app.py

#!/usr/bin/env python

import os
from threading import Thread
from queue import Queue, Empty
from typing import Iterator

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

DESCRIPTION = "# Sakalti/anchobi-4b"
DESCRIPTION += "\n<p>現在の環境に合わせて最適化されています。</p>"

MAX_MAX_NEW_TOKENS = 2048
DEFAULT_MAX_NEW_TOKENS = 1024
MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "32768"))

model_id = "Sakalti/anchobi-4b"
if torch.cuda.is_available():
    model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
else:
    model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float32)
model.eval()

tokenizer = AutoTokenizer.from_pretrained(model_id)


def apply_chat_template(conversation: list[dict[str, str]]) -> str:
    prompt = "\n".join([f"{c['role']}: {c['content']}" for c in conversation])
    prompt = f"{prompt}\nASSISTANT: "
    return prompt


@torch.inference_mode()
def generate(
    message: str,
    chat_history: list[tuple[str, str]],
    max_new_tokens: int = 1024,
    temperature: float = 0.7,
    top_p: float = 0.95,
    top_k: int = 50,
    repetition_penalty: float = 1.0,
) -> Iterator[str]:
    conversation = []
    for user, assistant in chat_history:
        conversation.extend([{"role": "USER", "content": user}, {"role": "ASSISTANT", "content": assistant}])
    conversation.append({"role": "USER", "content": message})

    prompt = apply_chat_template(conversation)
    input_ids = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt")
    if input_ids.shape[1] > MAX_INPUT_TOKEN_LENGTH:
        input_ids = input_ids[:, -MAX_INPUT_TOKEN_LENGTH:]
        gr.Warning(f"Trimmed input from conversation as it was longer than {MAX_INPUT_TOKEN_LENGTH} tokens.")
    input_ids = input_ids.to(model.device)

    output_queue = Queue()
    def inference():
        outputs = model.generate(
            input_ids=input_ids,
            max_new_tokens=max_new_tokens,
            do_sample=True,
            top_p=top_p,
            top_k=top_k,
            temperature=temperature,
            repetition_penalty=repetition_penalty,
            pad_token_id=tokenizer.eos_token_id,
        )
        for token in tokenizer.decode(outputs[0], skip_special_tokens=True).split():
            output_queue.put(token)
        output_queue.put(None)  # 終了シグナル

    Thread(target=inference).start()

    outputs = []
    while True:
        try:
            token = output_queue.get(timeout=20.0)  # タイムアウト設定
            if token is None:
                break
            outputs.append(token)
            yield "".join(outputs)
        except Empty:
            yield "現在応答を生成中です。しばらくお待ちください。"


demo = gr.ChatInterface(
    fn=generate,
    type="tuples",
    additional_inputs_accordion=gr.Accordion(label="詳細設定", open=False),
    additional_inputs=[
        gr.Slider(
            label="Max new tokens",
            minimum=1,
            maximum=MAX_MAX_NEW_TOKENS,
            step=1,
            value=DEFAULT_MAX_NEW_TOKENS,
        ),
        gr.Slider(
            label="Temperature",
            minimum=0.1,
            maximum=4.0,
            step=0.1,
            value=0.7,
        ),
        gr.Slider(
            label="Top-p (nucleus sampling)",
            minimum=0.05,
            maximum=1.0,
            step=0.05,
            value=0.95,
        ),
        gr.Slider(
            label="Top-k",
            minimum=1,
            maximum=1000,
            step=1,
            value=50,
        ),
        gr.Slider(
            label="Repetition penalty",
            minimum=1.0,
            maximum=2.0,
            step=0.05,
            value=1.0,
        ),
    ],
    stop_btn=None,
    examples=[
        ["東京の観光名所を教えて。"],
        ["落武者って何？"],
        ["暴れん坊将軍って誰のこと？"],
        ["人がヘリを食べるのにかかる時間は？"],
    ],
    description=DESCRIPTION,
    css_paths="style.css",
    fill_height=True,
)

if __name__ == "__main__":
    demo.launch()