# At the top level of your script, after initializing the tokenizer
tokenizer = AutoTokenizer.from_pretrained(tokenizer_id)
if tokenizer.pad_token == None:
    tokenizer.pad_token = tokenizer.eos_token
    tokenizer.pad_token_id = tokenizer.eos_token_id

@spaces.GPU
def generate(
    message: str,
    chat_history: list[tuple[str, str]],
    max_new_tokens: int = 1024,
    temperature: float = 0.6,
    top_p: float = 0.9,
    top_k: int = 50,
    repetition_penalty: float = 1.4,
) -> Iterator[str]:
    global tokenizer, model  # Add this line to access global variables
    
    input_ids = tokenizer([message], return_tensors="pt").input_ids
    # ... rest of the function ...

# The rest of your code remains the same