update

src/diffusers/pipelines/wan/pipeline_wan_vace.py

@@ -171,7 +171,7 @@ class WanVACEPipeline(DiffusionPipeline, WanLoraLoaderMixin):
             boundary_timestep. If `None`, only the available transformer is used for the entire denoising process.
     """
 
-    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    model_cpu_offload_seq = "text_encoder->transformer->transformer_2->vae"
     _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
     _optional_components = ["transformer", "transformer_2"]
 

tests/pipelines/wan/test_wan_vace.py

@@ -253,9 +253,15 @@ class WanVACEPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         assert video.shape == (17, 3, 16, 16)
 
     def test_save_load_optional_components(self, expected_max_difference=1e-4):
-        optional_component = ["transformer", "image_encoder", "image_processor"]
+        optional_component = ["transformer"]
 
         components = self.get_dummy_components()
+        components["transformer_2"] = components["transformer"]
+        # FlowMatchEulerDiscreteScheduler doesn't support running low noise only scheduler
+        # because starting timestep t == 1000 == boundary_timestep
+        components["scheduler"] = UniPCMultistepScheduler(
+            prediction_type="flow_prediction", use_flow_sigmas=True, flow_shift=3.0
+        )
         for component in optional_component:
             components[component] = None