rinnaが公開している「llama-3-youko-8b-instruct」をGradioを使ってローカルで使用する

はじめに

前回「CyberAgentLM3-22B-Chat」や「Llama-3-ELYZA-JP-8B」や「gemma-2-9b-it」で同じことをしました。
touch-sp.hatenablog.com
touch-sp.hatenablog.com
touch-sp.hatenablog.com
今回は「llama-3-youko-8b-instruct」です。

モデルの量子化

今回は量子化を行いませんでした。

float16でモデルを読み込みました。

Gradioで実行

import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
from threading import Thread
import torch

system_prompt_text = "あなたは誠実で優秀なアシスタントです。どうか、簡潔かつ正直に答えてください。"
init = {
    "role": "system",
    "content": system_prompt_text,
}

# model was downloaded from https://huggingface.co/rinna/llama-3-youko-8b-instruct
model = AutoModelForCausalLM.from_pretrained(
    "llama-3-youko-8b-instruct",
    device_map="auto",
    torch_dtype=torch.float16
)
tokenizer = AutoTokenizer.from_pretrained("llama-3-youko-8b-instruct")
streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)

def call_llm(
    message: str,
    history: list[dict],
    max_tokens: int,
    temperature: float,
    top_p: float,
    repetition_penalty: float
):
    history_openai_format = []
    if len(history) == 0:
        history_openai_format.append(init)
        history_openai_format.append({"role": "user", "content": message})
    else:
        history_openai_format.append(init)
        for human, assistant in history:
            history_openai_format.append({"role": "user", "content": human})
            history_openai_format.append({"role": "assistant", "content": assistant})
        history_openai_format.append({"role": "user", "content": message})

    input_ids = tokenizer.apply_chat_template(
        history_openai_format,
        add_generation_prompt=True,
        return_tensors="pt"
    ).to(model.device)

    terminators = [
    tokenizer.convert_tokens_to_ids("<|end_of_text|>"),
    tokenizer.convert_tokens_to_ids("<|eot_id|>")
    ]   
    generation_kwargs = dict(
        inputs=input_ids,
        streamer=streamer,
        eos_token_id=terminators,
        pad_token_id=tokenizer.eos_token_id,
        max_new_tokens=max_tokens,
        temperature=temperature,
        top_p=top_p,
        repetition_penalty=repetition_penalty
    )

    thread = Thread(target=model.generate, kwargs=generation_kwargs)
    thread.start()

    generated_text = ""
    for new_text in streamer:
        generated_text += new_text
        yield generated_text

def run():
    chatbot = gr.Chatbot(
        elem_id="chatbot",
        scale=1,
        show_copy_button=True,
        height="70%",
        layout="panel",
    )
    with gr.Blocks(fill_height=True) as demo:
        gr.Markdown("# llama-3-youko-8b-instruct")
        gr.ChatInterface(
            fn=call_llm,
            stop_btn="Stop Generation",
            cache_examples=False,
            multimodal=False,
            chatbot=chatbot,
            additional_inputs_accordion=gr.Accordion(
                label="Parameters", open=False, render=False
            ),
            additional_inputs=[
                gr.Slider(
                    minimum=1,
                    maximum=4096,
                    step=1,
                    value=512,
                    label="Max tokens",
                    visible=True,
                    render=False,
                ),
                gr.Slider(
                    minimum=0,
                    maximum=1,
                    step=0.1,
                    value=0.6,
                    label="Temperature",
                    visible=True,
                    render=False,
                ),
                gr.Slider(
                    minimum=0,
                    maximum=1,
                    step=0.1,
                    value=0.9,
                    label="Top-p",
                    visible=True,
                    render=False,
                ),
                gr.Slider(
                    minimum=0,
                    maximum=2,
                    step=0.1,
                    value=1.1,
                    label="repetition_penalty",
                    visible=True,
                    render=False,
                ),
            ],
        )
    demo.launch(share=False)

if __name__ == "__main__":
    run()