li-lizhe 3ef498e9fd fix(autoencoder_kl): use x.device instead of hardcoded cuda in tiled_encode ...

Tile tensor is a slice of the input x, so tile.device equals x.device.
Using x.device instead of hardcoded "cuda" makes the tiled encoder
work on any accelerator (CUDA, NPU, MPS, XPU) without crashing.

The tile's device is already inherited from x, so .to(x.device) is
a no-op on CUDA — the change only affects non-CUDA backends.

2026-09-19 22:11:26 +08:00

..

__init__.py

major refactoring of modules

2025-07-03 09:18:38 -04:00

autoencoder_kl.py

fix(autoencoder_kl): use x.device instead of hardcoded cuda in tiled_encode

2026-09-19 22:11:26 +08:00

pixelsmith_pipeline.py

- all guidance values set to -1 to enable using model defaults

2026-05-16 11:06:23 +02:00

vae.py

major refactoring of modules

2025-07-03 09:18:38 -04:00