app.py

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
from collections import defaultdict
import gradio as gr
from optimum.onnxruntime import ORTModelForCausalLM


import itertools
import regex as re
import logging



user_token = "<User>"
eos_token = "<EOS>"
bos_token = "<BOS>"
bot_token = "<Assistant>"


logger = logging.getLogger()
handler = logging.StreamHandler()
formatter = logging.Formatter(
        '%(asctime)s %(name)-12s %(levelname)-8s %(message)s')
handler.setFormatter(formatter)
logger.addHandler(handler)
logger.setLevel(logging.INFO)

max_context_length = 750

def format(history):
    prompt = bos_token

    for idx, txt in enumerate(history):
        if idx % 2 == 0:
            prompt += f"{user_token}{txt}{eos_token}"
        else:
            prompt += f"{bot_token}{txt}"
    prompt += bot_token
    logger.info(prompt)
    return prompt

def remove_spaces_between_chinese(text):
    rex = r"(?<![a-zA-Z]{2})(?<=[a-zA-Z]{1})[ ]+(?=[a-zA-Z] |.$)|(?<=\p{Han}) +"
    return re.sub(rex, "", text, 0, re.MULTILINE | re.UNICODE)

def gradio(model, tokenizer):
    def response(
        user_input,
        chat_history,
        top_k,
        top_p,
        temperature,
        repetition_penalty,
        no_repeat_ngram_size,
    ):
        history = list(itertools.chain(*chat_history))
        history.append(user_input)

        prompt = format(history)

        input_ids = tokenizer.encode(
            prompt,
            return_tensors="pt",
            add_special_tokens=False,
        )[:, -max_context_length:]

        prompt_length = input_ids.shape[1]

        beam_output = model.generate(
            input_ids,
            pad_token_id=tokenizer.pad_token_id,
            max_new_tokens=250,
            num_beams=1, # with cpu
            top_k=top_k,
            top_p=top_p,
            no_repeat_ngram_size=no_repeat_ngram_size,
            temperature=temperature,
            repetition_penalty=repetition_penalty,
            early_stopping=True,
            do_sample=True
        )
        output = beam_output[0][prompt_length:]

        generated = remove_spaces_between_chinese(tokenizer.decode(output, skip_special_tokens=True, clean_up_tokenization_spaces=True))
        
        return generated

    bot = gr.Chatbot(show_copy_button=True, show_share_button=True, height="2000")

    with gr.Blocks() as demo:
        gr.Markdown("GPT2 chatbot | Powered by nlp-greyfoss")

        with gr.Accordion("Parameters in generation", open=False):
            with gr.Row():
                top_k = gr.Slider(
                    2.0,
                    100.0,
                    label="top_k",
                    step=1,
                    value=50,
                    info="Limit the number of candidate tokens considered during decoding.",
                )
                top_p = gr.Slider(
                    0.1,
                    1.0,
                    label="top_p",
                    value=0.9,
                    info="Control the diversity of the output by selecting tokens with cumulative probabilities up to the Top-P threshold.",
                )
                temperature = gr.Slider(
                    0.1,
                    2.0,
                    label="temperature",
                    value=0.9,
                    info="Control the randomness of the generated text. A higher temperature results in more diverse and unpredictable outputs, while a lower temperature produces more conservative and coherent text.",
                )
                repetition_penalty = gr.Slider(
                    0.1,
                    2.0,
                    label="repetition_penalty",
                    value=1.2,
                    info="Discourage the model from generating repetitive tokens in a sequence.",
                )
                no_repeat_ngram_size = gr.Slider(
                    0,
                    100,
                    label="no_repeat_ngram_size",
                    step=1,
                    value=5,
                    info="Prevent the model from generating sequences of n consecutive tokens that have already been generated in the context. ",
                )

        gr.ChatInterface(
            response,
            chatbot=bot,
            additional_inputs=[
                top_k,
                top_p,
                temperature,
                repetition_penalty,
                no_repeat_ngram_size,
            ],
            retry_btn = "🔄 Regenerate",
            undo_btn = "↩️ Remove last turn",
            clear_btn = "➕ New conversation",
            examples=[
                ["写一篇介绍人工智能的文章。", 30, 0.9, 0.95, 1.2, 5],
                ["给我讲一个笑话。", 50, 0.8, 0.9, 1.2, 6],
                ["Can you describe spring in English?", 50, 0.9, 1.0, 1, 5]
            ]
        )

    demo.queue().launch()




tokenizer = AutoTokenizer.from_pretrained("greyfoss/gpt2-chatbot-chinese")

model = ORTModelForCausalLM.from_pretrained("greyfoss/gpt2-chatbot-chinese", export=True)

gradio(model, tokenizer)