Simplify model loading and remove unnecessary caching

app.py

@@ -7,7 +7,6 @@ Any new model should implement the following functions:
 
 import os
 import glob
-import functools
 import spaces
 import gradio as gr
 from huggingface_hub import get_token
@@ -19,9 +18,8 @@ from utils import (
 )
 from flagging import HuggingFaceDatasetSaver
 
-import install_private_repos
-from seavision import load_model, AHOY
-from seavision.utils.general import LOGGER
+import install_private_repos  # noqa: F401
+from seavision import load_model
 
 
 TITLE = """
@@ -48,14 +46,12 @@ h1 {
 }
 """
 
-@functools.lru_cache(maxsize=1)
-def get_model():
-    return load_model("ahoy-RGB-b2")
+model = load_model("ahoy-RGB-b2")
+model.to("cuda")
 
 @spaces.GPU
 def inference(image):
     """Run inference on image and return annotated image."""
-    model = get_model()
     results = model(image)
     return results.draw(image, diameter=4)
 
@@ -95,6 +91,7 @@ with gr.Blocks(theme=theme, css=css, title="SEA.AI Vision Demo") as demo:
         outputs=img_output,
         fn=inference,
         cache_examples=True,
+        cache_mode="lazy",
     )
 
     # add components to clear when clear button is clicked