refactor(ai): 统一 ONNX 推理后端选择并新增 OpenVINO 支持

将分散在 face/embedding/image_classification/ticket 四个 service 的硬编码
provider 列表收敛到新增的 onnx_providers.get_onnx_providers()，按
CUDA > OpenVINO > CPU 优先级动态探测可用 EP，避免未安装 EP 报警告。

- 新增 openvino extra（onnxruntime-openvino + openvino 2025.4.1），与
  cpu/gpu 互斥；uv.lock 同步更新
- _openvino_device() 按 openvino.Core().available_devices() 探测：NPU 优先、
  否则 CPU，避免无 NPU 机器因 device_type=NPU 在会话创建时硬挂；OPENVINO_DEVICE
  环境变量保留为强制覆盖口
- image_classification_service 补回 provider_options，与其他 service 一致
- get_onnx_providers() 加 lru_cache，进程内只探测与打日志一次
- ocr_service 注明 OCR 走 RapidOCR 自有引擎口，不接入 OpenVINO EP

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: LC044

package/ai/app/services/embedding_service.py

@@ -10,6 +10,7 @@
 from app.services.model_downloader import model_downloader
 from app.services.model_manager import model_manager
 from app.services.ai_config_manager import ai_config_manager
+from app.services.onnx_providers import get_onnx_providers
 
 class ONNXCLIPTextWrapper:
     def __init__(self, model_dir):
@@ -22,7 +23,7 @@ def __init__(self, model_dir):
 
         self.tokenizer = AutoTokenizer.from_pretrained(model_dir)
 
-        providers = ['CUDAExecutionProvider', 'CPUExecutionProvider']
+        providers, _ = get_onnx_providers()
         text_model_path = os.path.join(model_dir, "textual.onnx")
         self.text_session = ort.InferenceSession(text_model_path, providers=providers)
 
@@ -52,8 +53,8 @@ def __init__(self, model_dir):
 
         self.processor = AutoImageProcessor.from_pretrained(model_dir)
 
-        # CUDA 优先：onnxruntime-gpu 在 CUDA 运行时库可用时走 GPU，否则回退 CPU。
-        providers = ['CUDAExecutionProvider', 'CPUExecutionProvider']
+        # 推理后端按 CUDA -> OpenVINO -> CPU 自动选择。
+        providers, _ = get_onnx_providers()
         vision_model_path = os.path.join(model_dir, "visual.onnx")
         self.vision_session = ort.InferenceSession(vision_model_path, providers=providers)
 

package/ai/app/services/face_service.py

@@ -9,23 +9,23 @@
 from app.services.model_downloader import model_downloader
 from app.services.model_manager import model_manager
 from app.services.ai_config_manager import ai_config_manager
+from app.services.onnx_providers import get_onnx_providers
 
 def load_insightface_model():
     try:
         import insightface
         from insightface.app import FaceAnalysis
-        
+
         model_name = ai_config_manager.get_model_selection("face")
         logging.info(f"Initializing InsightFace with model: {model_name}")
 
-        # Initialize InsightFace analysis
-        # providers=['CUDAExecutionProvider', 'CPUExecutionProvider'] if GPU available
-        provider_options = [{"device_id": 0}, {}]
+        # 推理后端按 CUDA -> OpenVINO -> CPU 自动选择，与安装的 extra (gpu/openvino/cpu) 对齐。
+        providers, provider_options = get_onnx_providers()
         model_path = settings.MODEL_PATH.rstrip("/").rstrip("models")
         app = FaceAnalysis(
             name=model_name, root=model_path,
-            providers=['CUDAExecutionProvider', 'CPUExecutionProvider'],
-            provider_options = provider_options  # 传递 CUDA 配置
+            providers=providers,
+            provider_options=provider_options,
         )
         app.prepare(ctx_id=0, det_size=(640, 640))
         logging.info("InsightFace model initialized successfully.")

package/ai/app/services/image_classification_service.py

@@ -12,12 +12,14 @@
 from app.config import settings
 from app.services.model_downloader import model_downloader
 from app.services.model_manager import model_manager
+from app.services.onnx_providers import get_onnx_providers
 
 class ONNXModelWrapper:
     def __init__(self, model_path):
         import onnxruntime as ort
         import ast
-        self.session = ort.InferenceSession(model_path, providers=['CUDAExecutionProvider', 'CPUExecutionProvider'])
+        providers, provider_options = get_onnx_providers()
+        self.session = ort.InferenceSession(model_path, providers=providers, provider_options=provider_options)
         meta = self.session.get_modelmeta()
         names_str = meta.custom_metadata_map.get('names', '{}')
         try:

package/ai/app/services/ocr_service.py

@@ -15,6 +15,8 @@ def load_paddleocr_model():
         # GPU 优先：ONNXRUNTIME 引擎开启 use_cuda，RapidOCR 会把 CUDAExecutionProvider
         # 插到 provider 列表首位；CUDA 运行时库不可用时自动回退 CPU（onnxruntime.get_device()
         # 非 GPU 即跳过，不会报错）。安装了 torch 时仍走下面的 TORCH 引擎 GPU 路径。
+        # 注意：OCR 走 RapidOCR 自己的引擎配置口，不经过 onnx_providers.get_onnx_providers()，
+        # 因此 openvino extra 不会让 OCR 走 OpenVINO EP（仅人脸 / 分类 / 票据 / embedding 受影响）。
         params = {
             "Det.engine_type": EngineType.ONNXRUNTIME,
             "Det.lang_type": LangDet.CH,

package/ai/app/services/onnx_providers.py

@@ -0,0 +1,64 @@
+import functools
+import logging
+import os
+
+logger = logging.getLogger("app.services.onnx_providers")
+
+# 推理后端优先级：CUDA (gpu extra) > OpenVINO (openvino extra) > CPU (cpu extra)。
+# 通过 ort.get_available_providers() 动态探测，避免硬编码导致未安装的 EP 报警告。
+_PRIORITY = ["CUDAExecutionProvider", "OpenVINOExecutionProvider", "CPUExecutionProvider"]
+
+
+def _openvino_device() -> str:
+    """选择 OpenVINO EP 的 device_type：NPU 优先，否则 CPU。
+
+    OPENVINO_DEVICE 环境变量可强制覆盖（设为 NPU / GPU / CPU）；留空时按
+    openvino.Core().available_devices() 探测——有 NPU 走 NPU，否则回退 CPU，
+    避免在无 NPU 的机器上指定 NPU 导致会话创建失败。
+    """
+    override = os.getenv("OPENVINO_DEVICE", "").strip()
+    if override:
+        return override
+    try:
+        from openvino import Core
+        devices = set(Core().available_devices)
+        if "NPU" in devices:
+            return "NPU"
+    except Exception as e:
+        logger.warning(f"Failed to probe OpenVINO devices, fallback to CPU: {e}")
+    return "CPU"
+
+
+@functools.lru_cache(maxsize=1)
+def get_onnx_providers():
+    """返回 (providers, provider_options)，按可用性筛选并保持优先级。
+
+    结果在进程生命周期内缓存（可用 EP 不会变），避免每次加载模型都重复探测与打日志。
+
+    provider_options 与 providers 一一对齐：
+      - CUDAExecutionProvider     -> {"device_id": 0}
+      - OpenVINOExecutionProvider -> {"device_type": "NPU" | "CPU"}（NPU 优先）
+      - CPUExecutionProvider      -> {}
+    """
+    try:
+        import onnxruntime as ort
+        available = set(ort.get_available_providers())
+    except Exception as e:
+        logger.warning(f"Failed to probe onnxruntime providers, fallback to CPU: {e}")
+        return ["CPUExecutionProvider"], [{}]
+
+    providers = [p for p in _PRIORITY if p in available]
+    if not providers:
+        providers = ["CPUExecutionProvider"]
+
+    options = []
+    for p in providers:
+        if p == "CUDAExecutionProvider":
+            options.append({"device_id": 0})
+        elif p == "OpenVINOExecutionProvider":
+            options.append({"device_type": _openvino_device()})
+        else:
+            options.append({})
+
+    logger.info(f"ONNX Runtime providers selected: {providers} with options {options}")
+    return providers, options

package/ai/app/services/ticket_service.py

@@ -10,6 +10,7 @@
 from app.services.ai_config_manager import ai_config_manager
 from app.services.ticket_parser import parse_ticket_info, extract_text
 from app.services.fly_ticket_parser import extract_flight_info
+from app.services.onnx_providers import get_onnx_providers
 
 def load_modelscope_model():
     """
@@ -36,7 +37,7 @@ def load_onnx_model():
         model_path = os.path.join(model_dir, "ticket-recognition.onnx")
 
         # Initialize ONNX model
-        providers = ['CPUExecutionProvider', 'CUDAExecutionProvider']
+        providers, _ = get_onnx_providers()
         session = ort.InferenceSession(model_path, providers=providers)
 
         # Set names attribute manually since ONNX model might not have it in the same format

package/ai/pyproject.toml

@@ -15,16 +15,21 @@ dependencies = [
   "rapidocr>=3.0.0",
   "uvicorn>=0.27.0",
   "httpx>=0.27.0",
+  "transformers>=4.37.0"
 ]
 
 [project.optional-dependencies]
 cpu = [
-  "onnxruntime>=1.23.2",
-  "transformers>=4.37.0"
+  "onnxruntime>=1.23.2"
+]
+
+openvino = [
+  "onnxruntime-openvino>=1.24.1",
+  "openvino==2025.4.1"
 ]
+
 gpu = [
-  "onnxruntime-gpu>=1.23.2",
-  "transformers>=4.37.0"
+  "onnxruntime-gpu>=1.23.2"
 ]
 
 [dependency-groups]
@@ -37,6 +42,7 @@ dev = [
 conflicts = [
   [
     { extra = "cpu" },
+    { extra = "openvino" },
     { extra = "gpu" },
   ],
 ]