Allow for vae tiling onle without low vram on wan models with a model kwarg

toolkit/models/wan21/wan21.py

@@ -520,6 +520,13 @@ class Wan21(BaseModel):
 
         return pipeline
 
+    @property
+    def use_vae_tiling(self):
+        # tile the vae decode when sampling if in low vram or explicitly enabled
+        return self.model_config.low_vram or self.model_config.model_kwargs.get(
+            "vae_tiling", False
+        )
+
     def generate_single_image(
         self,
         pipeline: WanPipeline,
@@ -532,6 +539,11 @@ class Wan21(BaseModel):
         # reactivate progress bar since this is slooooow
         pipeline.set_progress_bar_config(disable=False)
         pipeline = pipeline.to(self.device_torch)
+
+        if self.use_vae_tiling:
+            # set vae to tile decode
+            pipeline.vae.enable_tiling()
+
         # todo, figure out how to do video
         output = pipeline(
             prompt_embeds=conditional_embeds.text_embeds.to(
@@ -550,6 +562,10 @@ class Wan21(BaseModel):
             **extra
         )[0]
 
+        if self.use_vae_tiling:
+            # restore no tiling
+            pipeline.vae.disable_tiling()
+
         # shape = [1, frames, channels, height, width]
         batch_item = output[0]  # list of pil images
         if gen_config.num_frames > 1: