Fix loading Img2Img refiner components in from_single_file (huggingface#7282)

* update

* update

* update

* update

Author: DN6

src/diffusers/loaders/single_file.py

@@ -56,6 +56,8 @@ def build_sub_model_components(
 
     if component_name == "unet":
         num_in_channels = kwargs.pop("num_in_channels", None)
+        upcast_attention = kwargs.pop("upcast_attention", None)
+
         unet_components = create_diffusers_unet_model_from_ldm(
             pipeline_class_name,
             original_config,
@@ -64,6 +66,7 @@ def build_sub_model_components(
             image_size=image_size,
             torch_dtype=torch_dtype,
             model_type=model_type,
+            upcast_attention=upcast_attention,
         )
         return unet_components
 
@@ -300,7 +303,9 @@ def from_single_file(cls, pretrained_model_link_or_path, **kwargs):
                     continue
                 init_kwargs.update(components)
 
-        additional_components = set_additional_components(class_name, original_config, model_type=model_type)
+        additional_components = set_additional_components(
+            class_name, original_config, checkpoint=checkpoint, model_type=model_type
+        )
         if additional_components:
             init_kwargs.update(additional_components)
 

src/diffusers/loaders/single_file_utils.py

@@ -410,7 +410,7 @@ def is_valid_url(url):
     return original_config
 
 
-def infer_model_type(original_config, checkpoint=None, model_type=None):
+def infer_model_type(original_config, checkpoint, model_type=None):
     if model_type is not None:
         return model_type
 
@@ -1279,7 +1279,7 @@ def create_diffusers_unet_model_from_ldm(
     original_config,
     checkpoint,
     num_in_channels=None,
-    upcast_attention=False,
+    upcast_attention=None,
     extract_ema=False,
     image_size=None,
     torch_dtype=None,
@@ -1307,7 +1307,8 @@ def create_diffusers_unet_model_from_ldm(
     )
     unet_config = create_unet_diffusers_config(original_config, image_size=image_size)
     unet_config["in_channels"] = num_in_channels
-    unet_config["upcast_attention"] = upcast_attention
+    if upcast_attention is not None:
+        unet_config["upcast_attention"] = upcast_attention
 
     diffusers_format_unet_checkpoint = convert_ldm_unet_checkpoint(checkpoint, unet_config, extract_ema=extract_ema)
     ctx = init_empty_weights if is_accelerate_available() else nullcontext

tests/pipelines/stable_diffusion_xl/test_stable_diffusion_xl_img2img.py

@@ -838,9 +838,11 @@ def test_single_file_component_configs(self):
         for param_name, param_value in single_file_pipe.unet.config.items():
             if param_name in PARAMS_TO_IGNORE:
                 continue
+            if param_name == "upcast_attention" and pipe.unet.config[param_name] is None:
+                pipe.unet.config[param_name] = False
             assert (
                 pipe.unet.config[param_name] == param_value
-            ), f"{param_name} differs between single file loading and pretrained loading"
+            ), f"{param_name} is differs between single file loading and pretrained loading"
 
         for param_name, param_value in single_file_pipe.vae.config.items():
             if param_name in PARAMS_TO_IGNORE: