Deploy

README.md

@@ -0,0 +1,32 @@
+---
+title: Murasaki LLM 在线演示
+emoji: 🌸
+colorFrom: purple
+colorTo: indigo
+sdk: gradio
+sdk_version: "4.44.0"
+app_file: app.py
+pinned: true
+license: apache-2.0
+---
+
+# Murasaki LLM 在线演示
+
+**ACGN 日中翻译模型** - 专为轻小说、Galgame 剧本设计的高质量翻译引擎。
+
+## ✨ 特性
+
+- 🎯 **深度思考 (CoT)**: 翻译前先分析文风、补全主语、梳理逻辑
+- 📖 **双模式**: 轻小说文学风 / 剧本口语化
+- ⚡ **双模型**: 8B 轻量快速 / 14B 高质量
+
+## 📊 限制
+
+- 每日 30 次请求 / 20,000 字符
+- 单次 100-5,000 字符
+- 首次冷启动约 30 秒
+
+## 🔗 链接
+
+- [GitHub](https://github.com/soundstarrain/Murasaki-Translator)
+- [模型仓库](https://huggingface.co/Murasaki-Project)

app.py

@@ -0,0 +1,510 @@
+"""
+Murasaki 翻译模型在线演示
+布局: 两栏 + 底部可折叠思考面板
+"""
+
+import os, re, json, requests, time
+from datetime import date
+from typing import Generator, Tuple
+import gradio as gr
+
+# ============================================================
+# 配置
+# ============================================================
+MODAL_ENDPOINT_URL = os.getenv("MODAL_ENDPOINT_URL", "http://localhost:8000")
+DEMO_SECRET_TOKEN = os.getenv("DEMO_SECRET_TOKEN", "")
+UPSTASH_URL = os.getenv("UPSTASH_REDIS_REST_URL", "")
+UPSTASH_TOKEN = os.getenv("UPSTASH_REDIS_REST_TOKEN", "")
+
+DAILY_CHAR_QUOTA = 20000
+DAILY_REQ_QUOTA = 30
+MIN_TEXT_LENGTH = 100
+MAX_REQUEST_SIZE = 5000
+MAX_CHUNK_SIZE = 1500
+MIN_CHUNK_SIZE = 1000
+REQUEST_TIMEOUT = 180
+
+GITHUB_URL = "https://github.com/soundstarrain/Murasaki-Translator"
+HF_REPO_URL = "https://huggingface.co/Murasaki-Project"
+
+# ============================================================
+# 辅助函数
+# ============================================================
+def estimate_tokens(text: str) -> int:
+    return max(1, int(len(text) / 1.5))
+
+def post_process(text: str) -> str:
+    text = text.replace('"', '「').replace('"', '」').replace(''', '『').replace(''', '』')
+    lines = []
+    for line in text.splitlines():
+        if line.count('"') > 0 and line.count('"') % 2 == 0:
+            line = re.sub(r'"([^"]*)"', r'「\1」', line)
+        lines.append(line)
+    return "\n\n".join([l.rstrip() for l in "\n".join(lines).splitlines() if l.strip()])
+
+def stream_parse(raw: str) -> Tuple[str, str]:
+    if '<think>' in raw:
+        parts = raw.split('<think>', 1)[1]
+        if '</think>' in parts:
+            t = parts.split('</think>', 1)
+            return t[0].strip(), t[1].lstrip()
+        return parts.strip(), ""
+    return "", raw.strip()
+
+# ============================================================
+# Redis
+# ============================================================
+def get_redis():
+    if not UPSTASH_URL or not UPSTASH_TOKEN: return None
+    try:
+        from upstash_redis import Redis
+        return Redis(url=UPSTASH_URL, token=UPSTASH_TOKEN)
+    except: return None
+
+def check_quota(ip, chars):
+    r = get_redis()
+    if not r: return True, DAILY_CHAR_QUOTA, DAILY_REQ_QUOTA, ""
+    try:
+        ck, rk = f"demo:c:{ip}:{date.today()}", f"demo:r:{ip}:{date.today()}"
+        uc, ur = int(r.get(ck) or 0), int(r.get(rk) or 0)
+        rc, rr = DAILY_CHAR_QUOTA - uc, DAILY_REQ_QUOTA - ur
+        if rr <= 0: return False, rc, 0, "请求次数用尽"
+        if chars > rc: return False, rc, rr, "字符额度不足"
+        return True, rc, rr, ""
+    except: return True, 0, 0, ""
+
+def deduct_quota(ip, chars):
+    r = get_redis()
+    if not r: return 0, 0
+    try:
+        ck, rk = f"demo:c:{ip}:{date.today()}", f"demo:r:{ip}:{date.today()}"
+        r.incrby(ck, chars); r.incr(rk); r.expire(ck, 86400); r.expire(rk, 86400)
+        return max(0, DAILY_CHAR_QUOTA - int(r.get(ck) or 0)), max(0, DAILY_REQ_QUOTA - int(r.get(rk) or 0))
+    except: return 0, 0
+
+# ============================================================
+# 翻译逻辑
+# ============================================================
+def split_chunks(text):
+    if len(text) <= MAX_CHUNK_SIZE: return [text]
+    paragraphs = re.split(r'\n\n+', text); chunks, cur = [], ""
+    for p in paragraphs:
+        if len(p) > MAX_CHUNK_SIZE:
+            if cur: chunks.append(cur.strip()); cur = ""
+            for s in re.split(r'(?<=[。！？\n])', p):
+                if len(cur) + len(s) <= MAX_CHUNK_SIZE: cur += s
+                else: 
+                    if cur: chunks.append(cur.strip())
+                    cur = s
+        elif len(cur) + len(p) + 2 <= MAX_CHUNK_SIZE: cur += ("\n\n" if cur else "") + p
+        elif len(cur) >= MIN_CHUNK_SIZE: chunks.append(cur.strip()); cur = p
+        else: cur += ("\n\n" if cur else "") + p
+    if cur: chunks.append(cur.strip())
+    return chunks
+
+def translate_stream(text, model, preset, ip):
+    text = text.strip()
+    if not text: yield "", "", "输入为空"; return
+    if len(text) < MIN_TEXT_LENGTH: yield "", "", f"最少 {MIN_TEXT_LENGTH} 字"; return
+    if len(text) > MAX_REQUEST_SIZE: yield "", "", f"最多 {MAX_REQUEST_SIZE} 字"; return
+    ok, rc, rr, err = check_quota(ip, len(text))
+    if not ok: yield "", "", f"限额已用尽: {err}"; return
+    
+    start_time = time.time()
+    input_tokens = estimate_tokens(text)
+    
+    chunks = split_chunks(text)
+    headers = {"Content-Type": "application/json"}
+    if DEMO_SECRET_TOKEN: headers["Authorization"] = f"Bearer {DEMO_SECRET_TOKEN}"
+    trans, think, total_output = "", "", ""
+    
+    for i, chunk in enumerate(chunks):
+        try:
+            resp = requests.post(f"{MODAL_ENDPOINT_URL}/translate", 
+                json={"text": chunk, "model": model, "preset": preset, "stream": True},
+                headers=headers, stream=True, timeout=REQUEST_TIMEOUT)
+            if resp.status_code != 200: yield trans, think, "服务错误"; return
+            raw = ""
+            for line in resp.iter_lines(decode_unicode=True):
+                if line and line.startswith("data: "):
+                    d = line[6:]
+                    if d == "[DONE]": break
+                    data = json.loads(d)
+                    if "error" in data:
+                        yield trans, think, f"服务错误: {data['error']}"
+                        return
+                    raw += data.get("content", "")
+                    ct, cr = stream_parse(raw)
+                    elapsed = time.time() - start_time
+                    status = f"处理中 {i+1}/{len(chunks)} · {elapsed:.1f}s"
+                    yield trans + ("\n\n" if trans and cr else "") + cr, think + ("\n" if think and ct else "") + ct, status
+            t, th = stream_parse(raw)
+            trans += ("\n\n" if trans else "") + post_process(t)
+            total_output += raw
+            if th: think += ("\n\n" if think else "") + th
+        except Exception as e: yield trans, think, str(e); return
+    
+    elapsed = time.time() - start_time
+    output_tokens = estimate_tokens(total_output)
+    speed = output_tokens / elapsed if elapsed > 0 else 0
+    
+    rc, rr = deduct_quota(ip, len(text))
+    yield trans, think, f"完成 · 输入 {input_tokens} tokens · 输出 {output_tokens} tokens · {elapsed:.1f}s ({speed:.1f} t/s) · 剩余 {rc:,}字/{rr}次"
+
+def wrapper(text, model, preset, req: gr.Request):
+    yield from translate_stream(text, model, preset, req.client.host if req else "x")
+
+# ============================================================
+# 自定义主题 - 白色背景 + 浅紫色关键文字
+# ============================================================
+theme = gr.themes.Soft(
+    primary_hue="violet",
+    secondary_hue="purple", 
+    neutral_hue="slate",
+).set(
+    body_background_fill="#fafafa",
+    body_background_fill_dark="#fafafa",
+    block_background_fill="#ffffff",
+    block_background_fill_dark="#ffffff",
+    input_background_fill="#f8f9fa",
+    input_background_fill_dark="#f8f9fa",
+    button_primary_background_fill="#8b5cf6",
+    button_primary_background_fill_dark="#8b5cf6",
+    button_primary_background_fill_hover="#7c3aed",
+    button_primary_background_fill_hover_dark="#7c3aed",
+    body_text_color="#374151",
+    body_text_color_dark="#374151",
+    body_text_color_subdued="#6b7280",
+    body_text_color_subdued_dark="#6b7280",
+    block_label_background_fill="transparent",
+    block_label_background_fill_dark="transparent",
+    block_label_text_color="#8b5cf6",
+    block_label_text_color_dark="#8b5cf6",
+    block_title_text_color="#8b5cf6",
+    block_title_text_color_dark="#8b5cf6",
+    block_border_width="1px",
+    block_border_color="#e5e7eb",
+    block_border_color_dark="#e5e7eb",
+    input_border_width="1px",
+    input_border_color="#e5e7eb",
+    input_border_color_dark="#e5e7eb",
+    block_radius="12px",
+    input_radius="8px",
+)
+
+CSS = """
+/* ============================================
+   全局字体基准
+   ============================================ */
+.gradio-container {
+    font-size: 16px !important;
+}
+
+/* 标题区 */
+.app-header {
+    text-align: center;
+    padding: 20px 0 12px 0;
+}
+.app-header h1 {
+    font-size: 2rem;
+    font-weight: 700;
+    color: #8b5cf6;
+    margin: 0;
+    letter-spacing: -0.02em;
+}
+.app-header p {
+    color: #6b7280;
+    font-size: 1rem;
+    margin: 6px 0 0 0;
+}
+
+/* 说明框 */
+.info-box {
+    background: #f3f0ff;
+    border: 1px solid #e9d5ff;
+    border-radius: 8px;
+    padding: 12px 16px;
+    margin-bottom: 12px;
+}
+.info-box h3 {
+    color: #7c3aed;
+    font-size: 0.9rem;
+    font-weight: 600;
+    margin: 0 0 6px 0;
+}
+.info-box p {
+    color: #4b5563;
+    font-size: 0.9rem;
+    line-height: 1.5;
+    margin: 3px 0;
+}
+.info-box a {
+    color: #8b5cf6;
+    text-decoration: none;
+    font-weight: 500;
+}
+.info-box a:hover {
+    text-decoration: underline;
+}
+
+/* 面板标签 */
+.panel-label {
+    color: #8b5cf6;
+    font-size: 0.85rem;
+    font-weight: 600;
+    text-transform: uppercase;
+    letter-spacing: 0.08em;
+    margin-bottom: 6px;
+}
+
+/* ============================================
+   全局边框重置 - 移除所有嵌套边框
+   ============================================ */
+
+/* 所有 block 容器移除边框 */
+.gradio-container .block {
+    border: none !important;
+    box-shadow: none !important;
+    background: transparent !important;
+}
+
+/* Form 容器移除边框 */
+.gradio-container .form {
+    border: none !important;
+    box-shadow: none !important;
+    background: transparent !important;
+}
+
+/* 嵌套的 wrap 容器移除边框 */
+.gradio-container .wrap {
+    border: none !important;
+    box-shadow: none !important;
+}
+
+/* ============================================
+   文本框 - 只在 textarea 本身保留边框
+   ============================================ */
+.gradio-textbox {
+    border: none !important;
+    box-shadow: none !important;
+    background: transparent !important;
+}
+.gradio-textbox > .wrap,
+.gradio-textbox > div {
+    border: none !important;
+    box-shadow: none !important;
+    background: transparent !important;
+}
+textarea {
+    background: #f8f9fa !important;
+    border: 1px solid #e5e7eb !important;
+    border-radius: 8px !important;
+    color: #374151 !important;
+    font-size: 0.95rem !important;
+    line-height: 1.6 !important;
+    padding: 10px 12px !important;
+}
+textarea::placeholder {
+    color: #9ca3af !important;
+    font-size: 0.95rem !important;
+}
+textarea:focus {
+    border-color: #a78bfa !important;
+    box-shadow: 0 0 0 2px rgba(139, 92, 246, 0.1) !important;
+}
+
+/* ============================================
+   下拉框 - 只保留外层单一边框
+   ============================================ */
+.gradio-dropdown {
+    border: none !important;
+    box-shadow: none !important;
+    background: transparent !important;
+}
+.gradio-dropdown > .wrap,
+.gradio-dropdown > div:not(ul) {
+    border: 1px solid #e5e7eb !important;
+    border-radius: 8px !important;
+    background: #ffffff !important;
+    box-shadow: none !important;
+    min-height: 38px !important;
+}
+.gradio-dropdown .wrap .wrap,
+.gradio-dropdown .secondary-wrap,
+.gradio-dropdown .border-none {
+    border: none !important;
+    box-shadow: none !important;
+    background: transparent !important;
+}
+.gradio-dropdown input,
+.gradio-dropdown select {
+    border: none !important;
+    box-shadow: none !important;
+    background: transparent !important;
+    color: #374151 !important;
+    font-size: 0.95rem !important;
+}
+/* 下拉菜单列表 */
+.gradio-dropdown ul[role="listbox"],
+.gradio-container ul[role="listbox"] {
+    background: #ffffff !important;
+    border: 1px solid #e5e7eb !important;
+    border-radius: 8px !important;
+    box-shadow: 0 4px 12px rgba(0,0,0,0.1) !important;
+}
+.gradio-dropdown ul li,
+.gradio-container ul[role="listbox"] li {
+    background: #ffffff !important;
+    color: #374151 !important;
+}
+.gradio-dropdown ul li:hover,
+.gradio-container ul[role="listbox"] li:hover {
+    background: #f3f0ff !important;
+}
+
+/* ============================================
+   折叠面板 - 简洁单边框
+   ============================================ */
+.gradio-accordion {
+    border: 1px solid #e5e7eb !important;
+    border-radius: 8px !important;
+    background: #ffffff !important;
+    box-shadow: none !important;
+    overflow: hidden;
+    margin-top: 8px !important;
+}
+.gradio-accordion > div,
+.gradio-accordion .label-wrap,
+.gradio-accordion button {
+    border: none !important;
+    box-shadow: none !important;
+    background: #ffffff !important;
+    color: #374151 !important;
+    font-size: 0.95rem !important;
+    padding: 8px 12px !important;
+}
+.gradio-accordion button span {
+    color: #8b5cf6 !important;
+    font-size: 0.95rem !important;
+    font-weight: 600 !important;
+}
+.gradio-accordion .content {
+    border-top: 1px solid #e5e7eb !important;
+    padding: 8px 12px !important;
+}
+
+/* ============================================
+   按钮
+   ============================================ */
+button.primary {
+    margin-top: 16px !important;
+    font-weight: 600 !important;
+    font-size: 0.95rem !important;
+    padding: 8px 20px !important;
+    min-height: 40px !important;
+    box-shadow: 0 2px 8px rgba(139, 92, 246, 0.25) !important;
+    border: none !important;
+    border-radius: 8px !important;
+    transition: all 0.15s ease !important;
+}
+button.primary:hover {
+    box-shadow: 0 3px 12px rgba(139, 92, 246, 0.35) !important;
+    transform: translateY(-1px) !important;
+}
+
+/* ============================================
+   状态栏和禁用输入
+   ============================================ */
+input:disabled, textarea:disabled {
+    background: #f8f9fa !important;
+    color: #6b7280 !important;
+    border: 1px solid #e5e7eb !important;
+}
+
+/* Row 容器 */
+.gradio-row {
+    border: none !important;
+    box-shadow: none !important;
+}
+
+/* Column 容器 */
+.gradio-column {
+    border: none !important;
+    box-shadow: none !important;
+}
+
+/* 标签文字 */
+.gradio-container label,
+.gradio-container .label-wrap span {
+    font-size: 0.95rem !important;
+    font-weight: 500 !important;
+    color: #8b5cf6 !important;
+}
+
+/* 状态栏 */
+.gradio-container input[type="text"]:disabled {
+    font-size: 0.95rem !important;
+    color: #6b7280 !important;
+}
+
+/* 全局深色覆盖 */
+.dark, [data-theme="dark"] {
+    --background-fill-primary: #fafafa !important;
+    --background-fill-secondary: #f8f9fa !important;
+    --block-background-fill: transparent !important;
+    --input-background-fill: #f8f9fa !important;
+    --body-text-color: #374151 !important;
+}
+
+footer { display: none !important; }
+"""
+
+with gr.Blocks(title="Murasaki LLM 在线演示", theme=theme, css=CSS) as demo:
+    gr.HTML("""
+        <div class='app-header'>
+            <h1>Murasaki LLM 在线演示</h1>
+            <p>ACGN日中翻译模型</p>
+        </div>
+    """)
+    
+    gr.HTML(f"""
+        <div class='info-box'>
+            <h3>使用说明</h3>
+            <p>1. 在线演示限额：每日 {DAILY_REQ_QUOTA} 次请求 / {DAILY_CHAR_QUOTA:,} 字符，单次 {MIN_TEXT_LENGTH}-{MAX_REQUEST_SIZE} 字符</p>
+            <p>2. 首次冷启动需要约 30 秒加载模型，请耐心等待</p>
+            <p>3. 推荐从 <a href='{HF_REPO_URL}'>模型仓库</a> 下载模型配合 <a href='{GITHUB_URL}'>Murasaki Translator</a> 本地运行</p>
+        </div>
+    """)
+    
+    with gr.Row():
+        with gr.Column():
+            gr.HTML("<div class='panel-label'>原文</div>")
+            src = gr.Textbox(placeholder="粘贴日语原文...", lines=16, show_label=False)
+        with gr.Column():
+            gr.HTML("<div class='panel-label'>译文</div>")
+            res = gr.Textbox(placeholder="翻译结果将在此显示...", lines=16, show_label=False, interactive=False, show_copy_button=True)
+    
+    with gr.Row():
+        m = gr.Dropdown(choices=["Murasaki-8B", "Murasaki-14B"], value="Murasaki-8B", label="模型", interactive=True, scale=2)
+        p = gr.Dropdown(choices=["轻小说", "剧本"], value="轻小说", label="预设", interactive=True, scale=2)
+        btn = gr.Button("翻译", variant="primary", scale=1)
+    
+    thinking_accordion = gr.Accordion("AI 推理过程", open=False)
+    with thinking_accordion:
+        cot = gr.Textbox(placeholder="模型思考过程...", lines=5, show_label=False, interactive=False)
+    
+    status = gr.Textbox(show_label=False, interactive=False, value="准备就绪", container=False)
+
+    def translate_wrapper(text, model, preset, req: gr.Request):
+        model_map = {"Murasaki-8B": "8b", "Murasaki-14B": "14b"}
+        preset_map = {"轻小说": "novel", "剧本": "script"}
+        for trans, think, st in translate_stream(text, model_map.get(model, "8b"), preset_map.get(preset, "novel"), req.client.host if req else "x"):
+            # 有思考内容时自动展开
+            should_open = bool(think and think.strip())
+            yield trans, think, st, gr.update(open=should_open)
+    
+    btn.click(fn=translate_wrapper, inputs=[src, m, p], outputs=[res, cot, status, thinking_accordion])
+    src.submit(fn=translate_wrapper, inputs=[src, m, p], outputs=[res, cot, status, thinking_accordion])
+
+if __name__ == "__main__":
+    demo.launch(server_name="0.0.0.0", server_port=7860)

mock_server.py

@@ -0,0 +1,79 @@
+"""
+本地测试 Mock 服务器 (V2)
+
+用于模拟 Murasaki 模型带 <think> 标签的流式输出，配合三栏 UI 测试。
+
+使用方法：
+    python mock_server.py
+"""
+
+import json
+import asyncio
+from fastapi import FastAPI
+from fastapi.responses import StreamingResponse
+from pydantic import BaseModel
+from typing import Literal
+
+app = FastAPI()
+
+class TranslateRequest(BaseModel):
+    text: str
+    model: Literal["8b", "14b"] = "8b"
+    preset: Literal["novel", "script", "short"] = "novel"
+    stream: bool = False
+
+MOCK_DATA = {
+    "novel": {
+        "think": "正在分析文风... 这是一个描写雨天的片段，语境忧郁。主语补全为『她』。采用轻小说优美文风。",
+        "trans": "她凝视着窗外。雨滴顺着明净的玻璃窗缓缓滑落，像是断了线的珍珠，在灰暗的天色中留下一道道模糊的痕迹。"
+    },
+    "script": {
+        "think": "对话场景分析。语气：平淡。角色：路人甲。采用简洁翻译。",
+        "trans": "「嗯，我知道了。」"
+    }
+}
+
+@app.post("/translate")
+async def translate(request: TranslateRequest):
+    if request.stream:
+        async def stream_generator():
+            data = MOCK_DATA.get(request.preset, MOCK_DATA["novel"])
+            
+            # 构造带标签的原始输出
+            full_text = f"<think>\n{data['think']}\n</think>\n{data['trans']}"
+            
+            # 逐字输出模拟流式
+            for char in full_text:
+                yield f"data: {json.dumps({'content': char})}\n\n"
+                await asyncio.sleep(0.01) # 快一点
+            
+            yield "data: [DONE]\n\n"
+        
+        return StreamingResponse(stream_generator(), media_type="text/event-stream")
+    
+    # 非流式
+    data = MOCK_DATA.get(request.preset, MOCK_DATA["novel"])
+    return {
+        "translation": data["trans"],
+        "thinking": data["think"],
+        "model_used": f"Murasaki {request.model.upper()} (Mock)"
+    }
+
+@app.get("/health")
+async def health(): return {"status": "ok"}
+
+@app.get("/models")
+async def list_models():
+    return {
+        "models": [
+            {"id": "8b", "display_name": "Murasaki 8B (Mock)"},
+            {"id": "14b", "display_name": "Murasaki 14B (Mock)"}
+        ],
+        "default": "8b"
+    }
+
+if __name__ == "__main__":
+    import uvicorn
+    print("\n🚀 Mock 服务器 V2 启动中...")
+    print("📍 地址: http://localhost:8000")
+    uvicorn.run(app, host="0.0.0.0", port=8000)

requirements.txt

@@ -0,0 +1,8 @@
+# HF Space 依赖
+gradio>=4.0.0
+requests>=2.28.0
+upstash-redis>=1.0.0
+
+# 本地测试用（可选）
+fastapi>=0.100.0
+uvicorn>=0.20.0