Disable VAE tile parallelism in workers + post-inference logging

app.py

@@ -278,11 +278,16 @@ def _generate_impl(input_image, prompt, negative_prompt, guidance_scale, seed, s
                 generate_image=True,
                 ref_latents=ref_latents,
                 guidance_scale=ff_args.guidance_scale,
+                # tiled VAE decode uses threaded CUDA work and can deadlock
+                # in forked ZeroGPU workers; 512x512x1 frame fits VRAM easily
+                enable_vae_tile_parallelism=False,
             ).videos[0]
+        print("[frameforge] pipe() returned", flush=True)
 
         image = torch.from_numpy(np.array(sample.squeeze(1)))
         out_path = tempfile.NamedTemporaryFile(delete=False, suffix=".png").name
         ip.save_image(image, out_path)
+        print(f"[frameforge] saved {out_path}", flush=True)
         return out_path