Make gradio_chat a non-generator GPU function for ZeroGPU detection

- Add @spaces.GPU(duration=120) decorator to gradio_chat
- Change from generator (yield) to regular function (return)
- Use generate_text instead of zerogpu_generate_stream
- This ensures ZeroGPU detects GPU usage at startup

app.py

@@ -204,6 +204,7 @@ async def serverless_generate(
 # --- Gradio Interface ---
 
 
+@spaces.GPU(duration=120)
 def gradio_chat(
     message: str,
     history: list[list[str]],
@@ -211,11 +212,10 @@ def gradio_chat(
     temperature: float,
     max_tokens: int,
 ):
-    """Gradio chat interface handler."""
+    """Gradio chat interface handler - GPU decorated for ZeroGPU."""
     # Validate model_id
     if not model_id:
-        yield "Please select a model first."
-        return
+        return "Please select a model first."
 
     # Build messages from history
     messages = []
@@ -229,24 +229,21 @@ def gradio_chat(
     try:
         prompt = apply_chat_template(model_id, messages)
     except Exception as e:
-        yield f"Error loading model: {str(e)}"
-        return
+        return f"Error loading model: {str(e)}"
 
-    # Generate response (streaming)
-    response = ""
+    # Generate response (non-streaming for simplicity with ZeroGPU)
     try:
-        for token in zerogpu_generate_stream(
+        response = generate_text(
             model_id=model_id,
             prompt=prompt,
             max_new_tokens=max_tokens,
             temperature=temperature,
             top_p=0.95,
             stop_sequences=None,
-        ):
-            response += token
-            yield response
+        )
+        return response
     except Exception as e:
-        yield f"Error generating response: {str(e)}"
+        return f"Error generating response: {str(e)}"
 
 
 # --- FastAPI app for OpenAI-compatible routes ---