change compile options

app.py

@@ -20,19 +20,26 @@ model = AutoModelForObjectDetection.from_pretrained(
     disable_custom_kernels=True,
     torch_dtype=torch.float16,
 ).to(device)
-model_compiled = torch.compile(model, mode="reduce-overhead")
+model_compiled = torch.compile(
+    model,
+    options={"triton.cudagraphs": True, "max_autotune": True},
+)
 
 url = "http://images.cocodataset.org/val2017/000000039769.jpg"
 image = Image.open(requests.get(url, stream=True).raw)
 inputs = processor(images=image, return_tensors="pt").to(device).to(torch.float16)
 
-print("Compiling model...")
-start_time = time.time()
-with torch.no_grad():
-    for _ in range(10):
-        outputs = model_compiled(**inputs)
-        _ = outputs[0].cpu()
-print(f"Model compiled in {time.time() - start_time:.2f} seconds.")
+
+@spaces.GPU
+def init_compiled_model():
+    print("Compiling model...")
+    start_time = time.time()
+    with torch.no_grad():
+        for _ in range(10):
+            outputs = model_compiled(**inputs)
+            _ = outputs[0].cpu()
+    print(f"Model compiled in {time.time() - start_time:.2f} seconds.")
+
 
 css = """
 .feedback textarea {font-size: 24px !important}
@@ -175,4 +182,5 @@ with gr.Blocks(theme=gr.themes.Soft(), css=css) as demo:
     )
 
 if __name__ == "__main__":
+    init_compiled_model()
     demo.launch(show_error=True)