* new OffloadingDevice loads one model at a time, on demand
* fixup! new OffloadingDevice loads one model at a time, on demand
* fix(prompt_to_embeddings): call the text encoder directly instead of its forward method
allowing any associated hooks to run with it.
* more attempts to get things on the right device from the offloader
* more attempts to get things on the right device from the offloader
* make offloading methods an explicit part of the pipeline interface
* inlining some calls where device is only used once
* ensure model group is ready after pipeline.to is called
* fixup! Strategize slicing based on free [V]RAM (#2572)
* doc(offloading): docstrings for offloading.ModelGroup
* doc(offloading): docstrings for offloading-related pipeline methods
* refactor(offloading): s/SimpleModelGroup/FullyLoadedModelGroup
* refactor(offloading): s/HotSeatModelGroup/LazilyLoadedModelGroup
to frame it is the same terms as "FullyLoadedModelGroup"
---------
Co-authored-by: Damian Stewart <null@damianstewart.com>
Tensors with diffusers no longer have to be multiples of 8. This broke Perlin noise generation. We now generate noise for the next largest multiple of 8 and return a cropped result. Fixes#2674.
`generator` now asks `InvokeAIDiffuserComponent` to do postprocessing work on latents after every step. Thresholding - now implemented as replacing latents outside of the threshold with random noise - is called at this point. This postprocessing step is also where we can hook up symmetry and other image latent manipulations in the future.
Note: code at this layer doesn't need to worry about MPS as relevant torch functions are wrapped and made MPS-safe by `generator.py`.
Strategize slicing based on free [V]RAM when not using xformers. Free [V]RAM is evaluated at every generation. When there's enough memory, the entire generation occurs without slicing. If there is not enough free memory, we use diffusers' sliced attention.
- implement the following pattern for finding data files under both
regular and editable install conditions:
import invokeai.foo.bar as bar
path = bar.__path__[0]
- this *seems* to work reliably with Python 3.9. Testing on 3.10 needs
to be performed.
1) Downgrade numpy to avoid dependency conflict with numba
2) Move all non ldm/invoke files into `invokeai`. This includes assets, backend, frontend, and configs.
3) Fix up way that the backend finds the frontend and the generator finds the NSFW caution.png icon.
* Update --hires_fix
Change `--hires_fix` to calculate initial width and height based on the model's resolution (if available) and with a minimum size.
