Update app.py

app.py

@@ -53,7 +53,8 @@ def clear_memory():
 def load_t2v_pipeline():
     """Load and optimize the T2V pipeline."""
     if t2v_pipe_cache[0] is None:
-        vae = AutoencoderKLWan.from_pretrained(T2V_MODEL_ID, subfolder="vae", torch_dtype=torch.float32)
+        vae = AutoencoderKLWan.from_pretrained("Wan-AI/Wan2.2-T2V-A14B-Diffusers",subfolder="vae",torch_dtype=torch.bfloat16).to('cuda')  # also move VAE to GPU immediately
+
         t2v_pipe_cache[0] = WanPipeline.from_pretrained(T2V_MODEL_ID,
             transformer=WanTransformer3DModel.from_pretrained('linoyts/Wan2.2-T2V-A14B-Diffusers-BF16',
                 subfolder='transformer',