| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | """Custom exceptions for vLLM.""" |
| 5 | |
| 6 | from http import HTTPStatus |
| 7 | from typing import Any |
| 8 | |
| 9 | |
| 10 | class VLLMError(Exception): |
| 11 | """Base class for all vLLM-specific errors. |
| 12 | |
| 13 | Subclasses are split into `VLLMClientError` (caused by the request, mapped |
| 14 | to 4xx) and `VLLMServerError` (caused by the server, mapped to 5xx). |
| 15 | Dispatching on this hierarchy lets the entrypoints decide the HTTP status |
| 16 | without relying on raw Python exception types such as `ValueError`. |
| 17 | """ |
| 18 | |
| 19 | |
| 20 | class VLLMClientError(VLLMError): |
| 21 | """Base class for errors caused by the client request (4xx).""" |
| 22 | |
| 23 | |
| 24 | class VLLMServerError(VLLMError): |
| 25 | """Base class for errors caused by the server (5xx).""" |
| 26 | |
| 27 | |
| 28 | class VLLMValidationError(VLLMClientError): |
| 29 | """vLLM-specific validation error for request validation failures. |
| 30 | |
| 31 | Args: |
| 32 | message: The error message describing the validation failure. |
| 33 | parameter: Optional parameter name that failed validation. |
| 34 | value: Optional value that was rejected during validation. |
| 35 | |
| 36 | """ |
| 37 | |
| 38 | def __init__( |
| 39 | self, |
| 40 | message: str, |
| 41 | *, |
| 42 | parameter: str | None = None, |
| 43 | value: Any = None, |
| 44 | ) -> None: |
| 45 | super().__init__(message) |
| 46 | self.parameter = parameter |
| 47 | self.value = value |
| 48 | |
| 49 | def __str__(self): |
| 50 | base = super().__str__() |
| 51 | extras = [] |
| 52 | if self.parameter is not None: |
| 53 | extras.append(f"parameter={self.parameter}") |
| 54 | if self.value is not None: |
| 55 | extras.append(f"value={self.value}") |
| 56 | return f"{base} ({', '.join(extras)})" if extras else base |
| 57 | |
| 58 | |
| 59 | class VLLMNotFoundError(VLLMClientError): |
| 60 | """vLLM-specific NotFoundError.""" |
| 61 | |
| 62 | pass |
| 63 | |
| 64 | |
| 65 | class LoRAAdapterNotFoundError(VLLMNotFoundError): |
| 66 | """Exception raised when a LoRA adapter is not found. |
| 67 | |
| 68 | This exception is thrown when a requested LoRA adapter does not exist |
| 69 | in the system. |
| 70 | |
| 71 | Attributes: |
| 72 | message: The error message string describing the exception |
| 73 | |
| 74 | """ |
| 75 | |
| 76 | message: str |
| 77 | |
| 78 | def __init__( |
| 79 | self, |
| 80 | lora_name: str, |
| 81 | lora_path: str, |
| 82 | ) -> None: |
| 83 | message = f"Loading lora {lora_name} failed: No adapter found for {lora_path}" |
| 84 | self.message = message |
| 85 | |
| 86 | def __str__(self): |
| 87 | return self.message |
| 88 | |
| 89 | |
| 90 | class VLLMUnprocessableEntityError(VLLMClientError): |
| 91 | """vLLM-specific error for unprocessable entity requests. |
| 92 | |
| 93 | This exception is raised when the request content is invalid or cannot be |
| 94 | processed, such as when an image URL points to a non-existent or inaccessible |
| 95 | resource (404, 403, DNS failure, etc.). |
| 96 | |
| 97 | Args: |
| 98 | message: The error message describing the unprocessable entity. |
| 99 | parameter: Optional parameter name that failed validation. |
| 100 | value: Optional value that was rejected during validation. |
| 101 | |
| 102 | """ |
| 103 | |
| 104 | def __init__( |
| 105 | self, |
| 106 | message: str, |
| 107 | *, |
| 108 | parameter: str | None = None, |
| 109 | value: Any = None, |
| 110 | ) -> None: |
| 111 | super().__init__(message) |
| 112 | self.parameter = parameter |
| 113 | self.value = value |
| 114 | |
| 115 | def __str__(self): |
| 116 | base = super().__str__() |
| 117 | extras = [] |
| 118 | if self.parameter is not None: |
| 119 | extras.append(f"parameter={self.parameter}") |
| 120 | if self.value is not None: |
| 121 | extras.append(f"value={self.value}") |
| 122 | return f"{base} ({', '.join(extras)})" if extras else base |
| 123 | |
| 124 | |
| 125 | class GenerationError(VLLMServerError): |
| 126 | """raised when finish_reason indicates internal server error (500)""" |
| 127 | |
| 128 | def __init__(self, message: str = "Internal server error"): |
| 129 | super().__init__(message) |
| 130 | self.status_code = HTTPStatus.INTERNAL_SERVER_ERROR |
| 131 | |
| 132 | |
| 133 | class RetryableRequestError(GenerationError): |
| 134 | """A transient request-level failure that is safe for clients to retry.""" |
| 135 | |
| 136 | def __init__(self, message: str = "Request failed; please retry"): |
| 137 | super().__init__(message) |
| 138 | self.status_code = HTTPStatus.SERVICE_UNAVAILABLE |
| 139 | |
| 140 | |
| 141 | class GracefulHTTPError(VLLMError): |
| 142 | """Exception that should be translated into an HTTP error response. |
| 143 | |
| 144 | These are expected to occur during normal operation (e.g. admission |
| 145 | control rejections) and should be surfaced to the client with the |
| 146 | explicit HTTP status code they carry, rather than being mapped to a |
| 147 | generic 4xx/5xx by the client/server split. |
| 148 | """ |
| 149 | |
| 150 | def __init__(self, message: str, http_status: HTTPStatus): |
| 151 | super().__init__(message) |
| 152 | self.message = message |
| 153 | self.http_status = http_status |
| 154 | |
| 155 | |
| 156 | class QueueOverflowError(GracefulHTTPError): |
| 157 | """Raised when admitting a request would exceed the request queue limit. |
| 158 | |
| 159 | Returns HTTP 503 (Service Unavailable) so that load balancers and |
| 160 | client SDKs retry the request on a different instance. |
| 161 | """ |
| 162 | |
| 163 | def __init__(self): |
| 164 | super().__init__( |
| 165 | "The engine is currently busy and cannot accept new requests. " |
| 166 | "Please try again later or on a different instance.", |
| 167 | HTTPStatus.SERVICE_UNAVAILABLE, |
| 168 | ) |
| 169 | |
| 170 | |
| 171 | class MaxQueuedTokensError(GracefulHTTPError): |
| 172 | """Raised when the pending prefill tokens exceed the configured limit. |
| 173 | |
| 174 | Returns HTTP 503 (Service Unavailable) so that load balancers and |
| 175 | client SDKs retry the request on a different instance. |
| 176 | """ |
| 177 | |
| 178 | def __init__(self): |
| 179 | super().__init__( |
| 180 | "The engine has reached its prefill token backlog limit. " |
| 181 | "Please try again later or on a different instance.", |
| 182 | HTTPStatus.SERVICE_UNAVAILABLE, |
| 183 | ) |