Adds logic to `DiskLatentsStorage.start()` to empty the latents folder on startup.
Adds start and stop methods to `ForwardCacheLatentsStorage`. This is required for `DiskLatentsStorage.start()` to be called, due to how this particular service breaks the direct DI pattern, wrapping the underlying storage with a cache.
## What type of PR is this? (check all applicable)
- [ ] Refactor
- [ ] Feature
- [X] Bug Fix
- [ ] Optimization
- [ ] 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 adds support for at least some of the SDXL embeddings currently
available on Civitai. The embeddings I have tested include:
- https://civitai.com/models/154898/marblingtixl?modelVersionId=173668
- https://civitai.com/models/148131?modelVersionId=167640
-
https://civitai.com/models/123485/hannah-ferguson-or-sdxl-or-comfyui-only-or-embedding?modelVersionId=134674
(said to be "comfyui only")
-
https://civitai.com/models/185938/kendall-jenner-sdxl-embedding?modelVersionId=208785
I am _not entirely sure_ that I have implemented support in the most
elegant way. The issue is that these embeddings have two weight tensors,
`clip_g` and `clip_l`, which correspond to `text_encoder` and
`text_encoder_2` in the main model. When the patcher calls the
ModelPatcher's `apply_ti()` method, I simply check the dimensions of the
incoming text encoder and choose the weights that match the dimensions
of the encoder.
While writing this, I also ran into a possible issue with the Compel
library's `get_pooled_embeddings()` call. It pads the input token list
to the model's max token length and then calls the TI manager to add the
additional tokens from the embedding. However, this ends up making the
input token list longer than the max length, and CLIPTextEncoder crashes
with a tensor size mismatch. I worked around this behavior by making the
TI manager's `expand_textual_inversion_token_ids_if_necessary()` method
remove the excess pads at the end of the token list.
Also note that I have made similar changes to `apply_ti()` in the
ONNXModelPatcher, but haven't tested them yet.
## 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#4401
## QA Instructions, Screenshots, Recordings
<!--
Please provide steps on how to test changes, any hardware or
software specifications as well as any other pertinent information.
-->
## Added/updated tests?
- [ ] Yes
- [X] No : We need to create tests for model patching...
## [optional] Are there any post deployment tasks we need to perform?
IndexedDB has a much larger storage limit than LocalStorage, and is widely supported.
Implemented as a custom storage driver for `redux-remember` via `idb-keyval`. `idb-keyval` is a simple wrapper for IndexedDB that allows it to be used easily as a key-value store.
The logic to clear persisted storage has been updated throughout the app.
