sgl-project/sglang · error · DS32EncodingError

Invalid thinking_mode `{thinking_mode}`

Error message

Invalid thinking_mode `{thinking_mode}`

What it means

DSv32 chat encoding rejects an unknown thinking_mode. render_message only supports 'chat' and 'thinking'; any other string (or a typo) raises DS32EncodingError before any rendering.

Source

Thrown at python/sglang/srt/entrypoints/openai/encoding_dsv32.py:168

def find_last_user_index(messages: List[Dict[str, Any]]) -> int:
    last_user_index = -1
    for idx in range(len(messages) - 1, -1, -1):
        if messages[idx].get("role") in ["user", "developer"]:
            last_user_index = idx
            break
    return last_user_index


def render_message(
    index: int, messages: List[Dict[str, Any]], thinking_mode: str
) -> str:
    if not (0 <= index < len(messages)):
        raise DS32EncodingError(
            f"Index {index} out of range for messages list of length {len(messages)}"
        )
    if thinking_mode not in ["chat", "thinking"]:
        raise DS32EncodingError(f"Invalid thinking_mode `{thinking_mode}`")

    prompt = ""
    msg = messages[index]
    last_user_idx = find_last_user_index(messages)

    role = msg.get("role")
    content = msg.get("content")
    tools = msg.get("tools")
    response_format = msg.get("response_format")
    tool_calls = msg.get("tool_calls")
    reasoning_content = msg.get("reasoning_content")

    if tools:
        tools = tools_from_openai_format(tools)
    if tool_calls:
        tool_calls = tool_calls_from_openai_format(tool_calls)

    if role == "system":

View on GitHub (pinned to 0132848349)

Solutions

1. Set thinking_mode to exactly 'chat' or 'thinking' (lowercase).
2. Check where thinking_mode is derived (server args / chat completion request) and normalize with .lower() before calling encode_messages.
3. Upgrade sglang if a newer mode name was introduced and your copy is stale.

Example fix

# before
encode_messages(msgs, thinking_mode='Thinking')
# after
encode_messages(msgs, thinking_mode='thinking')

assert thinking_mode in ('chat','thinking'), f"bad thinking_mode: {thinking_mode!r}"

def is_valid_thinking_mode(m: str) -> bool:
    return isinstance(m, str) and m in ('chat','thinking')

When it happens

Trigger: Calling encode_messages/render_message for the deepseek-v3.2 chat template with thinking_mode set to something other than 'chat' or 'thinking' (e.g. 'Think', 'auto', None passed as string, or a future mode name).

Common situations: Passing server args like --reasoning-quotes or a chat template flag that maps to a thinking_mode string not in the allowlist; case mismatch ('Thinking'); version upgrade that renamed modes.

Related errors

• No tool calls but found tool output
• ThinkingMode: {thinking_mode}, invalid message without reaso
• Unknown role: {role}
• Invalid content format: {content_format}
• --default-chat-template-kwargs must decode to a JSON object

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/7f18b77a311d4fe2. Report an issue: GitHub.