Update main.py

main.py

@@ -54,7 +54,7 @@ async def generate_text(request: GenerationRequest):
         )
 
         # Generate response
-        generate_kwargs = {
+        params = {
             "temperature": max(request.temperature, 0.01),  # Ensure temperature isn't too low
             "max_new_tokens": 1048,
             "top_p": request.top_p,
@@ -63,13 +63,14 @@ async def generate_text(request: GenerationRequest):
             "seed": 42
         }
 
-        # Generate the response
-        stream = client.text_generation(formatted_prompt, **generate_kwargs)
-        output = ""
-        for response in stream:
-            output += response.token.text
+        # Generate the response - handling the response as a single string
+        response = client.text_generation(
+            formatted_prompt,
+            **params
+        )
 
-        return {"response": output}
+        # The response is now directly a string
+        return {"response": response}
 
     except Exception as e:
         raise HTTPException(status_code=500, detail=str(e))