Add Whisper offline loading for air-gapped servers.

Pre-download via HF mirror scripts so inner-network deploys avoid Hub Network is unreachable errors. Co-authored-by: Cursor <cursoragent@cursor.com>
2026-06-12 16:11:57 +08:00
parent 39e29fe6a9
commit 0f5277c22e
6 changed files with 310 additions and 164 deletions
@@ -10,4 +10,6 @@ OLLAMA_PORT=11434

 # ChatTTS 模型目录（预下载脚本写入）
 # CHATTTS_MODEL_DIR=/opt/Trading_Studio/models/ChatTTS
+# WHISPER_MODEL_DIR=/opt/Trading_Studio/models/whisper
+# WHISPER_MODEL_SIZE=small
 # HF_ENDPOINT=https://hf-mirror.com
@@ -361,9 +361,20 @@ cd /opt/Trading_Studio
 pip install -r requirements.txt
 ```

-### 6.1 Faster-Whisper
+### 6.1 Faster-Whisper（必须预下载）

-随 `requirements.txt` 安装。首次运行会自动下载 `small` 模型（约 500MB）至 HuggingFace 缓存。
+随 `requirements.txt` 安装。**内网服务器无法访问 HuggingFace 时会报 `Network is unreachable`。**
+
+部署后执行（推荐与 ChatTTS 一起）：
+
+```bash
+cd /opt/Trading_Studio
+bash scripts/download_all_models.sh
+# 或仅 Whisper: bash scripts/download_whisper_models.sh small
+pm2 restart trading_studio
+```
+
+模型目录：`/opt/Trading_Studio/models/whisper/small/`（含 `model.bin`）。

 ### 6.2 ChatTTS（必须预下载，勿依赖 GitHub）

@@ -76,23 +76,33 @@ SYSTEM_PROMPT = (
    "去掉所有无意义的口头禅，字数不做删减。"
 )

+# ---------------------------------------------------------------------------
+# 路径
+# ---------------------------------------------------------------------------
+BASE_DIR = Path(__file__).resolve().parent
+INSTALL_DIR = Path("/opt/Trading_Studio")
+
 # ---------------------------------------------------------------------------
 # Faster-Whisper 配置
 # ---------------------------------------------------------------------------
-WHISPER_MODEL_SIZE = "small"
+WHISPER_MODEL_SIZE = _env_str("WHISPER_MODEL_SIZE", "small")
 WHISPER_DEVICE = "cuda"
 WHISPER_COMPUTE_TYPE = "float16"
 WHISPER_LANGUAGE = "zh"
+WHISPER_MODEL_DIR = Path(_env_str("WHISPER_MODEL_DIR", str(BASE_DIR / "models" / "whisper")))
+
+WHISPER_HF_REPO = {
+    "tiny": "Systran/faster-whisper-tiny",
+    "base": "Systran/faster-whisper-base",
+    "small": "Systran/faster-whisper-small",
+    "medium": "Systran/faster-whisper-medium",
+    "large-v2": "Systran/faster-whisper-large-v2",
+    "large-v3": "Systran/faster-whisper-large-v3",
+}

 # ---------------------------------------------------------------------------
 # ChatTTS 配置
 # ---------------------------------------------------------------------------
-# 标准生产安装路径（/opt，root 部署）
-INSTALL_DIR = Path("/opt/Trading_Studio")
-
-# 项目根目录（开发/生产均自适应，以实际 app.py 所在目录为准）
-BASE_DIR = Path(__file__).resolve().parent
-
 # 固定音色 Embedding 存储路径
 SPEAKER_EMB_PATH = BASE_DIR / "speaker_emb.pt"

@@ -0,0 +1,15 @@
+#!/usr/bin/env bash
+# 一次性下载 Whisper + ChatTTS 全部模型（内网服务器部署必跑）
+set -euo pipefail
+ROOT="$(cd "$(dirname "$0")/.." && pwd)"
+cd "${ROOT}"
+
+echo "========== 下载 Whisper (small) =========="
+bash scripts/download_whisper_models.sh small
+
+echo ""
+echo "========== 下载 ChatTTS =========="
+bash scripts/download_chattts_models.sh
+
+echo ""
+echo "[OK] 全部模型下载完成，请: pm2 restart trading_studio"
@@ -0,0 +1,59 @@
+#!/usr/bin/env bash
+# 预下载 Faster-Whisper 模型（HF 镜像，内网服务器离线可用）
+# 用法: bash scripts/download_whisper_models.sh [tiny|base|small|medium|large-v3]
+set -euo pipefail
+
+ROOT="$(cd "$(dirname "$0")/.." && pwd)"
+SIZE="${1:-small}"
+VENV_PY="${ROOT}/venv/bin/python"
+MODEL_DIR="${WHISPER_MODEL_DIR:-${ROOT}/models/whisper}/${SIZE}"
+
+export HF_ENDPOINT="${HF_ENDPOINT:-https://hf-mirror.com}"
+export HF_HUB_DOWNLOAD_TIMEOUT="${HF_HUB_DOWNLOAD_TIMEOUT:-600}"
+export HF_HOME="${HF_HOME:-${ROOT}/models/hf_cache}"
+export MODEL_DIR
+export WHISPER_SIZE="${SIZE}"
+
+echo "[INFO] Whisper 模型: ${SIZE}"
+echo "[INFO] 保存目录: ${MODEL_DIR}"
+echo "[INFO] HF 镜像: ${HF_ENDPOINT}"
+
+if [[ ! -x "${VENV_PY}" ]]; then
+  echo "[ERROR] 未找到 venv，请先 bash deploy.sh deps"
+  exit 1
+fi
+
+"${VENV_PY}" -m pip install -q huggingface_hub
+
+"${VENV_PY}" << 'PY'
+import os
+from pathlib import Path
+from huggingface_hub import snapshot_download
+
+size = os.environ["WHISPER_SIZE"]
+repos = {
+    "tiny": "Systran/faster-whisper-tiny",
+    "base": "Systran/faster-whisper-base",
+    "small": "Systran/faster-whisper-small",
+    "medium": "Systran/faster-whisper-medium",
+    "large-v2": "Systran/faster-whisper-large-v2",
+    "large-v3": "Systran/faster-whisper-large-v3",
+}
+repo = repos.get(size)
+if not repo:
+    raise SystemExit(f"未知模型尺寸: {size}, 可选: {list(repos)}")
+
+target = Path(os.environ["MODEL_DIR"])
+target.mkdir(parents=True, exist_ok=True)
+
+print(f"[INFO] 正在下载 {repo} ...")
+snapshot_download(repo_id=repo, local_dir=str(target), local_dir_use_symlinks=False)
+
+if not (target / "model.bin").is_file():
+    raise SystemExit(f"[ERROR] 下载不完整，未找到 model.bin: {target}")
+
+print(f"[OK] Whisper 模型就绪: {target}")
+PY
+
+echo ""
+echo "[OK] 请执行: pm2 restart trading_studio"
@@ -6,43 +6,93 @@ Faster-Whisper CUDA 语音识别服务
 from __future__ import annotations

 import logging
+import os
 import traceback
+from pathlib import Path
 from typing import Optional, Tuple

 from config import (
+    BASE_DIR,
+    HF_ENDPOINT,
+    HF_HOME,
+    HF_HUB_DOWNLOAD_TIMEOUT,
    WHISPER_COMPUTE_TYPE,
    WHISPER_DEVICE,
+    WHISPER_HF_REPO,
    WHISPER_LANGUAGE,
+    WHISPER_MODEL_DIR,
    WHISPER_MODEL_SIZE,
 )

 logger = logging.getLogger(__name__)

-# 全局懒加载模型实例，避免 Gradio 重复初始化占用显存
 _model = None
 _model_error: Optional[str] = None


+def _ensure_hf_env() -> None:
+    os.environ.setdefault("HF_ENDPOINT", HF_ENDPOINT)
+    os.environ.setdefault("HF_HUB_DOWNLOAD_TIMEOUT", str(HF_HUB_DOWNLOAD_TIMEOUT))
+    os.environ.setdefault("HF_HOME", str(HF_HOME))
+    WHISPER_MODEL_DIR.mkdir(parents=True, exist_ok=True)
+
+
+def _whisper_local_path() -> Optional[Path]:
+    """返回已预下载的本地模型目录。"""
+    local = WHISPER_MODEL_DIR / WHISPER_MODEL_SIZE
+    if (local / "model.bin").is_file():
+        return local
+    return None
+
+
 def _is_cuda_error(exc: BaseException) -> bool:
-    """判断异常是否与 CUDA/GPU 相关。"""
    msg = str(exc).lower()
    cuda_keywords = (
-        "cuda",
-        "cudnn",
-        "cublas",
-        "gpu",
-        "out of memory",
-        "no kernel image",
-        "device-side assert",
+        "cuda", "cudnn", "cublas", "gpu",
+        "out of memory", "no kernel image", "device-side assert",
    )
    return any(k in msg for k in cuda_keywords)


+def _is_network_error(exc: BaseException) -> bool:
+    msg = str(exc).lower()
+    return any(
+        k in msg
+        for k in (
+            "network is unreachable",
+            "connection error",
+            "connecterror",
+            "timed out",
+            "couldn't connect",
+            "name resolution",
+            "hub",
+        )
+    )
+
+
+def _build_load_error(exc: BaseException) -> str:
+    lines = [
+        "Whisper 模型加载失败。",
+        f"详情: {exc}",
+        "",
+    ]
+    if _is_network_error(exc):
+        lines.extend([
+            "原因：服务器无法访问 HuggingFace 下载模型（内网/无外网常见）。",
+            "请在服务器执行（走 HF 镜像，仅需一次）：",
+            f"  cd {BASE_DIR}",
+            "  bash scripts/download_whisper_models.sh",
+            "  pm2 restart trading_studio",
+            "",
+            f"模型将保存到: {WHISPER_MODEL_DIR / WHISPER_MODEL_SIZE}",
+        ])
+    else:
+        lines.append(f"完整日志:\n{traceback.format_exc()}")
+    return "\n".join(lines)
+
+
 def get_whisper_model():
-    """
-    获取或初始化 Faster-Whisper 模型。
-    强制 device=cuda, compute_type=float16。
-    """
+    """获取或初始化 Faster-Whisper 模型（优先本地预下载）。"""
    global _model, _model_error

    if _model is not None:
@@ -52,18 +102,31 @@ def get_whisper_model():
        return None, _model_error

    try:
+        _ensure_hf_env()
        from faster_whisper import WhisperModel

+        local = _whisper_local_path()
+        if local:
+            model_id = str(local)
+            logger.info("Whisper 从本地加载: %s", model_id)
+        else:
+            model_id = WHISPER_MODEL_SIZE
+            logger.warning(
+                "未找到本地 Whisper 模型 (%s)，尝试在线下载（可能失败）…",
+                WHISPER_MODEL_DIR / WHISPER_MODEL_SIZE,
+            )
+
        logger.info(
-            "正在加载 Whisper 模型: size=%s, device=%s, compute_type=%s",
-            WHISPER_MODEL_SIZE,
+            "Whisper 加载: model=%s, device=%s, compute_type=%s",
+            model_id,
            WHISPER_DEVICE,
            WHISPER_COMPUTE_TYPE,
        )
        _model = WhisperModel(
-            WHISPER_MODEL_SIZE,
+            model_id,
            device=WHISPER_DEVICE,
            compute_type=WHISPER_COMPUTE_TYPE,
+            download_root=str(WHISPER_MODEL_DIR),
        )
        logger.info("Whisper 模型加载成功。")
        return _model, None
@@ -79,26 +142,17 @@ def get_whisper_model():
    except Exception as exc:
        if _is_cuda_error(exc):
            _model_error = (
-                "CUDA 初始化失败，请检查 NVIDIA 驱动、CUDA 运行时及 cuDNN 是否正确安装。\n"
-                f"错误详情: {exc}\n"
-                f"{traceback.format_exc()}"
+                "CUDA 初始化失败，请检查 NVIDIA 驱动、CUDA 运行时及 cuDNN。\n"
+                f"错误详情: {exc}"
            )
        else:
-            _model_error = f"Whisper 模型加载失败: {exc}\n{traceback.format_exc()}"
+            _model_error = _build_load_error(exc)
        logger.exception("Whisper 模型加载异常")
        return None, _model_error


 def transcribe_audio(audio_path: str) -> Tuple[bool, str]:
-    """
-    将音频文件转写为中文文本。
-
-    Args:
-        audio_path: 本地音频文件绝对或相对路径
-
-    Returns:
-        (success, text_or_error_message)
-    """
+    """将音频文件转写为中文文本。"""
    if not audio_path:
        return False, "未提供音频文件路径。"

@@ -114,10 +168,7 @@ def transcribe_audio(audio_path: str) -> Tuple[bool, str]:
            vad_filter=True,
        )

-        text_parts = []
-        for segment in segments:
-            text_parts.append(segment.text.strip())
-
+        text_parts = [segment.text.strip() for segment in segments]
        result_text = "".join(text_parts).strip()

        if not result_text:
@@ -137,8 +188,7 @@ def transcribe_audio(audio_path: str) -> Tuple[bool, str]:
    except Exception as exc:
        if _is_cuda_error(exc):
            err = (
-                "CUDA 推理异常：显存可能不足或 GPU 状态异常。"
-                "建议关闭其他占用显存的进程后重试。\n"
+                "CUDA 推理异常：显存可能不足或 GPU 状态异常。\n"
                f"错误详情: {exc}"
            )
        else:
@@ -149,7 +199,6 @@ def transcribe_audio(audio_path: str) -> Tuple[bool, str]:


 def reset_whisper_model() -> None:
-    """释放模型引用（用于调试或显存回收）。"""
    global _model, _model_error
    _model = None
    _model_error = None