[Feature] Support for `tool_calls` in Assistant Messages Input

[kibitzing]

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

SGLang currently supports tool_calls in output responses, but does not properly handle them in input messages. When users include tool_calls in the assistant turns within messages, these are not reflected in the tokens during model inference.

Current Implementation

The current implementation only allows for role and content in assistant messages:

sglang/python/sglang/srt/openai_api/protocol.py

Lines 253 to 255 in 1acca3a

	class ChatCompletionMessageGenericParam(BaseModel):
	role: Literal["system", "assistant", "tool"]
	content: Union[str, List[ChatCompletionMessageContentTextPart], None]

sglang/python/sglang/srt/openai_api/adapter.py

Lines 968 to 983 in 1acca3a

	openai_compatible_messages = []
	for message in request.messages:
	if message.content is None:
	message.content = ""
	if isinstance(message.content, str):
	openai_compatible_messages.append(
	{"role": message.role, "content": message.content}
	)
	else:
	content_list = message.dict()["content"]
	for content in content_list:
	if content["type"] == "text":
	openai_compatible_messages.append(
	{"role": message.role, "content": content["text"]}
	)

Impact

This limitation causes the framework to deviate from the intended behavior of models like Qwen3, which explicitly handle tool_calls in their chat templates. For example, Qwen's chat template includes specific handling for tool calls:

{%- if message.tool_calls %}
    {%- for tool_call in message.tool_calls %}
        {%- if (loop.first and content) or (not loop.first) %}
            {{- '\\n' }}
        {%- endif %}
        {%- if tool_call.function %}
            {%- set tool_call = tool_call.function %}
        {%- endif %}
        {{- '<tool_call>\\n{\"name\": \"' }}
        {{- tool_call.name }}
        {{- '\", \"arguments\": ' }}
        {%- if tool_call.arguments is string %}
            {{- tool_call.arguments }}
        {%- else %}
            {{- tool_call.arguments | tojson }}
        {%- endif %}
        {{- '}\\n</tool_call>' }}
    {%- endfor %}
{%- endif %}

Requested Enhancement

1. Update the ChatCompletionMessageGenericParam class to include support for tool_calls
2. Modify the openai_compatible_message list processing logic in the openai_api/adapter.py to handle tool_calls in assistant messages properly, ensuring they are correctly incorporated when formatting messages for the model

Without these capabilities, the inference operates differently from what both users and model creators would expect, potentially leading to reduced functionality and unexpected behavior in tool-using scenarios.

I have implemented a fix for this issue and am ready to submit a PR once this feature request is confirmed. This should help avoid duplicate effort from other contributors.

Related resources

• The full version of Qwen3 chat template:
  • https://huggingface.co/Qwen/Qwen3-4B/blob/main/tokenizer_config.json#L230

[kibitzing]

I think it has been fixed here: #5996