vllm_omni.errors ¶

Request-scoped client error types shared across vLLM-Omni entrypoints.

DEFAULT_CLIENT_ERROR_TYPE `module-attribute` ¶

DEFAULT_CLIENT_ERROR_TYPE = 'BadRequestError'

GuardrailViolationError ¶

Bases: OmniClientError

Raised when a model guardrail rejects request content.

OmniClientError ¶

Bases: ValueError

Request-scoped error that should be surfaced as a 4xx response. One example of using OmniClientError is GuardrailViolationError, which is captured and resurfaced as HTTP 400 error code, instead of a generic 500. OmniClientError should be used for any exceptions which need to be resurfaced as 4xx, as opposed to EngineDeadError/EngineGenerateError which are resurfaced as 500.

error_type `instance-attribute` ¶

error_type = error_type

message `instance-attribute` ¶

message = message

status_code `instance-attribute` ¶

status_code = int(status_code)

client_error_from_metadata ¶

client_error_from_metadata(
    message: str,
    *,
    status_code: int | None,
    error_type: str | None,
) -> OmniClientError

client_error_metadata ¶

client_error_metadata(
    exc: BaseException,
) -> tuple[int | None, str | None]

is_client_error_status ¶

is_client_error_status(status_code: int | None) -> bool

raise_client_error_or ¶

raise_client_error_or(
    message: str,
    *,
    status_code: int | None,
    error_type: str | None,
    fallback: Callable[[str], BaseException],
) -> NoReturn

Raise a client error for 4xx statuses, otherwise raise fallback(message).

Centralizes the "client-error-or-fallback" decision shared by the engine error paths so the status mapping lives in one place.