diff --git a/vllm/entrypoints/serve/render/serving.py b/vllm/entrypoints/serve/render/serving.py index c99c014413..5aa2449797 100644 --- a/vllm/entrypoints/serve/render/serving.py +++ b/vllm/entrypoints/serve/render/serving.py @@ -548,7 +548,9 @@ class OpenAIServingRender: if reasoning_parser is not None: tokenizer = renderer.get_tokenizer() - request = reasoning_parser(tokenizer).adjust_request(request=request) + request = reasoning_parser( + tokenizer, model_config=self.model_config + ).adjust_request(request=request) # tool parsing is done only if a tool_parser has been set and if # tool_choice is not "none" (if tool_choice is "none" but a tool_parser diff --git a/vllm/reasoning/abs_reasoning_parsers.py b/vllm/reasoning/abs_reasoning_parsers.py index 2ff72ea2f5..8edbc5f82e 100644 --- a/vllm/reasoning/abs_reasoning_parsers.py +++ b/vllm/reasoning/abs_reasoning_parsers.py @@ -14,6 +14,7 @@ from vllm.utils.collection_utils import is_list_of from vllm.utils.import_utils import import_from_path if TYPE_CHECKING: + from vllm.config import ModelConfig from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest from vllm.entrypoints.openai.engine.protocol import DeltaMessage from vllm.entrypoints.openai.responses.protocol import ResponsesRequest @@ -32,6 +33,9 @@ class ReasoningParser: def __init__(self, tokenizer: "TokenizerLike", *args, **kwargs): self.model_tokenizer = tokenizer + # Optional vLLM ModelConfig from the server. Use get (not pop) so composite + # parsers can forward **kwargs to nested parsers. + self._model_config: ModelConfig | None = kwargs.get("model_config") @cached_property def vocab(self) -> dict[str, int]: