サイバーエージェントが公開している「CyberAgentLM3-22B-Chat」をGradioを使ってローカルで使用する - パソコン関連もろもろ

PC環境

Ubuntu 24.04 on WSL2 (Windows 11)
CUDA 12.1
Python 3.12

Python環境構築

pip install torch==2.3.1+cu121 --index-url https://download.pytorch.org/whl/cu121
pip install transformers accelerate bitsandbytes gradio

モデルの量子化

from transformers import AutoModelForCausalLM, BitsAndBytesConfig

# model was downloaded from https://huggingface.co/cyberagent/calm3-22b-chat
quantization_config = BitsAndBytesConfig(load_in_4bit=True)
model = AutoModelForCausalLM.from_pretrained(
    "calm3-22b-chat",
    quantization_config=quantization_config
)
model.save_pretrained("calm3-22b-chat-4bit")

Gradioで実行

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

system_prompt_text = "あなたは親切なAIアシスタントです。"
init = {
    "role": "system",
    "content": system_prompt_text,
}

model = AutoModelForCausalLM.from_pretrained(
    "calm3-22b-chat-4bit",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("calm3-22b-chat")
streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)

defcall_llm(
    message: str,
    history: list[dict],
    max_tokens: int,
    temperature: float,
    top_p: float,
):
    history_openai_format = []
    iflen(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)
    
    generation_kwargs = dict(
        inputs=input_ids,
        streamer=streamer,
        max_new_tokens=max_tokens,
        temperature=temperature
    )

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

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

chatbot = gr.Chatbot(
    elem_id="chatbot",
    scale=1,
    show_copy_button=True,
    height="70%",
    layout="panel",
)

defrun():
    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("# CALM3-22B-Chat")
        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=1.0,
                    label="Top-p",
                    visible=True,
                    render=False,
                ),
            ],
            analytics_enabled=False,
        )
    demo.launch(share=False)

if __name__ == "__main__":
    run()

実際の画面

VRAM使用量

ランキング参加中

プログラミング