Merge branch 'main' into fix_syntax_qwenimageEditPlus

SahilCarterr · web-flow · commit 8801ea193049 · 2025-09-24T17:52:27.000+05:30
diff --git a/src/diffusers/models/auto_model.py b/src/diffusers/models/auto_model.py
@@ -194,7 +194,7 @@ def from_pretrained(cls, pretrained_model_or_path: Optional[Union[str, os.PathLi
 
         has_remote_code = "auto_map" in config and cls.__name__ in config["auto_map"]
         trust_remote_code = resolve_trust_remote_code(trust_remote_code, pretrained_model_or_path, has_remote_code)
-        if not (has_remote_code and trust_remote_code):
+        if not has_remote_code and trust_remote_code:
             raise ValueError(
                 "Selected model repository does not happear to have any custom code or does not have a valid `config.json` file."
             )
diff --git a/src/diffusers/modular_pipelines/components_manager.py b/src/diffusers/modular_pipelines/components_manager.py
@@ -25,6 +25,7 @@
     is_accelerate_available,
     logging,
 )
+from ..utils.torch_utils import get_device
 
 
 if is_accelerate_available():
@@ -161,7 +162,9 @@ def __call__(self, hooks, model_id, model, execution_device):
 
         current_module_size = model.get_memory_footprint()
 
-        mem_on_device = torch.cuda.mem_get_info(execution_device.index)[0]
+        device_type = execution_device.type
+        device_module = getattr(torch, device_type, torch.cuda)
+        mem_on_device = device_module.mem_get_info(execution_device.index)[0]
         mem_on_device = mem_on_device - self.memory_reserve_margin
         if current_module_size < mem_on_device:
             return []
@@ -301,7 +304,7 @@ class ComponentsManager:
         cm.add("vae", vae_model, collection="sdxl")
 
         # Enable auto offloading
-        cm.enable_auto_cpu_offload(device="cuda")
+        cm.enable_auto_cpu_offload()
 
         # Retrieve components
         unet = cm.get_one(name="unet", collection="sdxl")
@@ -490,6 +493,8 @@ def remove(self, component_id: str = None):
             gc.collect()
             if torch.cuda.is_available():
                 torch.cuda.empty_cache()
+            if torch.xpu.is_available():
+                torch.xpu.empty_cache()
 
     # YiYi TODO: rename to search_components for now, may remove this method
     def search_components(
@@ -678,7 +683,7 @@ def matches_pattern(component_id, pattern, exact_match=False):
 
         return get_return_dict(matches, return_dict_with_names)
 
-    def enable_auto_cpu_offload(self, device: Union[str, int, torch.device] = "cuda", memory_reserve_margin="3GB"):
+    def enable_auto_cpu_offload(self, device: Union[str, int, torch.device] = None, memory_reserve_margin="3GB"):
         """
         Enable automatic CPU offloading for all components.
 
@@ -704,6 +709,8 @@ def enable_auto_cpu_offload(self, device: Union[str, int, torch.device] = "cuda"
 
         self.disable_auto_cpu_offload()
         offload_strategy = AutoOffloadStrategy(memory_reserve_margin=memory_reserve_margin)
+        if device is None:
+            device = get_device()
         device = torch.device(device)
         if device.index is None:
             device = torch.device(f"{device.type}:{0}")
diff --git a/src/diffusers/modular_pipelines/modular_pipeline.py b/src/diffusers/modular_pipelines/modular_pipeline.py
@@ -323,7 +323,7 @@ def from_pretrained(
         trust_remote_code = resolve_trust_remote_code(
             trust_remote_code, pretrained_model_name_or_path, has_remote_code
         )
-        if not (has_remote_code and trust_remote_code):
+        if not has_remote_code and trust_remote_code:
             raise ValueError(
                 "Selected model repository does not happear to have any custom code or does not have a valid `config.json` file."
             )
diff --git a/src/diffusers/pipelines/wan/pipeline_wan_vace.py b/src/diffusers/pipelines/wan/pipeline_wan_vace.py
@@ -795,7 +795,7 @@ def __call__(
             callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
 
         # Simplification of implementation for now
-        if not isinstance(prompt, str):
+        if prompt is not None and not isinstance(prompt, str):
             raise ValueError("Passing a list of prompts is not yet supported. This may be supported in the future.")
         if num_videos_per_prompt != 1:
             raise ValueError(
diff --git a/tests/lora/test_lora_layers_auraflow.py b/tests/lora/test_lora_layers_auraflow.py
@@ -43,7 +43,6 @@
 class AuraFlowLoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
     pipeline_class = AuraFlowPipeline
     scheduler_cls = FlowMatchEulerDiscreteScheduler
-    scheduler_classes = [FlowMatchEulerDiscreteScheduler]
     scheduler_kwargs = {}
 
     transformer_kwargs = {
diff --git a/tests/lora/test_lora_layers_cogvideox.py b/tests/lora/test_lora_layers_cogvideox.py
@@ -21,7 +21,6 @@
 
 from diffusers import (
     AutoencoderKLCogVideoX,
-    CogVideoXDDIMScheduler,
     CogVideoXDPMScheduler,
     CogVideoXPipeline,
     CogVideoXTransformer3DModel,
@@ -44,7 +43,6 @@ class CogVideoXLoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
     pipeline_class = CogVideoXPipeline
     scheduler_cls = CogVideoXDPMScheduler
     scheduler_kwargs = {"timestep_spacing": "trailing"}
-    scheduler_classes = [CogVideoXDDIMScheduler, CogVideoXDPMScheduler]
 
     transformer_kwargs = {
         "num_attention_heads": 4,
diff --git a/tests/lora/test_lora_layers_cogview4.py b/tests/lora/test_lora_layers_cogview4.py
@@ -50,7 +50,6 @@ def from_pretrained(*args, **kwargs):
 class CogView4LoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
     pipeline_class = CogView4Pipeline
     scheduler_cls = FlowMatchEulerDiscreteScheduler
-    scheduler_classes = [FlowMatchEulerDiscreteScheduler]
     scheduler_kwargs = {}
 
     transformer_kwargs = {
@@ -124,30 +123,29 @@ def test_simple_inference_save_pretrained(self):
         """
         Tests a simple usecase where users could use saving utilities for LoRA through save_pretrained
         """
-        for scheduler_cls in self.scheduler_classes:
-            components, _, _ = self.get_dummy_components(scheduler_cls)
-            pipe = self.pipeline_class(**components)
-            pipe = pipe.to(torch_device)
-            pipe.set_progress_bar_config(disable=None)
-            _, _, inputs = self.get_dummy_inputs(with_generator=False)
+        components, _, _ = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        _, _, inputs = self.get_dummy_inputs(with_generator=False)
 
-            output_no_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
-            self.assertTrue(output_no_lora.shape == self.output_shape)
+        output_no_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
+        self.assertTrue(output_no_lora.shape == self.output_shape)
 
-            images_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
+        images_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
 
-            with tempfile.TemporaryDirectory() as tmpdirname:
-                pipe.save_pretrained(tmpdirname)
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            pipe.save_pretrained(tmpdirname)
 
-                pipe_from_pretrained = self.pipeline_class.from_pretrained(tmpdirname)
-                pipe_from_pretrained.to(torch_device)
+            pipe_from_pretrained = self.pipeline_class.from_pretrained(tmpdirname)
+            pipe_from_pretrained.to(torch_device)
 
-            images_lora_save_pretrained = pipe_from_pretrained(**inputs, generator=torch.manual_seed(0))[0]
+        images_lora_save_pretrained = pipe_from_pretrained(**inputs, generator=torch.manual_seed(0))[0]
 
-            self.assertTrue(
-                np.allclose(images_lora, images_lora_save_pretrained, atol=1e-3, rtol=1e-3),
-                "Loading from saved checkpoints should give same results.",
-            )
+        self.assertTrue(
+            np.allclose(images_lora, images_lora_save_pretrained, atol=1e-3, rtol=1e-3),
+            "Loading from saved checkpoints should give same results.",
+        )
 
     @parameterized.expand([("block_level", True), ("leaf_level", False)])
     @require_torch_accelerator
diff --git a/tests/lora/test_lora_layers_flux.py b/tests/lora/test_lora_layers_flux.py
@@ -55,9 +55,8 @@
 @require_peft_backend
 class FluxLoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
     pipeline_class = FluxPipeline
-    scheduler_cls = FlowMatchEulerDiscreteScheduler()
+    scheduler_cls = FlowMatchEulerDiscreteScheduler
     scheduler_kwargs = {}
-    scheduler_classes = [FlowMatchEulerDiscreteScheduler]
     transformer_kwargs = {
         "patch_size": 1,
         "in_channels": 4,
@@ -282,9 +281,8 @@ def test_simple_inference_with_text_denoiser_multi_adapter_block_lora(self):
 
 class FluxControlLoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
     pipeline_class = FluxControlPipeline
-    scheduler_cls = FlowMatchEulerDiscreteScheduler()
+    scheduler_cls = FlowMatchEulerDiscreteScheduler
     scheduler_kwargs = {}
-    scheduler_classes = [FlowMatchEulerDiscreteScheduler]
     transformer_kwargs = {
         "patch_size": 1,
         "in_channels": 8,
diff --git a/tests/lora/test_lora_layers_hunyuanvideo.py b/tests/lora/test_lora_layers_hunyuanvideo.py
@@ -51,7 +51,6 @@
 class HunyuanVideoLoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
     pipeline_class = HunyuanVideoPipeline
     scheduler_cls = FlowMatchEulerDiscreteScheduler
-    scheduler_classes = [FlowMatchEulerDiscreteScheduler]
     scheduler_kwargs = {}
 
     transformer_kwargs = {
@@ -254,6 +253,7 @@ def test_original_format_cseti(self):
         expected_slices = Expectations(
             {
                 ("cuda", 7): np.array([0.1013, 0.1924, 0.0078, 0.1021, 0.1929, 0.0078, 0.1023, 0.1919, 0.7402, 0.104, 0.4482, 0.7354, 0.0925, 0.4382, 0.7275, 0.0815]),
+                ("xpu", 3): np.array([0.1013, 0.1924, 0.0078, 0.1021, 0.1929, 0.0078, 0.1023, 0.1919, 0.7402, 0.104, 0.4482, 0.7354, 0.0925, 0.4382, 0.7275, 0.0815]),
             }
         )
         # fmt: on
diff --git a/tests/lora/test_lora_layers_ltx_video.py b/tests/lora/test_lora_layers_ltx_video.py
@@ -37,7 +37,6 @@
 class LTXVideoLoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
     pipeline_class = LTXPipeline
     scheduler_cls = FlowMatchEulerDiscreteScheduler
-    scheduler_classes = [FlowMatchEulerDiscreteScheduler]
     scheduler_kwargs = {}
 
     transformer_kwargs = {
diff --git a/tests/lora/test_lora_layers_lumina2.py b/tests/lora/test_lora_layers_lumina2.py
@@ -39,7 +39,6 @@
 class Lumina2LoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
     pipeline_class = Lumina2Pipeline
     scheduler_cls = FlowMatchEulerDiscreteScheduler
-    scheduler_classes = [FlowMatchEulerDiscreteScheduler]
     scheduler_kwargs = {}
 
     transformer_kwargs = {
@@ -141,33 +140,30 @@ def test_simple_inference_with_text_lora_save_load(self):
         strict=False,
     )
     def test_lora_fuse_nan(self):
-        for scheduler_cls in self.scheduler_classes:
-            components, text_lora_config, denoiser_lora_config = self.get_dummy_components(scheduler_cls)
-            pipe = self.pipeline_class(**components)
-            pipe = pipe.to(torch_device)
-            pipe.set_progress_bar_config(disable=None)
-            _, _, inputs = self.get_dummy_inputs(with_generator=False)
-
-            if "text_encoder" in self.pipeline_class._lora_loadable_modules:
-                pipe.text_encoder.add_adapter(text_lora_config, "adapter-1")
-                self.assertTrue(
-                    check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder"
-                )
-
-            denoiser = pipe.transformer if self.unet_kwargs is None else pipe.unet
-            denoiser.add_adapter(denoiser_lora_config, "adapter-1")
-            self.assertTrue(check_if_lora_correctly_set(denoiser), "Lora not correctly set in denoiser.")
-
-            # corrupt one LoRA weight with `inf` values
-            with torch.no_grad():
-                pipe.transformer.layers[0].attn.to_q.lora_A["adapter-1"].weight += float("inf")
-
-            # with `safe_fusing=True` we should see an Error
-            with self.assertRaises(ValueError):
-                pipe.fuse_lora(components=self.pipeline_class._lora_loadable_modules, safe_fusing=True)
-
-            # without we should not see an error, but every image will be black
-            pipe.fuse_lora(components=self.pipeline_class._lora_loadable_modules, safe_fusing=False)
-            out = pipe(**inputs)[0]
-
-            self.assertTrue(np.isnan(out).all())
+        components, text_lora_config, denoiser_lora_config = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        _, _, inputs = self.get_dummy_inputs(with_generator=False)
+
+        if "text_encoder" in self.pipeline_class._lora_loadable_modules:
+            pipe.text_encoder.add_adapter(text_lora_config, "adapter-1")
+            self.assertTrue(check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder")
+
+        denoiser = pipe.transformer if self.unet_kwargs is None else pipe.unet
+        denoiser.add_adapter(denoiser_lora_config, "adapter-1")
+        self.assertTrue(check_if_lora_correctly_set(denoiser), "Lora not correctly set in denoiser.")
+
+        # corrupt one LoRA weight with `inf` values
+        with torch.no_grad():
+            pipe.transformer.layers[0].attn.to_q.lora_A["adapter-1"].weight += float("inf")
+
+        # with `safe_fusing=True` we should see an Error
+        with self.assertRaises(ValueError):
+            pipe.fuse_lora(components=self.pipeline_class._lora_loadable_modules, safe_fusing=True)
+
+        # without we should not see an error, but every image will be black
+        pipe.fuse_lora(components=self.pipeline_class._lora_loadable_modules, safe_fusing=False)
+        out = pipe(**inputs)[0]
+
+        self.assertTrue(np.isnan(out).all())
diff --git a/tests/lora/test_lora_layers_mochi.py b/tests/lora/test_lora_layers_mochi.py
@@ -37,7 +37,6 @@
 class MochiLoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
     pipeline_class = MochiPipeline
     scheduler_cls = FlowMatchEulerDiscreteScheduler
-    scheduler_classes = [FlowMatchEulerDiscreteScheduler]
     scheduler_kwargs = {}
 
     transformer_kwargs = {
diff --git a/tests/lora/test_lora_layers_qwenimage.py b/tests/lora/test_lora_layers_qwenimage.py
@@ -37,7 +37,6 @@
 class QwenImageLoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
     pipeline_class = QwenImagePipeline
     scheduler_cls = FlowMatchEulerDiscreteScheduler
-    scheduler_classes = [FlowMatchEulerDiscreteScheduler]
     scheduler_kwargs = {}
 
     transformer_kwargs = {
diff --git a/tests/lora/test_lora_layers_sana.py b/tests/lora/test_lora_layers_sana.py
@@ -31,9 +31,8 @@
 @require_peft_backend
 class SanaLoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
     pipeline_class = SanaPipeline
-    scheduler_cls = FlowMatchEulerDiscreteScheduler(shift=7.0)
-    scheduler_kwargs = {}
-    scheduler_classes = [FlowMatchEulerDiscreteScheduler]
+    scheduler_cls = FlowMatchEulerDiscreteScheduler
+    scheduler_kwargs = {"shift": 7.0}
     transformer_kwargs = {
         "patch_size": 1,
         "in_channels": 4,
diff --git a/tests/lora/test_lora_layers_sd3.py b/tests/lora/test_lora_layers_sd3.py
@@ -55,7 +55,6 @@ class SD3LoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
     pipeline_class = StableDiffusion3Pipeline
     scheduler_cls = FlowMatchEulerDiscreteScheduler
     scheduler_kwargs = {}
-    scheduler_classes = [FlowMatchEulerDiscreteScheduler]
     transformer_kwargs = {
         "sample_size": 32,
         "patch_size": 1,
diff --git a/tests/lora/test_lora_layers_wan.py b/tests/lora/test_lora_layers_wan.py
@@ -42,7 +42,6 @@
 class WanLoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
     pipeline_class = WanPipeline
     scheduler_cls = FlowMatchEulerDiscreteScheduler
-    scheduler_classes = [FlowMatchEulerDiscreteScheduler]
     scheduler_kwargs = {}
 
     transformer_kwargs = {
diff --git a/tests/lora/test_lora_layers_wanvace.py b/tests/lora/test_lora_layers_wanvace.py
@@ -50,7 +50,6 @@
 class WanVACELoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
     pipeline_class = WanVACEPipeline
     scheduler_cls = FlowMatchEulerDiscreteScheduler
-    scheduler_classes = [FlowMatchEulerDiscreteScheduler]
     scheduler_kwargs = {}
 
     transformer_kwargs = {
@@ -165,9 +164,8 @@ def test_layerwise_casting_inference_denoiser(self):
 
     @require_peft_version_greater("0.13.2")
     def test_lora_exclude_modules_wanvace(self):
-        scheduler_cls = self.scheduler_classes[0]
         exclude_module_name = "vace_blocks.0.proj_out"
-        components, text_lora_config, denoiser_lora_config = self.get_dummy_components(scheduler_cls)
+        components, text_lora_config, denoiser_lora_config = self.get_dummy_components()
         pipe = self.pipeline_class(**components).to(torch_device)
         _, _, inputs = self.get_dummy_inputs(with_generator=False)
 
diff --git a/tests/lora/utils.py b/tests/lora/utils.py
diff --git a/tests/pipelines/easyanimate/test_easyanimate.py b/tests/pipelines/easyanimate/test_easyanimate.py
diff --git a/tests/pipelines/hidream_image/test_pipeline_hidream.py b/tests/pipelines/hidream_image/test_pipeline_hidream.py
diff --git a/tests/pipelines/marigold/test_marigold_depth.py b/tests/pipelines/marigold/test_marigold_depth.py
diff --git a/tests/pipelines/omnigen/test_pipeline_omnigen.py b/tests/pipelines/omnigen/test_pipeline_omnigen.py
diff --git a/tests/pipelines/qwenimage/test_qwenimage_controlnet.py b/tests/pipelines/qwenimage/test_qwenimage_controlnet.py

Original file line number	Diff line number	Diff line change
`@@ -194,7 +194,7 @@ def from_pretrained(cls, pretrained_model_or_path: Optional[Union[str, os.PathLi`
`194`	`194`
`195`	`195`	`has_remote_code = "auto_map" in config and cls.__name__ in config["auto_map"]`
`196`	`196`	`trust_remote_code = resolve_trust_remote_code(trust_remote_code, pretrained_model_or_path, has_remote_code)`
`197`		`- if not (has_remote_code and trust_remote_code):`
	`197`	`+ if not has_remote_code and trust_remote_code:`
`198`	`198`	`raise ValueError(`
`199`	`199`	"Selected model repository does not happear to have any custom code or does not have a valid `config.json` file."
`200`	`200`	`)`
Original file line number	Diff line number	Diff line change
`@@ -323,7 +323,7 @@ def from_pretrained(`
`323`	`323`	`trust_remote_code = resolve_trust_remote_code(`
`324`	`324`	`trust_remote_code, pretrained_model_name_or_path, has_remote_code`
`325`	`325`	`)`
`326`		`- if not (has_remote_code and trust_remote_code):`
	`326`	`+ if not has_remote_code and trust_remote_code:`
`327`	`327`	`raise ValueError(`
`328`	`328`	"Selected model repository does not happear to have any custom code or does not have a valid `config.json` file."
`329`	`329`	`)`
Original file line number	Diff line number	Diff line change
`@@ -51,7 +51,6 @@`
`51`	`51`	`class HunyuanVideoLoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):`
`52`	`52`	`pipeline_class = HunyuanVideoPipeline`
`53`	`53`	`scheduler_cls = FlowMatchEulerDiscreteScheduler`
`54`		`- scheduler_classes = [FlowMatchEulerDiscreteScheduler]`
`55`	`54`	`scheduler_kwargs = {}`
`56`	`55`
`57`	`56`	`transformer_kwargs = {`
`@@ -254,6 +253,7 @@ def test_original_format_cseti(self):`
`254`	`253`	`expected_slices = Expectations(`
`255`	`254`	`{`
`256`	`255`	`("cuda", 7): np.array([0.1013, 0.1924, 0.0078, 0.1021, 0.1929, 0.0078, 0.1023, 0.1919, 0.7402, 0.104, 0.4482, 0.7354, 0.0925, 0.4382, 0.7275, 0.0815]),`
	`256`	`+ ("xpu", 3): np.array([0.1013, 0.1924, 0.0078, 0.1021, 0.1929, 0.0078, 0.1023, 0.1919, 0.7402, 0.104, 0.4482, 0.7354, 0.0925, 0.4382, 0.7275, 0.0815]),`
`257`	`257`	`}`
`258`	`258`	`)`
`259`	`259`	`# fmt: on`