sgl-project/sglang · error · ValueError

{template_error}

Error message

{template_error}

What it means

Raised when applying the model's Jinja chat template fails (e.g. the template calls raise_exception, or a filter like tojson hits an Undefined variable). SGLang wraps the underlying template/TypeError into a ValueError so the OpenAI-compatible layer maps it to a 400 BadRequest client error. It originates in _apply_jinja_template, called while processing chat messages.

Source

Thrown at python/sglang/srt/entrypoints/openai/serving_chat.py:1423

                try:
                    rendered_prompt = (
                        self.tokenizer_manager.tokenizer.apply_chat_template(
                            openai_compatible_messages,
                            tokenize=False,
                            add_generation_prompt=True,
                            tools=tools,
                            return_dict=False,
                            **extra_template_kwargs,
                        )
                    )
                    prompt_ids = self.tokenizer_manager.tokenizer.encode(
                        rendered_prompt, **encode_kwargs
                    )
                except _CHAT_TEMPLATE_CLIENT_ERRORS as template_error:
                    # Template errors (e.g., from raise_exception in Jinja templates)
                    # and TypeError (e.g., tojson filter on Jinja2 Undefined variables)
                    # should be treated as client errors (400 BadRequest)
                    raise ValueError(str(template_error)) from template_error

            # Append assistant prefix if continue_final_message is enabled
            if assistant_prefix:
                prompt_ids = self._append_assistant_prefix_to_prompt_ids(
                    prompt_ids, assistant_prefix
                )

            if is_multimodal:
                prompt = self.tokenizer_manager.tokenizer.decode(prompt_ids)

        stop = request.stop
        image_data = image_data if image_data else None
        audio_data = audio_data if audio_data else None
        video_data = video_data if video_data else None
        modalities = modalities if modalities else []
        return MessageProcessingResult(
            prompt=prompt,
            prompt_ids=prompt_ids,

View on GitHub (pinned to 0132848349)

Solutions

1. Fix the request payload: ensure message roles/content shapes match what the model's chat template expects (e.g. include tool role support)
2. Inspect the wrapped template_error message (raise __cause__) to find the exact template line, then fix the custom chat template
3. If passing chat_template_kwargs, verify the template actually uses those variable names
4. Retry with a stock template / remove custom --chat-template to isolate the issue

Example fix

# before
curl /v1/chat/completions -d '{"model":"m","messages":[{"role":"tool","content":"42"}]}'  # template without tool support
# after
curl /v1/chat/completions -d '{"model":"m","messages":[{"role":"user","content":"hi"}]}'

required = {r.value for r in collections if r.role in ("system","user","assistant","tool")}
assert all(m.get("role") in required for m in messages)
assert all(not kw or kw for kw in (chat_template_kwargs or {}))

try: client.chat.completions.create(...)
except BadRequestError as e:
    if 'template' in str(e): fix messages/template kwargs

When it happens

Trigger: POST /v1/chat/completions where the model's chat template raises: invalid messages structure for the template (e.g. missing required roles), passing chat_template_kwargs the template doesn't expect, a tojson filter applied to an undefined variable, or a template that explicitly calls raise_exception on bad input.

Common situations: Custom --chat-template that assumes specific message shapes; multi-turn conversations with tool messages fed to a template lacking tool support; wrong or missing tokenizer_config chat_template; passing unsupported chat_template_kwargs.

Related errors

• Assistant tool call function.arguments must be valid JSON.
• {template_error}{suffix}
• Failed to render chat template for embedding input: {templat
• Invalid content format: {content_format}
• Invalid request body: {e}

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/f7dfbc5db4fe8d64. Report an issue: GitHub.