fix: get_trainable_info 检查 build_model 是否为 None，避免 103 个算法全显示为可训练

- registry.py: hasattr 改为 getattr(...) is None 判断，因 ModelAlgorithmAdapter.build_model property 对所有实例都返回非 None
- _torch_upgrade.py: 处理 checkpoint 格式异常时跳过 torch 推理

Author: jinyiwei2012

algorithms/models/deep_learning/_torch_upgrade.py

@@ -588,6 +588,11 @@ def try_torch_predict(
         model = getattr(algorithm, "_cached_torch_model", None)
         if model is None or (bvid and not getattr(algorithm, "_cached_bvid", "") == bvid):
             model = model_cls(**(model_kwargs or {}))
+            if isinstance(state, (tuple, list)):
+                state = state[0]
+            if not isinstance(state, dict):
+                logger.warning("[%s] checkpoint 格式异常 (type=%s)，跳过 torch 推理", algo_id, type(state).__name__)
+                return fallback_fn(video_data, threshold)
             model.load_state_dict(state)
             model.to(algorithm._device).eval()
             algorithm._cached_torch_model = model

algorithms/registry.py

@@ -256,5 +256,38 @@ def reset(cls):
         cls._model_adapters = {}
         cls._initialized = False
 
+    @classmethod
+    def get_trainable_info(cls) -> List[Dict]:
+        from algorithms.training.checkpoint_manager import CheckpointManager
+        if not cls._initialized:
+            cls.initialize()
+        result = []
+        for aid, adapter in cls._algorithms.items():
+            build_model_fn = getattr(adapter, "build_model", None)
+            if build_model_fn is None:
+                continue
+            ckpt = CheckpointManager(aid)
+            active = ckpt.active_version()
+            result.append({
+                "algorithm_id": aid,
+                "name": getattr(adapter, "name", aid),
+                "category": getattr(adapter, "category", ""),
+                "has_ckpt": ckpt.has_checkpoint(),
+                "active_version": active or "",
+            })
+        return result
+
+    @classmethod
+    def get_trainable_algorithms(cls) -> List:
+        if not cls._initialized:
+            cls.initialize()
+        result = []
+        for aid, adapter in cls._algorithms.items():
+            build_model_fn = getattr(adapter, "build_model", None)
+            if build_model_fn is None:
+                continue
+            result.append((aid, adapter.algo if hasattr(adapter, "algo") else adapter, adapter))
+        return result
+
 
 AlgorithmRegistry.initialize()