fix(runtime): harden embedding API behavior

requirements-test.txt

@@ -1,4 +1,7 @@
 # apps/embedding-runtime/requirements-test.txt
-fastapi==0.115.6
+fastapi==0.139.0
+starlette==1.3.1
 pytest==8.3.4
 httpx==0.27.2
+ruff==0.15.19
+ty==0.0.53

requirements.txt

@@ -1,8 +1,9 @@
 # apps/embedding-runtime/requirements.txt
-fastapi==0.115.6
-uvicorn[standard]==0.34.0
-sentence-transformers==3.3.1
-transformers==4.47.1
-torch==2.5.1
-peft==0.14.0
-numpy==2.2.1
+fastapi==0.139.0
+starlette==1.3.1
+uvicorn[standard]==0.51.0
+sentence-transformers==5.6.0
+transformers==5.13.1
+torch==2.13.0
+peft==0.19.1
+numpy==2.5.1

server.py

@@ -53,7 +53,7 @@ _active_adapter: Optional[str] = None
 def _load_model() -> SentenceTransformer:
     global _model
     if _model is None:
-        _model = SentenceTransformer(MODEL_NAME, device=DEVICE, trust_remote_code=True)
+        _model = SentenceTransformer(MODEL_NAME, device=DEVICE, trust_remote_code=False)
     return _model
 
 
@@ -71,11 +71,13 @@ def _apply_adapter(task: Optional[str]) -> None:
         return
     model = _load_model()
     inner = model[0].auto_model  # transformers model inside the ST wrapper
-    if hasattr(inner, "load_adapter"):
+    load_adapter = getattr(inner, "load_adapter", None)
+    if callable(load_adapter):
         try:
-            inner.load_adapter(adapter, adapter_name=task)
-            if hasattr(inner, "set_adapter"):
-                inner.set_adapter(task)
+            load_adapter(adapter, adapter_name=task)
+            set_adapter = getattr(inner, "set_adapter", None)
+            if callable(set_adapter):
+                set_adapter(task)
             _active_adapter = adapter
         except Exception as exc:  # noqa: BLE001 — surface a clean 400 to the caller
             raise HTTPException(status_code=400, detail=f"adapter load failed for task={task!r}: {exc}")

tests/test_server.py

@@ -69,24 +69,24 @@ class FakeCuda:
 
 def build_fake_torch_module() -> ModuleType:
     module = ModuleType("torch")
-    module.cuda = FakeCuda()
+    setattr(module, "cuda", FakeCuda())
     return module
 
 
 def build_fake_sentence_transformers_module() -> ModuleType:
     module = ModuleType("sentence_transformers")
-    module.SentenceTransformer = FailingSentenceTransformer
+    setattr(module, "SentenceTransformer", FailingSentenceTransformer)
     return module
 
 
 def build_fake_numpy_module() -> ModuleType:
     module = ModuleType("numpy")
-    module.ndarray = FakeMatrix
+    setattr(module, "ndarray", FakeMatrix)
 
     def argsort(scores: FakeScores) -> list[int]:
         return sorted(range(len(scores.values)), key=lambda index: scores.values[index])
 
-    module.argsort = argsort
+    setattr(module, "argsort", argsort)
     return module
 
 
@@ -100,6 +100,26 @@ def server_module(monkeypatch: pytest.MonkeyPatch) -> ModuleType:
     return importlib.import_module("server")
 
 
+def test_model_loading_disables_remote_code_TASK_8c19d6a7(
+    server_module: ModuleType, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    # Given: a recording model constructor at the real loader boundary.
+    constructor_calls: list[tuple[str, str, bool]] = []
+
+    class RecordingSentenceTransformer:
+        def __init__(self, model_name: str, *, device: str, trust_remote_code: bool) -> None:
+            constructor_calls.append((model_name, device, trust_remote_code))
+
+    monkeypatch.setattr(server_module, "SentenceTransformer", RecordingSentenceTransformer)
+    setattr(server_module, "_model", None)
+
+    # When: the runtime lazily loads its configured model.
+    server_module._load_model()
+
+    # Then: repository-supplied model code is never executed.
+    assert constructor_calls == [(server_module.MODEL_NAME, server_module.DEVICE, False)]
+
+
 def test_health_remains_available_without_loading_model_SPEC_KSERVE_002_SC_KSERVE_002(
     server_module: ModuleType,
 ) -> None: