app.py

# 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",
]

def merge_common_prefixes(suggestions, min_len=2):
    prefixes = []
    to_remove = set()
    for i in range(len(suggestions)):
        for j in range(i+1, len(suggestions)):
            s1, s2 = suggestions[i], suggestions[j]
            common = ''.join(c1 for c1, c2 in zip(s1, s2) if c1 == c2)
            if len(common) >= min_len:
                prefixes.append(common)
                to_remove.update([s1, s2])
    unique_prefixes = []
    for p in prefixes:
        if p not in unique_prefixes:
            unique_prefixes.append(p)
    remainder = [s for s in suggestions if s not in to_remove]
    return unique_prefixes + remainder

@lru_cache(maxsize=8)
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, num_beam_groups, diversity_penalty):
    gen_pipe = get_pipeline(model_name)
    # 構造 generate 參數字典，僅在 penalty>0 時加入 diversity 相關
    gen_kwargs = {
        "max_new_tokens": k,
        "num_beams": m,
        "num_return_sequences": m,
        "do_sample": False,
        "early_stopping": True,
    }
    if diversity_penalty and diversity_penalty > 0:
        gen_kwargs["num_beam_groups"] = num_beam_groups
        gen_kwargs["diversity_penalty"] = diversity_penalty

    outs = gen_pipe(text, **gen_kwargs)

    # 提取純下文、過濾空字串、繁體化
    suggestions = [
        cc.convert(out["generated_text"][len(text):].strip())
        for out in outs
        if out["generated_text"][len(text):].strip()
    ]

    # 去重
    unique_suggestions = []
    for s in suggestions:
        if s not in unique_suggestions:
            unique_suggestions.append(s)

    # 合併共同前綴
    final_suggestions = merge_common_prefixes(unique_suggestions, min_len=2)

    return update(choices=final_suggestions, value=None)

def append_suggestion(text, choice):
    return text + choice

with gr.Blocks(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: 6px; overflow-x: auto; white-space: nowrap;
}
#suggestions-bar .candidate-list label { cursor: pointer; }
""") as demo:
    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=20, elem_id="input-box"
        )

    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")

    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=10, label="K（最大新詞元數）"
        )
        m_slider = gr.Slider(
            minimum=1, maximum=30, step=1, value=30, label="M（建議數/Beam 數）"
        )
        group_slider = gr.Slider(
            minimum=1, maximum=30, step=1, value=30,
            label="Beam 群組數 (num_beam_groups)"
        )
        diversity_penalty_slider = gr.Slider(
            minimum=0.0, maximum=2.0, step=0.1, value=1.0,
            label="多樣性懲罰 (diversity_penalty)"
        )

    predict_button.click(
        fn=suggest_next,
        inputs=[
            input_text,
            model_selector,
            k_slider,
            m_slider,
            group_slider,
            diversity_penalty_slider
        ],
        outputs=suggestions,
    )
    input_text.change(
        fn=lambda txt, mdl, k, m, g, d, auto: (
            suggest_next(txt, mdl, k, m, g, d)
            if auto else update(choices=[], value=None)
        ),
        inputs=[
            input_text,
            model_selector,
            k_slider,
            m_slider,
            group_slider,
            diversity_penalty_slider,
            auto_predict
        ],
        outputs=suggestions,
    )
    suggestions.change(
        fn=append_suggestion,
        inputs=[input_text, suggestions],
        outputs=input_text,
    )

    demo.launch()