Googleが公開している「gemma-2-9b-it」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」で同じことをしました。
touch-sp.hatenablog.com
touch-sp.hatenablog.com
今回は「gemma-2-9b-it」です。

小規模でかつ日本語特化モデルでないにもかかわらず日本語性能は高い印象です。

モデルの量子化

今回は量子化を行いませんでした。

Gradioで実行

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

# model was downloaded from https://huggingface.co/google/gemma-2-9b-it
model = AutoModelForCausalLM.from_pretrained(
    "gemma-2-9b-it",
    device_map="auto",
    torch_dtype="auto"
)
tokenizer = AutoTokenizer.from_pretrained("gemma-2-9b-it")
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,
):
    history_openai_format = []
    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)
    
    generation_kwargs = dict(
        inputs=input_ids,
        streamer=streamer,
        max_new_tokens=max_tokens,
        do_sample=True,
        temperature=temperature,
        top_p=top_p
    )

    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("# gemma-2-9b-it")
        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=1024,
                    label="Max tokens",
                    visible=True,
                    render=False,
                ),
                gr.Slider(
                    minimum=0,
                    maximum=1,
                    step=0.1,
                    value=0.3,
                    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,
                ),
            ],
        )
    demo.launch(share=False)

if __name__ == "__main__":
    run()