"
# Create a Replicate API token at https://replicate.com/account/api-tokens
env:
replicate-api-token: ${{ secrets.REPLICATE_API_TOKEN }}
jobs:
train:
runs-on: ubuntu-latest
steps:
- name: Check secrets
if: ${{ env.replicate-api-token == '' }}
run: |
echo "🙈 Uh oh! Missing repository secret: REPLICATE_API_TOKEN"
echo "Go to https://replicate.com/account to copy your API token,"
echo "then visit https://github.com/${{ github.repository }}/settings/secrets/actions/new to set it."
echo
echo
echo
exit 1
- name: Checkout code
uses: actions/checkout@v3
- name: Zip training data
run: |
zip -r data.zip data
- name: Upload training data
id: upload-training-data
run: |
FILE_UPLOAD_RESPONSE=$(
curl -s -X POST "https://api.replicate.com/v1/files" \
-H "Authorization: Token $REPLICATE_API_TOKEN" \
-H "Content-Type: multipart/form-data" \
-F "[email protected];type=application/zip;filename=data.zip"
)
echo "INSTANCE_DATA_URL=$(jq -r '.urls.get' <<< $FILE_UPLOAD_RESPONSE)" >> $GITHUB_OUTPUT
- name: Start training
run: |
curl -s -X POST \
-H "Authorization: Token ${{ secrets.REPLICATE_API_TOKEN }}" \
-H "Content-Type: application/json" \
-d '{
"destination": "${{ inputs.destination }}",
"input": {
"trigger_word": "${{ inputs.trigger_word }}",
"input_images": "${{ steps.upload-training-data.outputs.INSTANCE_DATA_URL }}",
"steps": ${{ inputs.steps }}
},
}' \
:
https://api.replicate.com/v1/models/${{ inputs.trainer_model }}/versions/${{ inputs.trainer_version }}/trainings
- name: Link to model
run: |
echo "🚂 Your model is now training!"
echo "This takes a few minutes, depending on how many training steps you're running."
echo "To see your model, visit https://replicate.com/${{ inputs.destination }} and refresh until you see the prediction form."