How to make it forget everything during for loop? #519

Sorry I am new to this.
Say I have a book full of independent articles, and I am trying to run alpaca in a loop, but for every run, I want it to forget everything and start new. (Basically the opposite of chat, make model remember zero previous context) To do this, based on my limited understanding, I need to set last_n_tokens_size to 0 (Is this even correct?)

llm = Llama(model_path="some_model.bin",n_ctx=1024,n_batch=1024, last_n_tokens_size=0, n_gpu_layers=2000000)
for article in book:
    output=evaluate(llm,article)

to make predictions.

Unfortunately the code above will cause an error:

  File "/***/lib/python3.10/site-packages/llama_cpp/llama.py", line 1328, in __call__
    return self.create_completion(
  File "/***/lib/python3.10/site-packages/llama_cpp/llama.py", line 1280, in create_completion
    completion: Completion = next(completion_or_chunks)  # type: ignore
  File "/***/lib/python3.10/site-packages/llama_cpp/llama.py", line 872, in _create_completion
    for token in self.generate(
  File "/***/lib/python3.10/site-packages/llama_cpp/llama.py", line 695, in generate
    token = self.sample(
  File "/***/lib/python3.10/site-packages/llama_cpp/llama.py", line 620, in sample
    last_n_tokens_data=(llama_cpp.llama_token * self.last_n_tokens_size)(
IndexError: invalid index

Interestingly, if I change last_n_tokens_size to 1, it will run smoothly.

My question is:

1. Is last_n_tokens_size=0 even the correct way to make it forget everything before? If not, how?
2. Is it ok to put llm = Llama(model=***) outside the for loop if I want each run not interfere each other's result, or do I need to put it inside the for loop?
3. If anyone also making predictions that are independent for each run, could you share some examples?

I apologize if the questions are silly.
Thanks in advance!