Added README.md

README.md

@@ -0,0 +1,69 @@
+# flux-fine-tuner
+
+## Development
+
+Before submitting a PR, format the code and run the linter locally.
+
+```shell
+pip install -r requirements-test.txt  # only need to do this once
+ruff format
+ruff check
+```
+
+## Deploy to production
+
+### Push model
+
+Manually trigger the [Github Push action](https://github.com/replicate/flux-fine-tuner/actions/workflows/push.yaml). Uncheck the "Test only, without pushing to prod" button. You might also have to uncheck the "Compare outputs..." button.
+
+The push action takes half an hour since it tests both training and inference. But it's worth doing to be safe.
+
+Once you've deployed to production, make a test fine-tune and run predictions on the trained model, just to be super sure it works.
+
+### Update hotswap bases
+
+When you've tested the pushed model, you need to update all existing fine-tuned versions to use the predictor you just pushed.
+
+First, only update a single model or a subset of models to test that this step works.
+
+In your local checkout of [web](https://github.com/replicate/web), run
+
+```shell
+script/manage-prod update_hotswap_base_version \
+    --from-model ostris/flux-dev-lora-trainer \
+    --to-latest \
+    --trained-version-filter="<your-username>/<your-model-name>"
+```
+
+This is a dry run to list the versions that will be updated. The `--trained-version-filter` can be just "<your-username>" if you want to test all your models.
+
+When you're happy to with the list of models that will be updated, run
+
+```shell
+script/manage-prod update_hotswap_base_version \
+    --from-model ostris/flux-dev-lora-trainer \
+    --to-latest \
+    --trained-version-filter="<your-username>/<your-model-name>" \
+    --force
+```
+
+Now you can test that the predictor works for the updated model(s).
+
+When you're happy with that, do the same thing but for all models. First, a dry run:
+
+```shell
+script/manage-prod update_hotswap_base_version \
+    --from-model ostris/flux-dev-lora-trainer \
+    --to-latest \
+```
+
+And then actually update all the trained Flux versions:
+
+```shell
+script/manage-prod update_hotswap_base_version \
+    --from-model ostris/flux-dev-lora-trainer \
+    --to-latest \
+    --force
+```
+
+This process is being improved -- soon it will be possible to configure base versions in Django admin.