- min_overlap removed * restrictions and round_to_8
- min_overlap handles tile size > image size by clipping the num tiles to 1.
- Updated assert test on min_overlap.
## What type of PR is this? (check all applicable)
- [X] Refactor
- [X] Feature
- [ ] Bug Fix
- [ ] Optimization
- [X] Documentation Update
- [ ] Community Node Submission
## Have you discussed this change with the InvokeAI team?
- [X] Yes
- [ ] No, because:
## Have you updated all relevant documentation?
- [X] Yes
- [ ] No
## Description
This is the next phase of the model manager refactor, as discussed with
@psychedelicious and @RyanJDick. This implements the model installer,
which is responsible for managing model weights on disk and installing
new models.
Currently only installation of local files and directories is supported.
Remote installation will be implemented after the queued download
manager is reviewed and approved.
Please see the documentation located at
[docs/contributing/MODEL_MANAGER.md](8695ad6f59/docs/contributing/MODEL_MANAGER.md (model-installation))
for an explanation of how this module works.
Things that have changed relative to the current implementation.
1. Model importation runs in a background thread. Access to the
installation status is through a ModelInstallJob object returned by the
`import_model()` call. In addition, the installation process generates a
series of `model_install` events on the event bus.
2. `model_install_progress` events are documented, but not currently
issued. These will be issued when background downloading is implemented.
3. The model installer currently runs in parallel to the current model
manager. The frontend continues to use `configs/models.yaml` and ignores
what is in the `model_config` table of `invokeai.db`.
4. When the installer is initialized at app startup time, it
synchronizes its database to the contents of the InvokeAI `models`
directory. The current model manager does this as well, so you will see
two log messages indicating that this directory is being scanned.
## Related Tickets & Documents
<!--
For pull requests that relate or close an issue, please include them
below.
For example having the text: "closes #1234" would connect the current
pull
request to issue 1234. And when we merge the pull request, Github will
automatically close the issue.
-->
- Related Issue #
- Closes #
## QA Instructions, Screenshots, Recordings
You can test using the FastAPI swagger pages at
http://localhost:9090/docs. Use the calls listed under
`model_manager_v2`. Be aware that only installation of local models
(indicated by their file or directory path) are currently supported.
## Added/updated tests?
- [X] Yes -- see
`tests/app/services/model_install/test_model_install.py`
- [ ] No : _please replace this line with details on why tests
have not been included_
## [optional] Are there any post deployment tasks we need to perform?
In other words, build frontend when creating installer.
Changes to `create_installer.sh`
- If `python` is not in `PATH` but `python3` is, alias them (well, via function). This is needed on some machines to run the installer without symlinking to `python3`.
- Make the messages about pushing tags clearer. The script force-pushes, so it's possible to accidentally take destructive action. I'm not sure how to otherwise prevent damage, so I just added a warning.
- Print out `pwd` when prompting about being in the `installer` dir.
- Rebuild the frontend - if there is already a frontend build, first checks if the user wants to rebuild it.
- Checks for existence of `../build` dir before deleting - if the dir doesn't exist, the script errors and exits at this point.
- Format and spell check.
Other changes:
- Ignore `dist/` folder.
- Delete `dist/`.
**Note: you may need to use `git rm --cached invokeai/app/frontend/web/dist/` if git still wants to track `dist/`.**
Calling `inspect.getmembers()` on a pydantic field results in `getattr` being called on all members of the field. Pydantic has some attrs that are marked deprecated.
In our test suite, we do not filter deprecation warnings, so this is surfaced.
Use a context manager to ignore deprecation warnings when calling the function.
In the latest redux, unknown actions are typed as `unknown`. This forces type-safety upon us, requiring us to be more careful about the shape of actions.
In this case, we don't know if the rejection has a payload or what shape it may be in, so we need to do runtime checks. This is implemented with a simple zod schema, but probably the right way to handle this is to have consistent types in our RTK-Query error logic.
There are a few breaking changes, which I've addressed.
The vast majority of changes are related to new handling of `reselect`'s `createSelector` options.
For better or worse, we memoize just about all our selectors using lodash `isEqual` for `resultEqualityCheck`. The upgrade requires we explicitly set the `memoize` option to `lruMemoize` to continue using lodash here.
Doing that required changing our `defaultSelectorOptions`.
Instead of changing that and finding dozens of instances where we weren't using that and instead were defining selector options manually, I've created a pre-configured selector: `createMemoizedSelector`.
This is now used everywhere instead of `createSelector`.
- update all scripts
- update the frontend GH action
- remove yarn-related files
- update ignores
Yarn classic + storybook has some weird module resolution issue due to how it hoists dependencies.
See https://github.com/storybookjs/storybook/issues/22431#issuecomment-1630086092
When I did the `package.json` solution in this thread, it broke vite. Next option is to upgrade to yarn 3 or pnpm. I chose pnpm.
