wan2.2 i2v FirstBlockCache fix (#12013)

* enable caching for WanImageToVideoPipeline

* ruff format

src/diffusers/pipelines/wan/pipeline_wan_i2v.py

@@ -750,25 +750,27 @@ class WanImageToVideoPipeline(DiffusionPipeline, WanLoraLoaderMixin):
                     latent_model_input = torch.cat([latents, condition], dim=1).to(transformer_dtype)
                     timestep = t.expand(latents.shape[0])
 
-                noise_pred = current_model(
-                    hidden_states=latent_model_input,
-                    timestep=timestep,
-                    encoder_hidden_states=prompt_embeds,
-                    encoder_hidden_states_image=image_embeds,
-                    attention_kwargs=attention_kwargs,
-                    return_dict=False,
-                )[0]
-
-                if self.do_classifier_free_guidance:
-                    noise_uncond = current_model(
+                with current_model.cache_context("cond"):
+                    noise_pred = current_model(
                         hidden_states=latent_model_input,
                         timestep=timestep,
-                        encoder_hidden_states=negative_prompt_embeds,
+                        encoder_hidden_states=prompt_embeds,
                         encoder_hidden_states_image=image_embeds,
                         attention_kwargs=attention_kwargs,
                         return_dict=False,
                     )[0]
-                    noise_pred = noise_uncond + current_guidance_scale * (noise_pred - noise_uncond)
+
+                if self.do_classifier_free_guidance:
+                    with current_model.cache_context("uncond"):
+                        noise_uncond = current_model(
+                            hidden_states=latent_model_input,
+                            timestep=timestep,
+                            encoder_hidden_states=negative_prompt_embeds,
+                            encoder_hidden_states_image=image_embeds,
+                            attention_kwargs=attention_kwargs,
+                            return_dict=False,
+                        )[0]
+                        noise_pred = noise_uncond + current_guidance_scale * (noise_pred - noise_uncond)
 
                 # compute the previous noisy sample x_t -> x_t-1
                 latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]