[Feature] Reasoning model API support

[lambert0312]

Checklist

• 1. If the issue you raised is not a feature but a question, please raise a discussion at https://github.com/sgl-project/sglang/discussions/new/choose Otherwise, it will be closed.
• 2. Please use English, otherwise it will be closed.

Motivation

In order to better support reasoning models, such as DeepSeek-R1, etc., the API needs to support the reasoning_effort parameter. In addition, it is recommended to add reasoning_content to the output field mentioned in reasoning_model , used to display step information of reasoning thinking.
Similar to the dialogue completion interface parameters provided by openai. The parameter reasoning_effort support o1 model: "constrains effort on reasoning for reasoning models. Currently supported values are low, medium, and high. Reducing reasoning effort can result in faster responses and fewer tokens used on reasoning in a response."

Related resources

No response

[zhaochenyang20]

@lambert0312 I am pondering whether OpenAI has this parameter? SGLang aligns directly with OpenAI API.

[gaocegege]

As far as I know, there isn't such a parameter for OpenAI's reasoning model because it doesn't show the CoT tokens.

Ref https://github.com/openai/openai-openapi/blob/master/openapi.yaml

However, this doesn't break API compatibility since it's a newly added parameter. The code below still works even though the response.choices[0].message's data class, ChatCompletionMessage, doesn't include such a parameter. Given that, I think it’s safe to add such a parameter, especially since more reasoning models will likely output CoT tokens soon.

from openai import OpenAI
client = OpenAI(api_key="", base_url="https://api.deepseek.com")

# Round 1
messages = [{"role": "user", "content": "9.11 and 9.8, which is greater?"}]
response = client.chat.completions.create(
    model="deepseek-reasoner",
    messages=messages
)

+ reasoning_content = response.choices[0].message.reasoning_content
content = response.choices[0].message.content

class ChatCompletionMessage(BaseModel):
    content: Optional[str] = None
    """The contents of the message."""

    refusal: Optional[str] = None
    """The refusal message generated by the model."""

    role: Literal["assistant"]
    """The role of the author of this message."""

    audio: Optional[ChatCompletionAudio] = None
    """
    If the audio output modality is requested, this object contains data about the
    audio response from the model.
    [Learn more](https://platform.openai.com/docs/guides/audio).
    """

    function_call: Optional[FunctionCall] = None
    """Deprecated and replaced by `tool_calls`.

    The name and arguments of a function that should be called, as generated by the
    model.
    """

    tool_calls: Optional[List[ChatCompletionMessageToolCall]] = None
    """The tool calls generated by the model, such as function calls."""

[zhaochenyang20]

We will see if someone would love to work on this. Thansk!