app.py

import spaces
import gradio as gr
from gradio import update
from functools import lru_cache
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
from opencc import OpenCC  # 用於簡體轉繁體

# 初始化簡體到繁體轉換器
cc = OpenCC('s2t')

# 可選模型列表
MODEL_LIST = [
    "liswei/Taiwan-ELM-270M",
    "Mxode/SmolLM-Chinese-180M",
    "flyingfishinwater/chinese-baby-llama2",
    "unsloth/gemma-3-1b-pt",
    "ckiplab/gpt2-tiny-chinese",
    "ckiplab/gpt2-base-chinese",
    "liswei/Taiwan-ELM-1_1B",
    "benchang1110/Qwen2.5-Taiwan-1.5B-Instruct",
    "benchang1110/Taiwan-tinyllama-v1.0-base",
    "lianghsun/Llama-3.2-Taiwan-3B",
    "twinkle-ai/Llama-3.2-3B-F1-Instruct",
    "Epiculous/Violet_Twilight-v0.2",
]

@lru_cache(maxsize=None)
def get_pipeline(model_name):
    tok = AutoTokenizer.from_pretrained(model_name)
    mdl = AutoModelForCausalLM.from_pretrained(
        model_name, weights_only=False, trust_remote_code=True
    )
    mdl.to("cuda")
    return pipeline("text-generation", model=mdl, tokenizer=tok, device=0)

@spaces.GPU
def suggest_next(text, model_name, k, m):
    """
    使用 Beam Search 產生 m 條候選，並一次更新候選列表，轉繁體並編號。
    """
    gen_pipe = get_pipeline(model_name)
    outs = gen_pipe(
        text,
        max_new_tokens=k,
        num_beams=m,
        num_return_sequences=m,
        do_sample=False,
        early_stopping=True
    )
    suggestions = [out["generated_text"][len(text):].strip() for out in outs]
    suggestions = [s for s in suggestions if s]
    suggestions = [cc.convert(s) for s in suggestions]
    numbered = [f"{i+1}. {s}" for i, s in enumerate(suggestions)]
    return update(choices=numbered, value=None)


def append_suggestion(current, choice):
    if choice is None:
        return current
    text = choice.split(". ", 1)[1] if ". " in choice else choice
    return current + text

# 自訂 CSS：模擬經典中文輸入法候選欄樣式
custom_css = """
#suggestions-bar {
    width: 100%;
    margin-bottom: 8px;
}
#suggestions-bar .candidate-list {
    display: flex;
    gap: 8px;
    background: #fff;
    border: 1px solid #999;
    border-radius: 4px;
    padding: 4px 6px;
    overflow-x: auto;
    white-space: nowrap;
}
#suggestions-bar .candidate-list input[type=radio] {
    display: none;
}
#suggestions-bar .candidate-list label {
    position: relative;
    cursor: pointer;
    padding: 4px 8px;
    font-size: 14px;
}
#suggestions-bar .candidate-list label:hover {
    background: #f5f5f5;
}
#suggestions-bar .candidate-list input[type=radio]:checked + label {
    background: #e6f7ff;
    border: 1px solid #1890ff;
}
#input-box {
    width: 100%;
}
"""

with gr.Blocks(css=custom_css) as demo:
    # 標題與說明
    gr.Markdown(
        "## 🇹🇼 繁體中文 IME 加速器  \
"
        "結合小型語言模型與 ZeroGPU，提供即時輸入法風格候選欄。"
    )

    # 垂直排列：候選欄置於上方，輸入框置於下方
    with gr.Column():
        suggestions = gr.Radio(
            [], label="", interactive=True, type="value",
            elem_id="suggestions-bar", elem_classes="candidate-list"
        )
        input_text = gr.Textbox(
            label="", placeholder="請輸入拼音或文字…",
            lines=1, max_lines=1, elem_id="input-box"
        )

    # 自動預測開關及預測按鈕（預測按鈕初始隱藏）
    with gr.Row():
        auto_predict = gr.Checkbox(
            value=True, label="自動預測（內容變更時觸發）", elem_id="auto-predict"
        )
        predict_button = gr.Button(
            "預測", elem_id="predict-button", visible=False
        )

    # 進階參數設定（可摺疊）
    with gr.Accordion("進階設定", open=False):
        model_selector = gr.Dropdown(
            MODEL_LIST, value=MODEL_LIST[0], label="模型"
        )
        k_slider = gr.Slider(
            minimum=1, maximum=50, step=1, value=1, label="K（最大新詞元數）"
        )
        m_slider = gr.Slider(
            minimum=1, maximum=30, step=1, value=6, label="M（建議數/Beam 數）"
        )

    # 事件綁定
    auto_predict.change(
        fn=lambda auto: update(visible=not auto),
        inputs=[auto_predict],
        outputs=[predict_button],
    )
    predict_button.click(
        fn=suggest_next,
        inputs=[input_text, model_selector, k_slider, m_slider],
        outputs=suggestions,
    )
    input_text.change(
        fn=lambda txt, mdl, k, m, auto: suggest_next(txt, mdl, k, m) if auto else update(choices=[], value=None),
        inputs=[input_text, model_selector, k_slider, m_slider, auto_predict],
        outputs=suggestions,
    )
    suggestions.change(
        fn=append_suggestion,
        inputs=[input_text, suggestions],
        outputs=input_text,
    )

    demo.launch()