avoid memory leak

app.py

@@ -121,13 +121,17 @@ def caption_frame(frame, size, model_file, clip_file, interval_ms, sys_prompt, u
             ]}
         ]
         # re-init handler
-        llm.chat_handler.__init__(clip_model_path=clip_file, verbose=False)
+        llm.chat_handler = SmolVLM2ChatHandler(clip_model_path=clip_file, verbose=False)
         resp = llm.create_chat_completion(
             messages=messages,
             max_tokens=128,
             temperature=0.1,
             stop=["<end_of_utterance>"]
         )
+
+    import gc
+    gc.collect()
+
     return resp.get('choices', [{}])[0].get('message', {}).get('content', '').strip()
 
 # Gradio UI