PyTorch chose to make it Any because they expect its users' code
to be "highly dynamic": https://github.com/pytorch/pytorch/pull/104321
It is not the case for us, in Refiners having untyped code
goes contrary to one of our core principles.
Note that there is currently an open PR in PyTorch to
return `Module | Tensor`, but in practice this is not always
correct either: https://github.com/pytorch/pytorch/pull/115074
I also moved Residuals-related code from SD1 to latent_diffusion
because SDXL should not depend on SD1.
i.e. CLIPTextModel by default or CLIPTextModelWithProjection for SDXL
so-called text_encoder_2
This silent false positive warnings like:
Some weights of CLIPTextModelWithProjection were not initialized
from the model checkpoint [...]
Ran successfully to completion. But on a repeat run `convert_unclip` didn't pass the hash check for some reason.
- fix inpainting model download urls
- shows a progress bar for downloads
- skips downloading existing files
- uses a temporary file to prevent partial downloads
- can do a dry run to check if url is valid `DRY_RUN=1 python scripts/prepare_test_weights.py`
- displays the downloaded file hash
And replaced the remaining Sum-Identity layers by Residual.
The tolerance used to compare SAM's ViT models has been tweaked: for
some reasons there is a small difference (in float32) in the neck layer
(first conv2D)
Co-authored-by: Cédric Deltheil <cedric@deltheil.me>
This generalizes the Adapter abstraction to higher-level
constructs such as high-level LoRA (targeting e.g. the
SD UNet), ControlNet and Reference-Only Control.
Some adapters now work by adapting child models with
"sub-adapters" that they inject / eject when needed.
