[Bugfix] Add error handling for FINISHED_ERROR in OpenAIServing (#37148)

Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>

vllm/entrypoints/openai/api_server.py

@@ -29,11 +29,13 @@ from vllm.entrypoints.chat_utils import load_chat_template
 from vllm.entrypoints.launcher import serve_http
 from vllm.entrypoints.logger import RequestLogger
 from vllm.entrypoints.openai.cli_args import make_arg_parser, validate_parsed_serve_args
+from vllm.entrypoints.openai.engine.protocol import GenerationError
 from vllm.entrypoints.openai.models.protocol import BaseModelPath
 from vllm.entrypoints.openai.models.serving import OpenAIServingModels
 from vllm.entrypoints.openai.server_utils import (
     engine_error_handler,
     exception_handler,
+    generation_error_handler,
     get_uvicorn_log_config,
     http_exception_handler,
     lifespan,
@@ -263,6 +265,7 @@ def build_app(
     app.exception_handler(RequestValidationError)(validation_exception_handler)
     app.exception_handler(EngineGenerateError)(engine_error_handler)
     app.exception_handler(EngineDeadError)(engine_error_handler)
+    app.exception_handler(GenerationError)(generation_error_handler)
     app.exception_handler(Exception)(exception_handler)
 
     # Ensure --api-key option from CLI takes precedence over VLLM_API_KEY

vllm/entrypoints/openai/server_utils.py

@@ -21,7 +21,11 @@ from starlette.types import ASGIApp, Message, Receive, Scope, Send
 from vllm import envs
 from vllm.engine.protocol import EngineClient
 from vllm.entrypoints.launcher import terminate_if_errored
-from vllm.entrypoints.openai.engine.protocol import ErrorInfo, ErrorResponse
+from vllm.entrypoints.openai.engine.protocol import (
+    ErrorInfo,
+    ErrorResponse,
+    GenerationError,
+)
 from vllm.entrypoints.utils import create_error_response, sanitize_message
 from vllm.exceptions import VLLMValidationError
 from vllm.logger import init_logger
@@ -354,6 +358,17 @@ async def engine_error_handler(
     return JSONResponse(err.model_dump(), status_code=err.error.code)
 
 
+async def generation_error_handler(req: Request, exc: GenerationError):
+    """Handle GenerationError without logging stack traces.
+
+    GenerationError is a known, expected error (e.g. KV cache load failure)
+    that should be returned to the client as a 500 response without polluting
+    server logs with stack traces.
+    """
+    err = create_error_response(exc)
+    return JSONResponse(err.model_dump(), status_code=err.error.code)
+
+
 async def exception_handler(req: Request, exc: Exception):
     if req.app.state.args.log_error_stack:
         logger.exception(