Merge branch 'main' into transformers-v5-pr

Author: sayakpaul

src/diffusers/pipelines/prx/__init__.py

@@ -24,14 +24,25 @@
 else:
     _import_structure["pipeline_prx"] = ["PRXPipeline"]
 
-# Import T5GemmaEncoder for pipeline loading compatibility
+# Wrap T5GemmaEncoder to pass config.encoder (T5GemmaModuleConfig) instead of the
+# composite T5GemmaConfig, which lacks flat attributes expected by T5GemmaEncoder.__init__.
 try:
     if is_transformers_available():
         import transformers
-        from transformers.models.t5gemma.modeling_t5gemma import T5GemmaEncoder
+        from transformers.models.t5gemma.modeling_t5gemma import T5GemmaEncoder as _T5GemmaEncoder
+
+        class T5GemmaEncoder(_T5GemmaEncoder):
+            @classmethod
+            def from_pretrained(cls, pretrained_model_name_or_path, *args, **kwargs):
+                if "config" not in kwargs:
+                    from transformers.models.t5gemma.configuration_t5gemma import T5GemmaConfig
+
+                    config = T5GemmaConfig.from_pretrained(pretrained_model_name_or_path)
+                    if hasattr(config, "encoder"):
+                        kwargs["config"] = config.encoder
+                return super().from_pretrained(pretrained_model_name_or_path, *args, **kwargs)
 
         _additional_imports["T5GemmaEncoder"] = T5GemmaEncoder
-        # Patch transformers module directly for serialization
         if not hasattr(transformers, "T5GemmaEncoder"):
             transformers.T5GemmaEncoder = T5GemmaEncoder
 except ImportError: