app.py

import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# モデルとトークナイザーの読み込み
model_name = "EleutherAI/Pythia-1b"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, ignore_mismatched_sizes=True)

# 応答を生成する関数
def respond(message, history, max_tokens, temperature, top_p):
    # 入力履歴と新しいメッセージを連結
    if history is None:
        history = []

    input_text = ""
    for user_message, bot_response in history:
        input_text += f"User: {user_message}\nAssistant: {bot_response}\n"
    input_text += f"User: {message}\nAssistant:"

    # トークナイズ
    inputs = tokenizer(input_text, return_tensors="pt")

    # モデルによる応答生成
    with torch.no_grad():
        outputs = model.generate(
            inputs.input_ids,
            max_length=inputs.input_ids.shape[1] + max_tokens,
            do_sample=True,
            top_p=top_p,
            temperature=temperature,
        )

    # 応答をデコード
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    # 最後のユーザー入力以降の応答部分を抽出
    response = response.split("Assistant:")[-1].strip()

    # 応答と履歴を更新
    history.append((message, response))
    return response, history

# Gradioインターフェースの設定
with gr.Blocks() as demo:
    gr.Markdown("## AIチャット")
    chatbot = gr.Chatbot()
    msg = gr.Textbox(label="あなたのメッセージ", placeholder="ここにメッセージを入力...")
    max_tokens = gr.Slider(1, 2048, value=512, step=1, label="新規トークン最大")
    temperature = gr.Slider(0.1, 4.0, value=0.7, step=0.1, label="温度")
    top_p = gr.Slider(0.1, 1.0, value=0.95, step=0.05, label="Top-p (核サンプリング)")
    send_button = gr.Button("送信")
    clear = gr.Button("クリア")

    def clear_history():
        return [], []

    send_button.click(respond, inputs=[msg, chatbot, max_tokens, temperature, top_p], outputs=[chatbot, chatbot])
    clear.click(clear_history, outputs=[chatbot])

demo.launch()