Upload 3 files

README.md

@@ -12,7 +12,6 @@ hardware: zero-gpu
 license: apache-2.0
 ---
 
-
 # 🎵 ACE-Step v1.5 AI 音樂生成器
 
 基於 ACE-Step v1.5 模型的音樂生成 API（完全相容 OpenAI 格式）與 Web 使用者介面。
@@ -21,6 +20,3 @@ license: apache-2.0
 - `POST /v1/chat/completions` - 音樂生成
 - `GET /v1/models` - 獲取可用模型列表  
 - `GET /health` - 伺服器健康檢查
-
-## 認證設定
-請在 Hugging Face Space 的 Settings -> Secrets 中新增 `API_KEY` 變數，即可啟用 Bearer Token 認證。若未設定則為開放存取。

app.py

@@ -3,22 +3,31 @@ import time
 import base64
 import json
 import uuid
-import io
+import asyncio
+import traceback
 import torch
 import spaces
 import gradio as gr
 import numpy as np
-import soundfile as sf
 from fastapi import FastAPI, HTTPException, Request
 from fastapi.responses import StreamingResponse, JSONResponse
 from fastapi.middleware.cors import CORSMiddleware
 from pydantic import BaseModel
 from typing import Optional, List, Union, Any
+from gradio_client import Client
+
+# ─── 1. 在啟動時預先下載模型 ────────────────────────────
+from huggingface_hub import snapshot_download
+print("==== 正在預先下載模型到快取 (避免第一次請求超時) ====")
+try:
+    snapshot_download("ACE-Step/ACE-Step-v1-3.5B")
+    print("==== 模型下載完成！ ====")
+except Exception as e:
+    print(f"模型下載失敗: {e}")
 
 # ─── 環境變數 ───────────────────────────────────────────
 API_KEY = os.environ.get("API_KEY", None)
 MODEL_ID = "acemusic/acestep-v15-turbo"
-HF_MODEL_REPO = "ACE-Step/ACE-Step-v1-3.5B"
 
 # ─── 模型全局實例 ────────────────────────────────────────
 pipeline = None
@@ -26,12 +35,12 @@ pipeline = None
 def get_pipeline():
     global pipeline
     if pipeline is None:
+        print("初始化 ACE-Step Pipeline...")
         from acestep.pipeline_ace_step import ACEStepPipeline
-        # 自動從 Hugging Face Hub 下載模型到 ~/.cache
         pipeline = ACEStepPipeline(
             checkpoint_dir=None,   
             dtype="bfloat16",
-            device="cuda",
+            device="cuda",  # 在 ZeroGPU 的 wrapper 內呼叫時才安全
         )
     return pipeline
 
@@ -82,14 +91,11 @@ class ChatCompletionRequest(BaseModel):
     repainting_end: Optional[float] = None
     audio_cover_strength: float = 1.0
 
-# ─── 輸入解析 ────────────────────────────────────────────
 def parse_input(req: ChatCompletionRequest):
-    """解析 messages，提取出音樂描述 (prompt) 和歌詞 (lyrics)"""
     last_user_msg = ""
     for msg in reversed(req.messages):
         if msg.role == "user":
-            if isinstance(msg.content, str):
-                last_user_msg = msg.content
+            if isinstance(msg.content, str): last_user_msg = msg.content
             elif isinstance(msg.content, list):
                 for item in msg.content:
                     if isinstance(item, dict) and item.get("type") == "text":
@@ -100,345 +106,171 @@ def parse_input(req: ChatCompletionRequest):
     prompt = ""
     lyrics = req.lyrics or ""
 
-    # 標籤模式：包含 <prompt>...</prompt> 或 <lyrics>...</lyrics>
     if "<prompt>" in last_user_msg:
         import re
         p_match = re.search(r"<prompt>(.*?)</prompt>", last_user_msg, re.DOTALL)
         l_match = re.search(r"<lyrics>(.*?)</lyrics>", last_user_msg, re.DOTALL)
         prompt = p_match.group(1).strip() if p_match else ""
-        if not lyrics and l_match:
-            lyrics = l_match.group(1).strip()
-    elif lyrics:
-        # 歌詞與描述分離模式
-        prompt = last_user_msg
-    elif req.sample_mode:
-        # 讓 LLM 自動生成模式
+        if not lyrics and l_match: lyrics = l_match.group(1).strip()
+    elif lyrics or req.sample_mode:
         prompt = last_user_msg
     else:
-        # 自動偵測模式：如果包含 [Verse] 或 [Chorus] 等結構，視為純歌詞模式
         if any(tag in last_user_msg for tag in ["[Verse", "[verse", "[Chorus", "[chorus", "[Bridge"]):
             lyrics = last_user_msg
         else:
             prompt = last_user_msg
-
     return prompt, lyrics
 
-# ─── 音頻轉 Base64 ───────────────────────────────────────
-def audio_to_base64(audio_data: np.ndarray, sample_rate: int, fmt: str = "mp3") -> str:
-    """將生成的音頻陣列轉換為 Base64 Data URL"""
-    buf = io.BytesIO()
-    sf.write(buf, audio_data, sample_rate, format="WAV")
-    buf.seek(0)
-    audio_bytes = buf.read()
-    b64 = base64.b64encode(audio_bytes).decode("utf-8")
-    return f"data:audio/wav;base64,{b64}"
-
-# ─── 核心生成函數（使用 ZeroGPU 裝飾器）─────────────────
+# ─── Gradio 核心生成函數 (支援 ZeroGPU) ───────────────────
 @spaces.GPU(duration=120)
-def generate_music(
-    prompt: str,
-    lyrics: str,
-    duration: Optional[float],
-    bpm: Optional[int],
-    vocal_language: str,
-    instrumental: Optional[bool],
-    guidance_scale: float,
-    seed: Optional[int],
-    task_type: str,
-    repainting_start: float,
-    repainting_end: Optional[float],
-):
-    pipe = get_pipeline()
-
-    gen_kwargs = dict(
-        prompt=prompt if prompt else "instrumental music",
-        lyrics=lyrics,
-        audio_duration=duration or 30.0,
-        guidance_scale=guidance_scale,
-        infer_steps=27,
-        scheduler_type="euler",
-    )
-
-    if bpm is not None:
-        gen_kwargs["bpm"] = bpm
-    if vocal_language:
-        gen_kwargs["vocal_language"] = vocal_language
-    if instrumental is not None:
-        gen_kwargs["instrumental"] = instrumental
-    if seed is not None:
-        gen_kwargs["seed"] = int(seed)
-
-    result = pipe(**gen_kwargs)
-
-    # 提取音頻陣列數據
-    if hasattr(result, "audio"):
-        audio_data = result.audio
-    elif isinstance(result, tuple):
-        audio_data = result[0]
-    else:
-        audio_data = result
+def gradio_generate(prompt, lyrics, duration, bpm, vocal_language, instrumental, guidance_scale, seed):
+    try:
+        seed_val = int(seed) if seed is not None and str(seed).strip() != "" else None
+        bpm_val = int(bpm) if bpm is not None and str(bpm).strip() != "" else None
+        dur_val = float(duration) if duration is not None else 30.0
+
+        lang_map = {"英文 (en)": "en", "中文 (zh)": "zh", "日文 (ja)": "ja", "韓文 (ko)": "ko", "自動判定": "en"}
+        lang = lang_map.get(vocal_language, "en") if vocal_language else "en"
+        instr = True if instrumental in ["是", "Yes", True] else (False if instrumental in ["否", "No", False] else None)
+
+        pipe = get_pipeline()
+
+        gen_kwargs = dict(
+            prompt=prompt if prompt else "instrumental music",
+            lyrics=lyrics,
+            audio_duration=dur_val,
+            guidance_scale=float(guidance_scale) if guidance_scale else 7.0,
+            infer_steps=27,
+            scheduler_type="euler",
+        )
+        if bpm_val: gen_kwargs["bpm"] = bpm_val
+        if lang: gen_kwargs["vocal_language"] = lang
+        if instr is not None: gen_kwargs["instrumental"] = instr
+        if seed_val is not None: gen_kwargs["seed"] = seed_val
+
+        result = pipe(**gen_kwargs)
+
+        if hasattr(result, "audio"): audio_data = result.audio
+        elif isinstance(result, tuple): audio_data = result[0]
+        else: audio_data = result
 
-    sample_rate = getattr(result, "sample_rate", 44100)
-    if isinstance(audio_data, torch.Tensor):
-        audio_data = audio_data.cpu().numpy()
-    if audio_data.ndim > 1:
-        audio_data = audio_data.squeeze()
+        sample_rate = getattr(result, "sample_rate", 44100)
+        if isinstance(audio_data, torch.Tensor): audio_data = audio_data.cpu().numpy()
+        if audio_data.ndim > 1: audio_data = audio_data.squeeze()
 
-    return audio_data, sample_rate
+        return (sample_rate, audio_data)
+    except Exception as e:
+        traceback.print_exc()
+        raise gr.Error(f"生成失敗: {str(e)}")
 
 # ─── FastAPI 應用程式 ────────────────────────────────────
-fastapi_app = FastAPI(title="ACE-Step OpenRouter API", version="1.0")
-fastapi_app.add_middleware(
-    CORSMiddleware,
-    allow_origins=["*"],
-    allow_methods=["*"],
-    allow_headers=["*"],
-)
+fastapi_app = FastAPI(title="ACE-Step OpenRouter API")
+fastapi_app.add_middleware(CORSMiddleware, allow_origins=["*"], allow_methods=["*"], allow_headers=["*"])
 
 @fastapi_app.get("/health")
-async def health():
-    return {"status": "ok", "service": "ACE-Step OpenRouter API", "version": "1.0"}
+async def health(): return {"status": "ok"}
 
 @fastapi_app.get("/v1/models")
 async def list_models(request: Request):
     check_auth(request)
-    return {
-        "data": [{
-            "id": MODEL_ID,
-            "name": "ACE-Step v1.5",
-            "created": 1706688000,
-            "description": "高效能文字轉音樂生成模型",
-            "input_modalities": ["text", "audio"],
-            "output_modalities": ["audio", "text"],
-            "context_length": 4096,
-            "pricing": {"prompt": "0", "completion": "0", "request": "0"},
-            "supported_sampling_parameters": ["temperature", "top_p"]
-        }]
-    }
+    return {"data": [{"id": MODEL_ID, "name": "ACE-Step v1.5", "created": 1706688000, "pricing": {"prompt": "0", "completion": "0", "request": "0"}}]}
 
 @fastapi_app.post("/v1/chat/completions")
 async def chat_completions(req: ChatCompletionRequest, request: Request):
     check_auth(request)
-
     completion_id = f"chatcmpl-{uuid.uuid4().hex[:16]}"
     created_ts = int(time.time())
 
     try:
         prompt, lyrics = parse_input(req)
         audio_cfg = req.audio_config or AudioConfig()
-
-        seed_val = None
-        if req.seed is not None:
-            seed_str = str(req.seed)
-            seed_val = int(seed_str.split(",")[0].strip())
-
-        audio_data, sample_rate = generate_music(
-            prompt=prompt,
-            lyrics=lyrics,
-            duration=audio_cfg.duration,
-            bpm=audio_cfg.bpm,
-            vocal_language=audio_cfg.vocal_language,
-            instrumental=audio_cfg.instrumental,
-            guidance_scale=req.guidance_scale,
-            seed=seed_val,
-            task_type=req.task_type,
-            repainting_start=req.repainting_start,
-            repainting_end=req.repainting_end,
-        )
-
-        audio_url = audio_to_base64(audio_data, sample_rate, audio_cfg.format)
-
-        content_text = f"## 生成中繼資料 (Metadata)\n**風格描述:** {prompt}\n**時長:** {audio_cfg.duration or 30}秒\n"
-        if audio_cfg.bpm:
-            content_text += f"**BPM:** {audio_cfg.bpm}\n"
-        if lyrics:
-            content_text += f"\n## 歌詞\n{lyrics}"
+        seed_val = int(str(req.seed).split(",")[0].strip()) if req.seed is not None else None
+
+        # 透過 Gradio Client 橋接到 ZeroGPU 佇列
+        def _call_gradio():
+            client = Client("http://127.0.0.1:7860/")
+            return client.predict(
+                prompt=prompt,
+                lyrics=lyrics,
+                duration=audio_cfg.duration,
+                bpm=audio_cfg.bpm,
+                vocal_language=audio_cfg.vocal_language,
+                instrumental=audio_cfg.instrumental,
+                guidance_scale=req.guidance_scale,
+                seed=seed_val,
+                api_name="/generate_music"
+            )
+
+        # 非同步執行以防止 Uvicorn 死鎖
+        result_audio_path = await asyncio.to_thread(_call_gradio)
+
+        # 讀取生成的 .wav 轉換為 base64
+        with open(result_audio_path, "rb") as f:
+            audio_bytes = f.read()
+        b64 = base64.b64encode(audio_bytes).decode("utf-8")
+        audio_url = f"data:audio/wav;base64,{b64}"
+        try: os.remove(result_audio_path)
+        except: pass
+
+        content_text = f"## 生成中繼資料\n**風格:** {prompt}\n**時長:** {audio_cfg.duration or 30}s\n"
+        if lyrics: content_text += f"\n## 歌詞\n{lyrics}"
 
         response = {
-            "id": completion_id,
-            "object": "chat.completion",
-            "created": created_ts,
-            "model": MODEL_ID,
-            "choices": [{
-                "index": 0,
-                "message": {
-                    "role": "assistant",
-                    "content": content_text,
-                    "audio": [{
-                        "type": "audio_url",
-                        "audio_url": {"url": audio_url}
-                    }]
-                },
-                "finish_reason": "stop"
-            }],
-            "usage": {
-                "prompt_tokens": len(prompt.split()),
-                "completion_tokens": 100,
-                "total_tokens": len(prompt.split()) + 100
-            }
+            "id": completion_id, "object": "chat.completion", "created": created_ts, "model": MODEL_ID,
+            "choices": [{"index": 0, "message": {"role": "assistant", "content": content_text, "audio": [{"type": "audio_url", "audio_url": {"url": audio_url}}]}, "finish_reason": "stop"}],
+            "usage": {"prompt_tokens": len(prompt.split()), "completion_tokens": 100, "total_tokens": len(prompt.split()) + 100}
         }
 
-        # 處理 SSE 串流回應
         if req.stream:
             async def event_stream():
-                # 階段 1: 初始化
-                init_chunk = {"id": completion_id, "object": "chat.completion.chunk", "created": created_ts, "model": MODEL_ID, "choices": [{"index": 0, "delta": {"role": "assistant", "content": ""}, "finish_reason": None}]}
-                yield f"data: {json.dumps(init_chunk)}\n\n"
-                # 階段 2: 傳送文字資訊
-                content_chunk = {"id": completion_id, "object": "chat.completion.chunk", "created": created_ts, "model": MODEL_ID, "choices": [{"index": 0, "delta": {"content": content_text}, "finish_reason": None}]}
-                yield f"data: {json.dumps(content_chunk)}\n\n"
-                # 階段 3: 傳送音頻數據
-                audio_chunk = {"id": completion_id, "object": "chat.completion.chunk", "created": created_ts, "model": MODEL_ID, "choices": [{"index": 0, "delta": {"audio": [{"type": "audio_url", "audio_url": {"url": audio_url}}]}, "finish_reason": None}]}
-                yield f"data: {json.dumps(audio_chunk)}\n\n"
-                # 階段 4: 完成標記
-                finish_chunk = {"id": completion_id, "object": "chat.completion.chunk", "created": created_ts, "model": MODEL_ID, "choices": [{"index": 0, "delta": {}, "finish_reason": "stop"}]}
-                yield f"data: {json.dumps(finish_chunk)}\n\n"
+                chunks = [
+                    {"delta": {"role": "assistant", "content": ""}},
+                    {"delta": {"content": content_text}},
+                    {"delta": {"audio": [{"type": "audio_url", "audio_url": {"url": audio_url}}]}},
+                    {"delta": {}, "finish_reason": "stop"}
+                ]
+                for chunk in chunks:
+                    chunk_data = {"id": completion_id, "object": "chat.completion.chunk", "created": created_ts, "model": MODEL_ID, "choices": [{"index": 0, **chunk}]}
+                    yield f"data: {json.dumps(chunk_data)}\n\n"
                 yield "data: [DONE]\n\n"
-
             return StreamingResponse(event_stream(), media_type="text/event-stream")
 
         return JSONResponse(response)
 
     except Exception as e:
-        raise HTTPException(status_code=500, detail=f"伺服器內部錯誤: {str(e)}")
-
-# ─── Gradio Web UI 介面 ──────────────────────────────────
-@spaces.GPU(duration=120)
-def gradio_generate(prompt, lyrics, duration, bpm, vocal_language, instrumental, guidance_scale, seed):
-    try:
-        seed_val = int(seed) if seed else None
-        bpm_val = int(bpm) if bpm else None
-        dur_val = float(duration) if duration else 30.0
-        instr = True if instrumental == "是" else (False if instrumental == "否" else None)
-        # 對齊 API 需要的語言代碼
-        lang_map = {"英文 (en)": "en", "中文 (zh)": "zh", "日文 (ja)": "ja", "韓文 (ko)": "ko", "自動判定": "en"}
-        lang = lang_map.get(vocal_language, "en")
-
-        audio_data, sample_rate = generate_music(
-            prompt=prompt,
-            lyrics=lyrics,
-            duration=dur_val,
-            bpm=bpm_val,
-            vocal_language=lang,
-            instrumental=instr,
-            guidance_scale=float(guidance_scale),
-            seed=seed_val,
-            task_type="text2music",
-            repainting_start=0.0,
-            repainting_end=None,
-        )
-        return (sample_rate, audio_data)
-    except Exception as e:
-        raise gr.Error(f"生成失敗: {str(e)}")
-
-with gr.Blocks(
-    title="🎵 ACE-Step v1.5 音樂生成器",
-    theme=gr.themes.Soft(),
-    css="""
-    .header { text-align: center; padding: 20px; }
-    .api-info { background: #f0f4ff; padding: 15px; border-radius: 8px; margin-top: 15px; }
-    """
-) as demo:
-    gr.HTML("""
-    <div class="header">
-        <h1>🎵 ACE-Step v1.5 音樂生成器</h1>
-        <p>強大的 AI 音樂生成 · 完全相容 OpenAI API · 由 ZeroGPU 免費算力驅動</p>
-    </div>
-    """)
+        traceback.print_exc()
+        raise HTTPException(status_code=500, detail=f"內部伺服器錯誤: {str(e)}")
 
+# ─── Gradio Web UI ───────────────────────────────────────
+with gr.Blocks(title="🎵 ACE-Step v1.5 音樂生成器", theme=gr.themes.Soft()) as demo:
+    gr.HTML("<h1 style='text-align: center;'>🎵 ACE-Step v1.5 音樂生成器</h1>")
     with gr.Tab("🎼 生成音樂"):
         with gr.Row():
             with gr.Column(scale=1):
-                prompt_input = gr.Textbox(
-                    label="🏷️ 音樂風格描述 (Prompt)",
-                    placeholder="例如：節奏強烈的 EDM、包含重低音與合成器主旋律",
-                    lines=3
-                )
-                lyrics_input = gr.Textbox(
-                    label="📜 歌詞 (Lyrics，可選填)",
-                    placeholder="[Verse 1]\n這是一個美麗的早晨\n\n[Chorus]\n讓我們跟著節奏跳舞",
-                    lines=6
-                )
+                prompt_input = gr.Textbox(label="🏷️ 音樂風格描述 (Prompt)")
+                lyrics_input = gr.Textbox(label="📜 歌詞 (Lyrics，可選填)", lines=4)
                 with gr.Row():
-                    duration_input = gr.Number(label="⏱️ 生成時長 (秒)", value=30, minimum=5, maximum=240)
-                    bpm_input = gr.Number(label="🥁 BPM 節拍數 (可選)", value=None)
+                    duration_input = gr.Number(label="⏱️ 時長(秒)", value=30)
+                    bpm_input = gr.Number(label="🥁 BPM", value=None)
                 with gr.Row():
-                    lang_input = gr.Dropdown(
-                        label="🌍 人聲語言",
-                        choices=["英文 (en)", "中文 (zh)", "日文 (ja)", "韓文 (ko)", "自動判定"],
-                        value="英文 (en)"
-                    )
-                    instr_input = gr.Dropdown(
-                        label="🎸 純伴奏 (無人聲)",
-                        choices=["自動判定", "是", "否"],
-                        value="自動判定"
-                    )
+                    lang_input = gr.Dropdown(label="🌍 語言", choices=["英文 (en)", "中文 (zh)", "日文 (ja)", "韓文 (ko)"], value="英文 (en)")
+                    instr_input = gr.Dropdown(label="🎸 純伴奏", choices=["自動判定", "是", "否"], value="自動判定")
                 with gr.Row():
-                    cfg_input = gr.Slider(label="🎚️ 提示詞引導強度 (Guidance Scale)", minimum=1, maximum=15, value=7.0, step=0.5)
-                    seed_input = gr.Number(label="🎲 隨機種子 (Seed，可選)", value=None)
-
-                generate_btn = gr.Button("🚀 開始生成音樂", variant="primary", size="lg")
-
+                    cfg_input = gr.Slider(label="🎚️ Guidance Scale", minimum=1, maximum=15, value=7.0)
+                    seed_input = gr.Number(label="🎲 Seed", value=None)
+                generate_btn = gr.Button("🚀 開始生成音樂", variant="primary")
             with gr.Column(scale=1):
                 audio_output = gr.Audio(label="🎵 生成結果", type="numpy")
-                gr.HTML("""
-                <div class="api-info">
-                    <h3>📡 API 連線資訊</h3>
-                    <p>您可以使用與 OpenAI 完全相容的程式碼連接到此服務：</p>
-                    <code>POST /v1/chat/completions</code><br>
-                    <code>GET /v1/models</code><br>
-                    <code>GET /health</code>
-                </div>
-                """)
-
-        generate_btn.click(
-            fn=gradio_generate,
-            inputs=[prompt_input, lyrics_input, duration_input, bpm_input,
-                    lang_input, instr_input, cfg_input, seed_input],
-            outputs=audio_output
-        )
 
-    with gr.Tab("📖 API 使用教學"):
-        gr.Markdown("""
-## 快速整合指南
-
-此 Space 提供完全相容 OpenAI 的 API 接口。
-
-### cURL 請求範例
-```bash
-curl -X POST https://你的-SPACE-URL.hf.space/v1/chat/completions \\
-  -H "Content-Type: application/json" \\
-  -d '{
-    "messages": [{"role": "user", "content": "<prompt>Lo-fi hip hop beat</prompt>"}],
-    "audio_config": {"instrumental": true, "duration": 30}
-  }'
-```
-
-### JavaScript (前端呼叫) 範例
-```javascript
-const response = await fetch("https://你的-SPACE-URL.hf.space/v1/chat/completions", {
-  method: "POST",
-  headers: { "Content-Type": "application/json" },
-  body: JSON.stringify({
-    messages: [{ role: "user", content: "一首關於回憶的溫柔民謠" }],
-    sample_mode: true,
-    audio_config: { vocal_language: "zh", duration: 30 }
-  })
-});
-const data = await response.json();
-
-// 提取 Base64 音頻字串
-const audioUrl = data.choices[0].message.audio[0].audio_url.url;
-
-// 可以在網頁中直接播放
-const audio = new Audio(audioUrl);
-audio.play();
-```
-        """)
-
-# ─── 將 FastAPI 掛載至 Gradio ────────────────────────────
-app = gr.mount_gradio_app(fastapi_app, demo, path="/")
+        # UI 按鈕綁定
+        generate_btn.click(fn=gradio_generate, inputs=[prompt_input, lyrics_input, duration_input, bpm_input, lang_input, instr_input, cfg_input, seed_input], outputs=audio_output)
 
+        # API 專用隱藏按鈕 (供 FastAPI 呼叫)
+        api_btn = gr.Button("API", visible=False)
+        api_btn.click(fn=gradio_generate, inputs=[prompt_input, lyrics_input, duration_input, bpm_input, lang_input, instr_input, cfg_input, seed_input], outputs=audio_output, api_name="generate_music")
+
+# ─── 掛載並執行 ──────────────────────────────────────────
+app = gr.mount_gradio_app(fastapi_app, demo, path="/")
 if __name__ == "__main__":
     import uvicorn
     uvicorn.run(app, host="0.0.0.0", port=7860)

requirements.txt

@@ -1,4 +1,5 @@
 gradio>=4.44.0
+gradio_client>=1.3.0
 fastapi>=0.111.0
 uvicorn>=0.30.0
 httpx>=0.27.0