fix(backend): 部署韧性——模型自愈/就绪门禁/全局代理/启动诊断

- whisper: model.bin 截断/损坏时删目录重下重试一次,修「Unable to
  open file model.bin」死循环;mlx 同样按 config.json 判完整性
- /generate_note 加就绪门禁:本地转写引擎模型没下好直接拦截,返回
  reason=transcriber_model_not_ready,不让任务静默卡在首次下载
- 全局代理:新增 ProxyConfigManager(JSON 配置 + HTTP_PROXY env 兜底)
  + build_openai_client,统一注入代理到 LLM/Groq 客户端;yt-dlp 与
  youtube-transcript-api 也走代理
- build_openai_client 校验 api_key 非空,空 key 给「xxx 的 API Key
  未配置」而不是天书般的 Illegal header value b'Bearer '
- universal_gpt: 模型拒绝自定义 temperature(o1/o3/gpt-5 系列)时
  就地去掉参数重试,不消耗重试预算
- connect_test 改用真实 chat completion 而非 /v1/models 探测
- main.py: lifespan 拆 [startup 1/5..5/5] 分段日志 + 异常清晰定位
- /sys_health 重构为结构化返回 {backend,ffmpeg,db,whisper_model}

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
huangjianwu
2026-05-14 19:01:14 +08:00
co-authored by Claude Opus 4.7
parent 88d25f8cc1
commit 41f17592c2
16 changed files with 534 additions and 92 deletions
+33 -13
View File
@@ -10,6 +10,7 @@ from app.utils.path_helper import get_model_dir
from events import transcription_finished
from pathlib import Path
import os
import shutil
from tqdm import tqdm
from modelscope import snapshot_download
@@ -50,22 +51,41 @@ class WhisperTranscriber(Transcriber):
model_dir = get_model_dir("whisper")
model_path = os.path.join(model_dir, f"whisper-{model_size}")
if not Path(model_path).exists():
logger.info(f"模型 whisper-{model_size} 不存在,开始下载...")
repo_id = MODEL_MAP[model_size]
model_path = snapshot_download(
repo_id,
repo_id = MODEL_MAP[model_size]
local_dir=model_path,
)
# 第一步:目录 / model.bin 不在 → 下载。
# 关键判据用 model.bin 而不是目录存在:首次下载若被打断(网络中断 / 磁盘满 /
# 容器被 kill)会留下半成品目录,只看目录存在会跳过下载。
model_bin = Path(model_path) / "model.bin"
if not model_bin.exists():
if Path(model_path).exists():
logger.warning(f"模型目录 {model_path} 存在但 model.bin 缺失(上次下载未完成),重新下载")
else:
logger.info(f"模型 whisper-{model_size} 不存在,开始下载...")
model_path = snapshot_download(repo_id, local_dir=model_path)
logger.info("模型下载完成")
self.model = WhisperModel(
model_size_or_path=model_path,
device=self.device,
compute_type=self.compute_type,
download_root=model_dir
)
# 第二步:加载。model.bin 可能存在但【内容截断】(下载到一半被 kill),
# 此时 WhisperModel() 会抛 "File model.bin is incomplete: failed to read a buffer..."。
# 捕获后删掉损坏目录、重新下载、再试一次——自愈,避免 500 死循环。
try:
self.model = WhisperModel(
model_size_or_path=model_path,
device=self.device,
compute_type=self.compute_type,
download_root=model_dir,
)
except Exception as e:
logger.warning(f"加载 whisper-{model_size} 失败(疑似模型文件损坏 / 截断):{e};删除后重新下载")
shutil.rmtree(model_path, ignore_errors=True)
model_path = snapshot_download(repo_id, local_dir=model_path)
logger.info("模型重新下载完成,重试加载")
self.model = WhisperModel(
model_size_or_path=model_path,
device=self.device,
compute_type=self.compute_type,
download_root=model_dir,
)
@staticmethod
def is_torch_installed() -> bool:
try: