md/20251011.md

问题原因

vllm在处理请求时，调用了 get_structured_output_key 函数，在处理 structured_outputs 参数时，由于不满足所有的情况，导致了抛出No valid structured output parameter found异常，该异常未被 EngineCore 捕获导致引擎崩溃，APIServer 发现引擎死了，自己也退出了进程。

函数 get_structured_output_key 实现

分支：release/v0.11.0

https://github.com/vllm-project/vllm/blob/releases/v0.11.0/vllm/v1/structured_output/request.py

def get_structured_output_key(sampling_params: SamplingParams) -> StructuredOutputKey:
    params = sampling_params.structured_outputs
    assert params is not None, "params can't be None."
    if params.json is not None:
        if not isinstance(params.json, str):
            json_str = json.dumps(params.json)
        else:
            json_str = params.json
        return (StructuredOutputOptions.JSON, json_str)
    elif params.json_object:
        return (StructuredOutputOptions.JSON_OBJECT, "")
    elif params.regex is not None:
        return (StructuredOutputOptions.REGEX, params.regex)
    elif params.choice is not None:
        if not isinstance(params.choice, str):
            json_str = json.dumps(params.choice)
        else:
            json_str = params.choice
        return (StructuredOutputOptions.CHOICE, json_str)
    elif params.grammar is not None:
        return (StructuredOutputOptions.GRAMMAR, params.grammar)
    elif params.structural_tag is not None:
        return (StructuredOutputOptions.STRUCTURAL_TAG, params.structural_tag)
    else:
        raise ValueError("No valid structured output parameter found")

分支：release/v0.10.2

https://github.com/vllm-project/vllm/blob/releases/v0.10.2/vllm/v1/structured_output/request.py

def get_structured_output_key(
        sampling_params: SamplingParams) -> StructuredOutputKey:    
    params = sampling_params.guided_decoding
    assert params is not None, "params can't be None."
    if params.json is not None:
        if not isinstance(params.json, str):
            json_str = json.dumps(params.json)
        else:
            json_str = params.json
        return (StructuredOutputOptions.JSON, json_str)
    elif params.json_object:
        return (StructuredOutputOptions.JSON_OBJECT, "")
    elif params.regex is not None:
        return (StructuredOutputOptions.REGEX, params.regex)
    elif params.choice is not None:
        if not isinstance(params.choice, str):
            json_str = json.dumps(params.choice)
        else:
            json_str = params.choice
        return (StructuredOutputOptions.CHOICE, json_str)
    elif params.grammar is not None:
        return (StructuredOutputOptions.GRAMMAR, params.grammar)
    elif params.structural_tag is not None:
        return (StructuredOutputOptions.STRUCTURAL_TAG, params.structural_tag)
    else:
        raise ValueError("No valid structured output parameter found")

2个版本的唯一区别，仅仅是提取SamplingParams的属性不一样，其它判断逻辑都是一致的

SamplingParams

https://github.com/vllm-project/vllm/blob/releases/v0.11.0/vllm/sampling_params.py

# 其它参数已省略
class SamplingParams(
    💡初始化逻辑是没问题的，默认值是None
    structured_outputs: Optional[StructuredOutputsParams] = None
    guided_decoding: Optional[GuidedDecodingParams] = None

    @staticmethod
    def from_optional(
         💡默认值也是None
        structured_outputs: Optional[StructuredOutputsParams] = None,
    ) -> "SamplingParams":
        if guided_decoding is not None:
            warnings.warn(
                "guided_decoding is deprecated. This will be removed in "
                💡官方将在 v0.12.0 废弃 guided_decoding 参数，使用 structured_outputs 参数替代，在 v0.11.0 版本做了兼容，
                "v0.12.0 or v1.0.0, which ever is soonest. Please use "
                "structured_outputs instead.",
                DeprecationWarning,
                stacklevel=2)
            structured_outputs = guided_decoding
            guided_decoding = None
        return SamplingParams(
            structured_outputs=structured_outputs,
        )

StructuredOutputOptions

class StructuredOutputOptions(enum.Enum):
    JSON = enum.auto()
    JSON_OBJECT = enum.auto()
    REGEX = enum.auto()
    GRAMMAR = enum.auto()
    CHOICE = enum.auto()
    STRUCTURAL_TAG = enum.auto()

💡只支持这6种类型，每个类型都对应 structured_outputs 下面的一个不同的参数。

崩溃日志片段

(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710] EngineCore encountered a fatal error.
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710] Traceback (most recent call last):
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710]   File "/aisoft/conda/env/vllm2/lib/python3.10/site-packages/vllm/v1/engine/core.py", line 701, in run_engine_core
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710]     engine_core.run_busy_loop()
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710]   File "/aisoft/conda/env/vllm2/lib/python3.10/site-packages/vllm/v1/engine/core.py", line 728, in run_busy_loop
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710]     self._process_engine_step()
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710]   File "/aisoft/conda/env/vllm2/lib/python3.10/site-packages/vllm/v1/engine/core.py", line 754, in _process_engine_step
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710]     outputs, model_executed = self.step_fn()
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710]   File "/aisoft/conda/env/vllm2/lib/python3.10/site-packages/vllm/v1/engine/core.py", line 283, in step
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710]     scheduler_output = self.scheduler.schedule()
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710]   File "/aisoft/conda/env/vllm2/lib/python3.10/site-packages/vllm/v1/core/sched/scheduler.py", line 359, in schedule
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710]     if structured_output_req and structured_output_req.grammar:
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710]   File "/aisoft/conda/env/vllm2/lib/python3.10/site-packages/vllm/v1/structured_output/request.py", line 45, in grammar
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710]     completed = self._check_grammar_completion()
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710]   File "/aisoft/conda/env/vllm2/lib/python3.10/site-packages/vllm/v1/structured_output/request.py", line 33, in _check_grammar_completion
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710]     self._grammar = self._grammar.result(timeout=0.0001)
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710]   File "/aisoft/conda/env/vllm2/lib/python3.10/concurrent/futures/_base.py", line 458, in result
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710]     return self.__get_result()
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710]   File "/aisoft/conda/env/vllm2/lib/python3.10/concurrent/futures/_base.py", line 403, in __get_result
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710]     raise self._exception
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710]   File "/aisoft/conda/env/vllm2/lib/python3.10/concurrent/futures/thread.py", line 58, in run
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710]     result = self.fn(*self.args, **self.kwargs)
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710]   File "/aisoft/conda/env/vllm2/lib/python3.10/site-packages/vllm/v1/structured_output/__init__.py", line 128, in _async_create_grammar
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710]     key = request.structured_output_request.structured_output_key  # type: ignore[union-attr]
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710]   File "/aisoft/conda/env/vllm2/lib/python3.10/functools.py", line 981, in __get__
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710]     val = self.func(instance)
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710]   File "/aisoft/conda/env/vllm2/lib/python3.10/site-packages/vllm/v1/structured_output/request.py", line 58, in structured_output_key
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710]     return get_structured_output_key(self.sampling_params)
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710]   File "/aisoft/conda/env/vllm2/lib/python3.10/site-packages/vllm/v1/structured_output/request.py", line 86, in get_structured_output_key
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710]     raise ValueError("No valid structured output parameter found")
(EngineCore_DP0 pid=2738693) ERROR 10-10 10:43:10 [core.py:710] ValueError: No valid structured output parameter found
(EngineCore_DP0 pid=2738693) Process EngineCore_DP0:
(EngineCore_DP0 pid=2738693) Traceback (most recent call last):
(APIServer pid=2738423) ERROR 10-10 10:43:10 [async_llm.py:480] AsyncLLM output_handler failed.
(APIServer pid=2738423) ERROR 10-10 10:43:10 [async_llm.py:480] Traceback (most recent call last):
(APIServer pid=2738423) ERROR 10-10 10:43:10 [async_llm.py:480]   File "/aisoft/conda/env/vllm2/lib/python3.10/site-packages/vllm/v1/engine/async_llm.py", line 439, in output_handler
(APIServer pid=2738423) ERROR 10-10 10:43:10 [async_llm.py:480]     outputs = await engine_core.get_output_async()
(APIServer pid=2738423) ERROR 10-10 10:43:10 [async_llm.py:480]   File "/aisoft/conda/env/vllm2/lib/python3.10/site-packages/vllm/v1/engine/core_client.py", line 846, in get_output_async
(APIServer pid=2738423) ERROR 10-10 10:43:10 [async_llm.py:480]     raise self._format_exception(outputs) from None
(APIServer pid=2738423) ERROR 10-10 10:43:10 [async_llm.py:480] vllm.v1.engine.exceptions.EngineDeadError: EngineCore encountered an issue. See stack trace (above) for the root cause.
(EngineCore_DP0 pid=2738693)   File "/aisoft/conda/env/vllm2/lib/python3.10/multiprocessing/process.py", line 314, in _bootstrap
(EngineCore_DP0 pid=2738693)     self.run()
(EngineCore_DP0 pid=2738693)   File "/aisoft/conda/env/vllm2/lib/python3.10/multiprocessing/process.py", line 108, in run
(EngineCore_DP0 pid=2738693)     self._target(*self._args, **self._kwargs)
(EngineCore_DP0 pid=2738693)   File "/aisoft/conda/env/vllm2/lib/python3.10/site-packages/vllm/v1/engine/core.py", line 712, in run_engine_core
(EngineCore_DP0 pid=2738693)     raise e
(EngineCore_DP0 pid=2738693)   File "/aisoft/conda/env/vllm2/lib/python3.10/site-packages/vllm/v1/engine/core.py", line 701, in run_engine_core
(EngineCore_DP0 pid=2738693)     engine_core.run_busy_loop()
(EngineCore_DP0 pid=2738693)   File "/aisoft/conda/env/vllm2/lib/python3.10/site-packages/vllm/v1/engine/core.py", line 728, in run_busy_loop
(EngineCore_DP0 pid=2738693)     self._process_engine_step()
(EngineCore_DP0 pid=2738693)   File "/aisoft/conda/env/vllm2/lib/python3.10/site-packages/vllm/v1/engine/core.py", line 754, in _process_engine_step
(EngineCore_DP0 pid=2738693)     outputs, model_executed = self.step_fn()
(EngineCore_DP0 pid=2738693)   File "/aisoft/conda/env/vllm2/lib/python3.10/site-packages/vllm/v1/engine/core.py", line 283, in step
(EngineCore_DP0 pid=2738693)     scheduler_output = self.scheduler.schedule()
(EngineCore_DP0 pid=2738693)   File "/aisoft/conda/env/vllm2/lib/python3.10/site-packages/vllm/v1/core/sched/scheduler.py", line 359, in schedule
(EngineCore_DP0 pid=2738693)     if structured_output_req and structured_output_req.grammar:
(EngineCore_DP0 pid=2738693)   File "/aisoft/conda/env/vllm2/lib/python3.10/site-packages/vllm/v1/structured_output/request.py", line 45, in grammar
(EngineCore_DP0 pid=2738693)     completed = self._check_grammar_completion()
(EngineCore_DP0 pid=2738693)   File "/aisoft/conda/env/vllm2/lib/python3.10/site-packages/vllm/v1/structured_output/request.py", line 33, in _check_grammar_completion
(EngineCore_DP0 pid=2738693)     self._grammar = self._grammar.result(timeout=0.0001)
(EngineCore_DP0 pid=2738693)   File "/aisoft/conda/env/vllm2/lib/python3.10/concurrent/futures/_base.py", line 458, in result
(EngineCore_DP0 pid=2738693)     return self.__get_result()
(EngineCore_DP0 pid=2738693)   File "/aisoft/conda/env/vllm2/lib/python3.10/concurrent/futures/_base.py", line 403, in __get_result
(EngineCore_DP0 pid=2738693)     raise self._exception
(EngineCore_DP0 pid=2738693)   File "/aisoft/conda/env/vllm2/lib/python3.10/concurrent/futures/thread.py", line 58, in run
(EngineCore_DP0 pid=2738693)     result = self.fn(*self.args, **self.kwargs)
(EngineCore_DP0 pid=2738693)   File "/aisoft/conda/env/vllm2/lib/python3.10/site-packages/vllm/v1/structured_output/__init__.py", line 128, in _async_create_grammar
(EngineCore_DP0 pid=2738693)     key = request.structured_output_request.structured_output_key  # type: ignore[union-attr]
(EngineCore_DP0 pid=2738693)   File "/aisoft/conda/env/vllm2/lib/python3.10/functools.py", line 981, in __get__
(EngineCore_DP0 pid=2738693)     val = self.func(instance)
(EngineCore_DP0 pid=2738693)   File "/aisoft/conda/env/vllm2/lib/python3.10/site-packages/vllm/v1/structured_output/request.py", line 58, in structured_output_key
(EngineCore_DP0 pid=2738693)     return get_structured_output_key(self.sampling_params)
(EngineCore_DP0 pid=2738693)   File "/aisoft/conda/env/vllm2/lib/python3.10/site-packages/vllm/v1/structured_output/request.py", line 86, in get_structured_output_key
(EngineCore_DP0 pid=2738693)     raise ValueError("No valid structured output parameter found")
(EngineCore_DP0 pid=2738693) ValueError: No valid structured output parameter found
(APIServer pid=2738423) ERROR 10-10 10:43:10 [serving_chat.py:1145] Error in chat completion stream generator.
(APIServer pid=2738423) ERROR 10-10 10:43:10 [serving_chat.py:1145] Traceback (most recent call last):
(APIServer pid=2738423) ERROR 10-10 10:43:10 [serving_chat.py:1145]   File "/aisoft/conda/env/vllm2/lib/python3.10/site-packages/vllm/entrypoints/openai/serving_chat.py", line 574, in chat_completion_stream_generator
(APIServer pid=2738423) ERROR 10-10 10:43:10 [serving_chat.py:1145]     async for res in result_generator:
(APIServer pid=2738423) ERROR 10-10 10:43:10 [serving_chat.py:1145]   File "/aisoft/conda/env/vllm2/lib/python3.10/site-packages/vllm/v1/engine/async_llm.py", line 387, in generate
(APIServer pid=2738423) ERROR 10-10 10:43:10 [serving_chat.py:1145]     out = q.get_nowait() or await q.get()
(APIServer pid=2738423) ERROR 10-10 10:43:10 [serving_chat.py:1145]   File "/aisoft/conda/env/vllm2/lib/python3.10/site-packages/vllm/v1/engine/output_processor.py", line 59, in get
(APIServer pid=2738423) ERROR 10-10 10:43:10 [serving_chat.py:1145]     raise output
(APIServer pid=2738423) ERROR 10-10 10:43:10 [serving_chat.py:1145]   File "/aisoft/conda/env/vllm2/lib/python3.10/site-packages/vllm/v1/engine/async_llm.py", line 439, in output_handler
(APIServer pid=2738423) ERROR 10-10 10:43:10 [serving_chat.py:1145]     outputs = await engine_core.get_output_async()
(APIServer pid=2738423) ERROR 10-10 10:43:10 [serving_chat.py:1145]   File "/aisoft/conda/env/vllm2/lib/python3.10/site-packages/vllm/v1/engine/core_client.py", line 846, in get_output_async
(APIServer pid=2738423) ERROR 10-10 10:43:10 [serving_chat.py:1145]     raise self._format_exception(outputs) from None
(APIServer pid=2738423) ERROR 10-10 10:43:10 [serving_chat.py:1145] vllm.v1.engine.exceptions.EngineDeadError: EngineCore encountered an issue. See stack trace (above) for the root cause.
[rank0]:[W1010 10:43:10.666166970 ProcessGroupNCCL.cpp:1538] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
(APIServer pid=2738423) INFO:     Shutting down
(APIServer pid=2738423) INFO:     Waiting for application shutdown.
(APIServer pid=2738423) INFO:     Application shutdown complete.
(APIServer pid=2738423) INFO:     Finished server process [2738423]