Although the changes in #2091 got things to failing at the point of iree-compile of unet for me, the merged version considers vae_encode before unet. It turns out vae_encode has a similar problem with including a precision argument when main branch turbine_models.custom_models.sd_inference doesn't yet include it.
This comments out precision for all the blocks in the the static_kwargs variable which defines how the the various pipeline models are called, and gets us back to an iree-compile failure again.
Changes
Comment out precision being set in static_kwargs for vae_encode when setting up a StableDiffusion pipeline
Comment out precision being set in static_kwargs for vae_decode when setting up a StableDiffusion pipeline
Possible Problems/Concerns
This is will need to be reverted once support for precision is landed in the Turbine models, I've added a TODO comment saying that.
Motivation
Although the changes in #2091 got things to failing at the point of
iree-compile
of unet for me, the merged version considers vae_encode before unet. It turns out vae_encode has a similar problem with including aprecision
argument when main branchturbine_models.custom_models.sd_inference
doesn't yet include it.This comments out
precision
for all the blocks in the thestatic_kwargs
variable which defines how the the various pipeline models are called, and gets us back to aniree-compile
failure again.Changes
Possible Problems/Concerns
precision
is landed in the Turbine models, I've added a TODO comment saying that.