torch compile

app.py

@@ -86,6 +86,8 @@ if gr.NO_RELOAD:
     
     # experimental speedup?
     # pipe.compile()
+    pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
+    print("Model Compiled!")
     
     # torch.cuda.empty_cache()
     # gc.collect()
@@ -350,7 +352,6 @@ def process_image(
 ):
     print("processing image")
     start = time.time()
-    torch.cuda.empty_cache()
     preprocessor.load("NormalBae")
     # preprocessor.load("Canny") #20 steps, 9 guidance, 512, 512