Upload folder using huggingface_hub

app.py

@@ -547,12 +547,14 @@ def generate(
 
     pred_latent = torch.cat(pred_latent_lst, dim=2)
 
-    # Clean up caches and free VRAM before VAE decode
+    # Free VRAM before VAE decode: move transformer and text encoder to CPU
     transformer.clear_cache("pos")
     streaming_vae.clear_cache()
+    transformer.to("cpu")
+    text_encoder.to("cpu")
     torch.cuda.empty_cache()
 
-    # Decode video - move pred_latent to CPU, then decode frame-by-frame
+    # Decode video
     pred_latent_cpu = pred_latent.cpu()
     del pred_latent
     torch.cuda.empty_cache()