Fixed attention precision not being cast in attention_pytorch and others functions. This led to the functions not being able to respect the `--dont-upcast-attention` flag.
change calls to `.float()` to `.to(dtype=torch.float32)` in several locations, as it profiles much faster.
removed unneeded check for `attn_precision == torch.float32`, as the change from `.float()` to `.to(dtype=cast_to_type)` does not cast or copy if `tensor.dtype == cast_to_type`
I will not add this dependency to the default ones because models in the
spandrel_extra_arches package are non commercial and therefore not
compatible with free software licenses like the one ComfyUI uses.
If you don't mind this you can install it manually yourself.
Extends the core Comfy.SnapToGrid behavior for nodes to apply to LiteGraph's LGraphGroup with the same behavior. Also, pulls out redundant rounding code into util function.
On some keyboards it's apparently too easy to accidentally do CTRL-Delete
when pressing CTRL-Enter repeatedly.
CTRL-Backspace can still be used to clear the workflow.
* Fix to LoadImage Node for #3416 HDR images loading additional smaller images.
Added a blocking if statement in the ImageSequence.Iterator that checks if subsequent images after the first match dimensionally, and prevent them from being appended to output_images if they do not match.
This does not fix or change current behavior for PIL 10.2.0 where the images are loaded at the same size, but it does for 10.3.0 where they are loaded at their correct smaller sizes.
* added list of excluded formats that should return 1 image
added an explicit check for the image format so that additional formats can be added to the list that have problematic behavior.
