[Frontend] Clean Client Facing 500 Report (#58043)

Co-authored-by: Claude <noreply@anthropic.com>

tests/entrypoints/serve/exception_handling/test_error_sanitization.py

@@ -135,3 +135,21 @@ class TestAffectedModulesUseSanitize:
         source = Path(spec.origin).read_text()
         assert "sanitize_message" in source, f"{module} does not call sanitize_message"
         assert "import" in source and "sanitize_message" in source
+
+
+def test_engine_dead_error_response_hides_log_oriented_message():
+    """Clients get a generic message with no reference to server internals."""
+    from vllm.entrypoints.serve.exception_handling.error_response import (
+        create_error_response,
+    )
+    from vllm.v1.engine.exceptions import EngineDeadError
+
+    exc = EngineDeadError()
+    err = create_error_response(exc)
+
+    assert err.error.code == 500
+    assert err.error.type == "InternalServerError"
+    assert err.error.message == EngineDeadError.CLIENT_MESSAGE
+    assert "stack trace" not in err.error.message
+    assert "log" not in err.error.message
+    assert "stack trace (above)" in str(exc)

vllm/entrypoints/serve/exception_handling/error_response.py

@@ -21,6 +21,7 @@ def create_error_response(
 
     if isinstance(message, Exception):
         exc = message
+        message = str(exc)
         logger.debug(
             "create_error_response called with %s: %s", type(exc).__name__, exc
         )
@@ -33,6 +34,7 @@ def create_error_response(
             VLLMUnprocessableEntityError,
             VLLMValidationError,
         )
+        from vllm.v1.engine.exceptions import EngineDeadError
 
         if isinstance(exc, GracefulHTTPError):
             err_type = HTTPStatus(exc.http_status).phrase
@@ -63,6 +65,11 @@ def create_error_response(
             )
             status_code = exc.status_code
             param = None
+        elif isinstance(exc, EngineDeadError):
+            err_type = "InternalServerError"
+            status_code = HTTPStatus.INTERNAL_SERVER_ERROR
+            param = None
+            message = exc.CLIENT_MESSAGE
         elif isinstance(exc, VLLMServerError):
             # Any other server-caused error defaults to 500.
             err_type = "InternalServerError"
@@ -88,8 +95,6 @@ def create_error_response(
             status_code = HTTPStatus.INTERNAL_SERVER_ERROR
             param = None
 
-        message = str(exc)
-
     return ErrorResponse(
         error=ErrorInfo(
             message=sanitize_message(message),

vllm/v1/engine/exceptions.py

@@ -10,12 +10,19 @@ class EngineGenerateError(VLLMServerError):
 
 
 class EngineDeadError(VLLMServerError):
-    """Raised when the EngineCore dies. Unrecoverable."""
+    """Raised when the EngineCore dies. Unrecoverable.
+
+    `str(exc)` is the log-oriented message for server logs and offline
+    tracebacks; `CLIENT_MESSAGE` is what API clients receive.
+    """
+
+    ENGINE_DEAD_MESSAGE = (
+        "EngineCore encountered an issue. See stack trace (above) for the root cause."
+    )
+    CLIENT_MESSAGE = "The server had an error while processing your request."
 
     def __init__(self, *args, suppress_context: bool = False, **kwargs):
-        ENGINE_DEAD_MESSAGE = "EngineCore encountered an issue. See stack trace (above) for the root cause."  # noqa: E501
-
-        super().__init__(ENGINE_DEAD_MESSAGE, *args, **kwargs)
+        super().__init__(self.ENGINE_DEAD_MESSAGE, *args, **kwargs)
         # Make stack trace clearer when using with LLMEngine by
         # silencing irrelevant ZMQError.
         self.__suppress_context__ = suppress_context