* attention maps saving to /tmp
* tidy up diffusers branch backporting of cross attention refactoring
* base64-encoding the attention maps image for generationResult
* cleanup/refactor conditioning.py
* attention maps and tokens being sent to web UI
* attention maps: restrict count to actual token count and improve robustness
* add argument type hint to image_to_dataURL function
Co-authored-by: psychedelicious <4822129+psychedelicious@users.noreply.github.com>
Co-authored-by: damian <git@damianstewart.com>
Co-authored-by: psychedelicious <4822129+psychedelicious@users.noreply.github.com>
* add windows to test runners
* disable fail-fast for debugging
* re-enable login shell for conda workflow
also fix expression to exclude windows from run tests
* enable fail-fast again
* fix condition, pin runner verisons
* remove feature branch from push trigger
since already being triggered now via PR
* use gfpgan from pypi for windows
curious if this would fix the installation here as well
since worked for #1802
* unpin basicsr for windows
* for curiosity enabling testing for windows as well
* disable pip cache
since windows failed with a memory error now
but was working before it had a cache
* use matrix.github-env
* set platform specific root and outdir
* disable tests for windows since memory error
I guess the windows installation uses more space than linux
and for this they have less swap memory
In the event where no `init_mask` is given and `invert_mask` is set to True, the script will raise the following error:
```bash
AttributeError: 'NoneType' object has no attribute 'mode'
```
The new implementation will only run inversion when both variables are valid.
* update docker build (cloud) action with additional metadata, new labels
* (docker) also add aarch64 cloud build and remove arch suffix
* (docker) architecture suffix is needed for now
* (docker) don't build aarch64 for now
* add docker build optimized for size; do not copy models to image
useful for cloud deployments. attempts to utilize docker layer
caching as effectively as possible. also some quick tools to help with
building
* add workflow to build cloud img in ci
* push cloud image in addition to building
* (ci) also tag docker images with git SHA
* (docker) rework Makefile for easy cache population and local use
* support the new conda-less install; further optimize docker build
* (ci) clean up the build-cloud-img action
* improve the Makefile for local use
* move execution of invoke script from entrypoint to cmd, allows overriding the cmd if needed (e.g. in Runpod
* remove unnecessary copyright statements
* (docs) add a section on running InvokeAI in the cloud using Docker
* (docker) add patchmatch to the cloud image; improve build caching; simplify Makefile
* (docker) fix pip requirements path to use binary_installer directory
prompt token sequences begin with a "beginning-of-sequence" marker <bos> and end with a repeated "end-of-sequence" marker <eos> - to make a default prompt length of <bos> + 75 prompt tokens + <eos>. the .swap() code was failing to take the column for <bos> at index 0 into account. the changes here do that, and also add extra handling for a single <eos> (which may be redundant but which is included for completeness).
based on my understanding and some assumptions about how this all works, the reason .swap() nevertheless seemed to do the right thing, to some extent, is because over multiple steps the conditioning process in Stable Diffusion operates as a feedback loop. a change to token n-1 has flow-on effects to how the [1x4x64x64] latent tensor is modified by all the tokens after it, - and as the next step is processed, all the tokens before it as well. intuitively, a token's conditioning effects "echo" throughout the whole length of the prompt. so even though the token at n-1 was being edited when what the user actually wanted was to edit the token at n, it nevertheless still had some non-negligible effect, in roughly the right direction, often enough that it seemed like it was working properly.
