fix(qwen-image):

- cond cache registry
- attention backend argument
- fix copies

Author: naykun

src/diffusers/hooks/_helpers.py

@@ -138,6 +138,7 @@ def _register_transformer_blocks_metadata():
     )
     from ..models.transformers.transformer_ltx import LTXVideoTransformerBlock
     from ..models.transformers.transformer_mochi import MochiTransformerBlock
+    from ..models.transformers.transformer_qwenimage import QwenImageTransformerBlock
     from ..models.transformers.transformer_wan import WanTransformerBlock
 
     # BasicTransformerBlock
@@ -240,6 +241,15 @@ def _register_transformer_blocks_metadata():
         ),
     )
 
+    # QwenImage
+    TransformerBlockRegistry.register(
+        model_class=QwenImageTransformerBlock,
+        metadata=TransformerBlockMetadata(
+            return_hidden_states_index=1,
+            return_encoder_hidden_states_index=0,
+        ),
+    )
+
 
 # fmt: off
 def _skip_attention___ret___hidden_states(self, *args, **kwargs):

src/diffusers/models/transformers/transformer_qwenimage.py

@@ -310,7 +310,13 @@ def __call__(
 
         # Compute joint attention
         joint_hidden_states = dispatch_attention_fn(
-            joint_query, joint_key, joint_value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+            joint_query,
+            joint_key,
+            joint_value,
+            attn_mask=attention_mask,
+            dropout_p=0.0,
+            is_causal=False,
+            backend=self._attention_backend,
         )
 
         # Reshape back

src/diffusers/utils/dummy_pt_objects.py

@@ -423,6 +423,21 @@ def from_pretrained(cls, *args, **kwargs):
         requires_backends(cls, ["torch"])
 
 
+class AutoencoderKLQwenImage(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
 class AutoencoderKLTemporalDecoder(metaclass=DummyObject):
     _backends = ["torch"]
 
@@ -1038,6 +1053,21 @@ def from_pretrained(cls, *args, **kwargs):
         requires_backends(cls, ["torch"])
 
 
+class QwenImageTransformer2DModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
 class SanaControlNetModel(metaclass=DummyObject):
     _backends = ["torch"]
 

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -1682,6 +1682,21 @@ def from_pretrained(cls, *args, **kwargs):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class QwenImagePipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
 class ReduxImageEncoder(metaclass=DummyObject):
     _backends = ["torch", "transformers"]