app.py

import spaces
import gradio as gr
from gradio import update
from functools import lru_cache
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

# 可選模型列表
MODEL_LIST = [
    "unsloth/gemma-3-1b-pt",
    "ckiplab/gpt2-tiny-chinese",
    "ckiplab/gpt2-base-chinese",
    "liswei/Taiwan-ELM-270M",
    "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 條最可能的下段建議，並一次更新候選列表。
    """
    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]

    # 更新候選條
    return update(choices=suggestions, value=None)

def append_suggestion(current, choice):
    if choice is None:
        return current
    # 模擬輸入法候選選中
    return current + choice

# 自定義 CSS：模擬經典中文輸入法候選欄樣式
custom_css = """
#suggestions-bar .candidate-list {
    display: flex;
    gap: 12px;
    background: #ffffff;
    border: 1px solid #ccc;
    border-radius: 4px;
    padding: 6px;
}
#suggestions-bar .candidate-list input[type=radio] {
    display: none;
}
#suggestions-bar .candidate-list label {
    cursor: pointer;
    padding: 2px 6px;
    border-radius: 4px;
}
#suggestions-bar .candidate-list label:hover {
    background: #f0f0f0;
}
#suggestions-bar .candidate-list input[type=radio]:checked + label {
    background: #e0e0e0;
    border: 1px solid #888;
}
"""

with gr.Blocks(css=custom_css) as demo:
    # 標題和說明
    gr.Markdown(
        "## 🇹🇼 台灣中文輸入法加速器  \n"
        "結合小型語言模型與 ZeroGPU，即時 IME 風格候選條。"
    )

    # 經典候選欄：水平排列
    suggestions = gr.Radio(
        [], label="", interactive=True, type="value",
        elem_id="suggestions-bar", elem_classes="candidate-list"
    )

    # 輸入區與按鈕：單行輸入框 + 小按鈕
    with gr.Row():
        input_text = gr.Textbox(
            label="", placeholder="請輸入拼音或文字…", lines=1, max_lines=1
        )
        gpu_button = gr.Button("建議")

    # 進階參數設定（可折疊）
    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=10, label="M（建議數/Beam 數）"
        )

    # 事件綁定
    gpu_button.click(
        fn=suggest_next,
        inputs=[input_text, model_selector, k_slider, m_slider],
        outputs=suggestions,
    )
    suggestions.change(
        fn=append_suggestion,
        inputs=[input_text, suggestions],
        outputs=input_text,
    )

    demo.launch()