PRXPipeline.from_pretrained() broken with transformers 5.1.0

[DavidBert]

Describe the bug

PRXPipeline.from_pretrained() fails when loading the T5GemmaEncoder text encoder component with transformers 5.1.0:

AttributeError: 'T5GemmaConfig' object has no attribute 'attention_dropout'

Root cause

In transformers 5.x, T5GemmaConfig was refactored into a composite config with encoder/decoder sub-configs (T5GemmaModuleConfig). T5GemmaEncoder.__init__ expects flat attributes like config.attention_dropout which only exist on T5GemmaModuleConfig, not on the composite T5GemmaConfig. When the pipeline loads the text encoder via from_pretrained, it passes the composite config instead of the encoder sub-config.

Reproduction

from diffusers import PRXPipeline

pipe = PRXPipeline.from_pretrained("Photoroom/prx-1024-t2i-beta")
# AttributeError: 'T5GemmaConfig' object has no attribute 'attention_dropout'

Logs

System Info

Environment
transformers 5.1.0

Who can help?

No response