Update main.py

main.py

@@ -54,6 +54,7 @@ def generate_text(request: Request, input_data: InputData):
     repetition_penalty = input_data.repetition_penalty
     input_text = generate_input_text(input_data)
     print(f"{datetime.now()} - Input Text: {input_text}")
+    history = []
     generated_response = generate(input_text, history, temperature, max_new_tokens, top_p, repetition_penalty)
     print(f"{datetime.now()} - Response Text: {generated_response}")
     return {"response": generated_response}