Update app_v3.py

app_v3.py

@@ -47,9 +47,12 @@ pipe = FluxControlNetPipeline.from_pretrained(
 )
 pipe.to("cuda")
 
-pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
 # For FLUX models, compiling VAE decode can also be beneficial if needed, though UNet is primary.
 # pipe.vae.decode = torch.compile(pipe.vae.decode, mode="reduce-overhead", fullgraph=True) # Uncomment if VAE compile helps
+try:
+    pipe.vae.decode = torch.compile(pipe.vae.decode, mode="reduce-overhead", fullgraph=True)
+except Exception as e:
+    print(f"Compile failed: {e}")
 
 # 2. Memory Efficient Attention (xFormers): Reduces memory usage and improves speed
 # Requires xformers library installation. Beneficial even with high VRAM.