Update app_v3.py

app_v3.py

@@ -47,6 +47,21 @@ 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
+
+# 2. Memory Efficient Attention (xFormers): Reduces memory usage and improves speed
+# Requires xformers library installation. Beneficial even with high VRAM.
+try:
+    pipe.enable_xformers_memory_efficient_attention()
+except Exception as e:
+    print(f"XFormers not available, skipping memory efficient attention: {e}")
+
+# 3. Attention Slicing: Recommended for FLUX models and high-resolution images,
+# even with ample VRAM, as it can sometimes help with very large tensors.
+pipe.enable_attention_slicing()
+
 @spaces.GPU(duration=10)
 @torch.no_grad()
 def generate_image(prompt, scale, steps, control_image, controlnet_conditioning_scale, guidance_scale, seed, guidance_end):
@@ -228,4 +243,4 @@ with gr.Blocks(title="FLUX Turbo Upscaler", fill_height=True) as demo:
         outputs=[prompt]
     )
 
-demo.launch(mcp_server=True, show_error=True)
+demo.launch(show_error=True)