feat: enhance GPU support detection and user feedback in the engine interface

server/api.py

@@ -24,6 +24,8 @@ from .core_bridge import (
     ProviderError,
     build_provider,
     detect_architecture,
+    engine_warning,
+    gpu_offload_supported,
     list_provider_names,
     load_engine_class,
     load_settings,
@@ -87,10 +89,16 @@ async def get_system() -> SystemInfo:
     except EngineLoadError as exc:
         runtime_name = f"unavailable ({exc})"
 
+    # Computed after load_engine_class above, which already imports llama_cpp,
+    # so the probe costs nothing extra here.
+    offload = gpu_offload_supported()
+
     return SystemInfo(
         version=__version__,
         engine_architecture=descriptor.architecture,
         engine_runtime=runtime_name,
+        engine_gpu_offload=offload,
+        engine_warning=engine_warning(descriptor, offload=offload),
         template_ok=TEMPLATE_PATH.exists(),
         python_version=platform.python_version(),
         metrics={

server/core_bridge.py

@@ -30,6 +30,8 @@ from config import (  # type: ignore[import-not-found]
 from engine_loader import (  # type: ignore[import-not-found]
     EngineLoadError,
     detect_architecture,
+    engine_warning,
+    gpu_offload_supported,
     load_engine_class,
 )
 from prompts import load_test_prompts  # type: ignore[import-not-found]
@@ -77,6 +79,8 @@ __all__ = [
     "DEFAULT_PROVIDER_NAME",
     "EngineLoadError",
     "detect_architecture",
+    "engine_warning",
+    "gpu_offload_supported",
     "load_engine_class",
     "DEFAULT_TEMPERATURE",
     "MODELS_DIR",

server/schemas.py

@@ -204,6 +204,13 @@ class SystemInfo(BaseModel):
     version: str
     engine_architecture: str
     engine_runtime: str
+    # None when llama-cpp-python is absent or too old to report it, which is a
+    # different thing from a confirmed False.
+    engine_gpu_offload: Optional[bool] = None
+    # Set when the detected hardware and the installed build disagree — the
+    # case where the interface would otherwise claim "cuda" while every token
+    # is generated on the CPU.
+    engine_warning: Optional[str] = None
     template_ok: bool
     python_version: str
     metrics: Dict[str, str] = {}