import torch, transformers  

def generate_response():
    model = transformers.AutoModelForCausalLM.from_pretrained("benchang1110/Taiwan-tinyllama-v1.0-chat", torch_dtype=torch.bfloat16, device_map=device,attn_implementation="flash_attention_2")
    tokenizer = transformers.AutoTokenizer.from_pretrained("benchang1110/Taiwan-tinyllama-v1.0-chat")
    streamer = transformers.TextStreamer(tokenizer,skip_prompt=True)
    while(1):
        prompt = input('USER:')
        if prompt == "exit":
            break
        print("Assistant: ")
        message = [
           {'content': prompt, 'role': 'user'},
        ]
        untokenized_chat = tokenizer.apply_chat_template(message,tokenize=False,add_generation_prompt=False)
        inputs = tokenizer.encode_plus(untokenized_chat, add_special_tokens=True, return_tensors="pt",return_attention_mask=True).to(device)
        outputs = model.generate(inputs["input_ids"],attention_mask=inputs['attention_mask'],streamer=streamer,use_cache=True,max_new_tokens=512,do_sample=True,temperature=0.1,repetition_penalty=1.2)


if __name__ == '__main__':
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
    generate_response()        