Add CUDA health checking to /healthz endpoint

inference/core/interfaces/http/http_api.py

@@ -2200,8 +2200,25 @@ def readiness(
 
         @app.get("/healthz", status_code=200)
         def healthz():
-            """Health endpoint for Kubernetes liveness probe."""
-            return {"status": "healthy"}
+            """Health endpoint for Kubernetes liveness probe.
+
+            Verifies CUDA context health when running on GPU. Returns 503 if
+            CUDA is corrupted (unrecoverable - requires process restart).
+            """
+            from inference.core.utils.cuda_health import check_cuda_health
+
+            is_healthy, error = check_cuda_health()
+            if is_healthy:
+                return {"status": "healthy"}
+            else:
+                logger.error("CUDA health check failed: %s", error)
+                return JSONResponse(
+                    content={
+                        "status": "unhealthy",
+                        "reason": "cuda_error",
+                    },
+                    status_code=503,
+                )
 
         if CORE_MODELS_ENABLED:
             if CORE_MODEL_CLIP_ENABLED:

inference/core/utils/cuda_health.py

@@ -0,0 +1,110 @@
+"""CUDA health checking utilities.
+
+Provides a fast, cached health check for GPU/CUDA state. Once CUDA fails,
+the context is permanently corrupted and cannot recover without process restart.
+The failure state is cached to avoid repeatedly calling into a broken CUDA runtime.
+"""
+
+import logging
+import threading
+import time
+from typing import Optional, Tuple
+
+logger = logging.getLogger(__name__)
+
+
+class CudaHealthChecker:
+    """Thread-safe CUDA health checker with failure caching.
+
+    Once a CUDA failure is detected, the result is cached permanently
+    (CUDA context corruption is unrecoverable). Subsequent calls return
+    the cached failure immediately without touching CUDA.
+    """
+
+    def __init__(self):
+        self._lock = threading.Lock()
+        self._cuda_failed: bool = False
+        self._failure_error: Optional[str] = None
+        self._failure_time: Optional[float] = None
+        self._gpu_available: Optional[bool] = None  # None = not yet checked
+
+    def _is_gpu_environment(self) -> bool:
+        """Check if we're running in a GPU environment. Cached after first call."""
+        if self._gpu_available is not None:
+            return self._gpu_available
+        try:
+            import torch
+
+            self._gpu_available = torch.cuda.is_available()
+        except ImportError:
+            self._gpu_available = False
+        except Exception:
+            self._gpu_available = False
+        return self._gpu_available
+
+    def check_health(self) -> Tuple[bool, Optional[str]]:
+        """Check CUDA health. Returns (is_healthy, error_message).
+
+        - If not a GPU environment: returns (True, None) immediately
+        - If CUDA previously failed: returns cached failure immediately
+        - Otherwise: runs synchronize + mem_get_info check
+
+        Thread-safe. The actual CUDA check is serialized by the lock to
+        prevent concurrent CUDA calls during health checking.
+        """
+        # Fast path: not a GPU environment
+        if not self._is_gpu_environment():
+            return True, None
+
+        # Fast path: already known to be failed (unrecoverable)
+        if self._cuda_failed:
+            return False, self._failure_error
+
+        # Slow path: actually check CUDA
+        with self._lock:
+            # Double-check after acquiring lock
+            if self._cuda_failed:
+                return False, self._failure_error
+
+            try:
+                import torch
+
+                # Synchronize to surface any pending async CUDA errors
+                torch.cuda.synchronize()
+                # Query runtime to verify it's still functional
+                torch.cuda.mem_get_info()
+                return True, None
+            except Exception as e:
+                error_msg = f"CUDA health check failed: {e}"
+                logger.error(error_msg)
+                self._cuda_failed = True
+                self._failure_error = error_msg
+                self._failure_time = time.time()
+                return False, error_msg
+
+    @property
+    def is_failed(self) -> bool:
+        return self._cuda_failed
+
+    @property
+    def failure_info(self) -> Optional[dict]:
+        if not self._cuda_failed:
+            return None
+        return {
+            "error": self._failure_error,
+            "failed_at": self._failure_time,
+        }
+
+
+# Module-level singleton
+_checker = CudaHealthChecker()
+
+
+def check_cuda_health() -> Tuple[bool, Optional[str]]:
+    """Module-level convenience function."""
+    return _checker.check_health()
+
+
+def get_cuda_health_checker() -> CudaHealthChecker:
+    """Return the singleton for dependency injection / testing."""
+    return _checker

tests/unit/core/test_cuda_health.py

@@ -0,0 +1,120 @@
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+from inference.core.utils.cuda_health import CudaHealthChecker
+
+
+class TestCudaHealthChecker:
+    def setup_method(self):
+        """Create a fresh checker for each test."""
+        self.checker = CudaHealthChecker()
+
+    def test_cpu_environment_no_torch(self):
+        """When torch is not installed, should always return healthy."""
+        with patch.dict("sys.modules", {"torch": None}):
+            self.checker._gpu_available = None  # reset cache
+            is_healthy, error = self.checker.check_health()
+            assert is_healthy is True
+            assert error is None
+
+    def test_cpu_environment_no_cuda(self):
+        """When torch is available but CUDA is not, should return healthy."""
+        mock_torch = MagicMock()
+        mock_torch.cuda.is_available.return_value = False
+        with patch.dict("sys.modules", {"torch": mock_torch}):
+            self.checker._gpu_available = None
+            is_healthy, error = self.checker.check_health()
+            assert is_healthy is True
+            assert error is None
+
+    def test_healthy_gpu(self):
+        """When CUDA operations succeed, should return healthy."""
+        mock_torch = MagicMock()
+        mock_torch.cuda.is_available.return_value = True
+        mock_torch.cuda.synchronize.return_value = None
+        mock_torch.cuda.mem_get_info.return_value = (4_000_000_000, 8_000_000_000)
+
+        self.checker._gpu_available = True
+        with patch.dict("sys.modules", {"torch": mock_torch}):
+            is_healthy, error = self.checker.check_health()
+            assert is_healthy is True
+            assert error is None
+            mock_torch.cuda.synchronize.assert_called_once()
+            mock_torch.cuda.mem_get_info.assert_called_once()
+
+    def test_cuda_synchronize_failure(self):
+        """When torch.cuda.synchronize() fails, should detect CUDA corruption."""
+        mock_torch = MagicMock()
+        mock_torch.cuda.is_available.return_value = True
+        mock_torch.cuda.synchronize.side_effect = RuntimeError(
+            "CUDA error: an illegal memory access was encountered"
+        )
+
+        self.checker._gpu_available = True
+        with patch.dict("sys.modules", {"torch": mock_torch}):
+            is_healthy, error = self.checker.check_health()
+            assert is_healthy is False
+            assert "illegal memory access" in error
+            assert self.checker.is_failed is True
+
+    def test_mem_get_info_failure(self):
+        """When mem_get_info fails (after synchronize succeeds), should detect failure."""
+        mock_torch = MagicMock()
+        mock_torch.cuda.is_available.return_value = True
+        mock_torch.cuda.synchronize.return_value = None
+        mock_torch.cuda.mem_get_info.side_effect = RuntimeError("CUDA runtime error")
+
+        self.checker._gpu_available = True
+        with patch.dict("sys.modules", {"torch": mock_torch}):
+            is_healthy, error = self.checker.check_health()
+            assert is_healthy is False
+            assert "CUDA runtime error" in error
+
+    def test_failure_is_cached(self):
+        """After first CUDA failure, subsequent checks should return cached failure
+        without calling torch again."""
+        mock_torch = MagicMock()
+        mock_torch.cuda.is_available.return_value = True
+        mock_torch.cuda.synchronize.side_effect = RuntimeError("CUDA error")
+
+        self.checker._gpu_available = True
+        with patch.dict("sys.modules", {"torch": mock_torch}):
+            # First call: detects failure
+            is_healthy1, error1 = self.checker.check_health()
+            assert is_healthy1 is False
+            assert mock_torch.cuda.synchronize.call_count == 1
+
+            # Second call: returns cached failure, no new CUDA calls
+            mock_torch.cuda.synchronize.reset_mock()
+            is_healthy2, error2 = self.checker.check_health()
+            assert is_healthy2 is False
+            assert error2 == error1
+            mock_torch.cuda.synchronize.assert_not_called()
+
+    def test_failure_info(self):
+        """failure_info should return error details after failure."""
+        assert self.checker.failure_info is None
+
+        mock_torch = MagicMock()
+        mock_torch.cuda.is_available.return_value = True
+        mock_torch.cuda.synchronize.side_effect = RuntimeError("CUDA error")
+
+        self.checker._gpu_available = True
+        with patch.dict("sys.modules", {"torch": mock_torch}):
+            self.checker.check_health()
+
+        info = self.checker.failure_info
+        assert info is not None
+        assert "CUDA error" in info["error"]
+        assert info["failed_at"] is not None
+
+    def test_gpu_available_is_cached(self):
+        """_is_gpu_environment() should only check torch once."""
+        mock_torch = MagicMock()
+        mock_torch.cuda.is_available.return_value = False
+        with patch.dict("sys.modules", {"torch": mock_torch}):
+            assert self.checker._is_gpu_environment() is False
+            assert self.checker._is_gpu_environment() is False
+            # Only called once despite two invocations
+            mock_torch.cuda.is_available.assert_called_once()