chat_template.jinja

{#- ======== Template Parameters ========  #}
{%- set add_generation_prompt = add_generation_prompt if add_generation_prompt is defined else true %}
{%- set provider_system_prompt = provider_system_prompt if provider_system_prompt else false %}
{%- set reasoning_effort = reasoning_effort if reasoning_effort is defined else "high" %}
{%- set think_render_option = think_render_option if think_render_option is defined else "interleaved" %}

{#- ======== System Block State ========  #}
{%- set sys_ns = namespace(is_first_block=true) -%}

{#- ======== Find last user message index ========  #}
{%- set last_user_idx = namespace(value=-1) -%}
{%- for message in messages -%}
    {%- if message.role == 'user' -%}
        {%- set last_user_idx.value = loop.index0 -%}
    {%- endif -%}
{%- endfor -%}

{#- ======== System messages renderers ========  #}
{%- macro render_system_message(user_system_messages) %}
    {%- if provider_system_prompt %}
        {%- if not sys_ns.is_first_block %}{{- "\n\n" }}{%- endif %}
        {%- set sys_ns.is_first_block = false %}
        {{- "## Provider System Prompt\n\nYou are Solar Open2 250B, a large language model trained by Upstage AI, a Korean startup. Your knowledge cutoff is 2026-02. The current date is " + strftime_now("%Y-%m-%d") + "." }}
    {%- endif -%}
    {%- if user_system_messages %}
        {%- if not sys_ns.is_first_block %}{{- "\n\n" }}{%- endif %}
        {%- set sys_ns.is_first_block = false %}
        {{- "## System Prompt" }}
        {%- for system_message in user_system_messages %}
            {{- "\n\n" }}
            {{- system_message }}
        {%- endfor %}
    {%- endif -%}
{%- endmacro %}

{%- macro render_tool_instruction(tools, is_last_block) %}
    {%- if not sys_ns.is_first_block %}{{- "\n\n" }}{%- endif %}
    {%- set sys_ns.is_first_block = false %}
    {{- "## Tools\n- You may invoke one or more tools to assist with the user's query." }}
    {{- "\n\n### Available Tools\n" }}
    {%- for tool in tools %}
        {{- "<|tool:start|>" }}
        {{- tool.function | tojson }}
        {{- "<|tool:end|>\n" }}
    {%- endfor %}
    {{- "\n### Tool Call Instruction\n" }}
    {{- "- If using a tool, any reasoning must strictly precede the call. Do not append any text after the tool call.\n" }}
    {{- "- If no tool is required, answer directly from your knowledge without ever mentioning the availability or absence of tools.\n" }}
    {{- "- Each tool call MUST use this following format: <|tool_call:start|>{example-tool-name}\n<|tool_arg:start|>{example-key-name-1}<|tool_arg:value|>{example-value-1}<|tool_arg:end|>\n<|tool_arg:start|>{example-key-name-2}<|tool_arg:value|>{example-value-2}<|tool_arg:end|>\n<|tool_call:end|>\n" }}
{%- endmacro %}

{#-
Input convention for `response_format`:
- Per the OpenAI Chat Completions API, `response_format` is an object of one of:
    {"type": "json_schema", "json_schema": {"name": ..., "schema": {...}, "strict": ...}}
    {"type": "json_object"}
- Only `response_format.json_schema.schema` (the inner JSON Schema) is injected
  into the prompt; the OpenAI wrapper fields (`name`, `strict`, ...) are dropped.
- For `json_object`, no schema exists, so we emit a generic JSON-only instruction.
- vLLM does NOT pass `response_format` to chat templates by default; it routes it
  to guided decoding instead. This template/whale opts into prompt injection,
  so callers serving the model directly via vLLM must opt in via
  `chat_template_kwargs` (or rely on whale's normalization).
-#}
{%- macro render_json_response_format_instruction(response_format) %}
    {%- if not sys_ns.is_first_block %}{{- "\n\n" }}{%- endif %}
    {%- set sys_ns.is_first_block = false %}
    {{- "## Output Format Constraint" }}
    {%- if response_format.type == "json_schema" %}
        {{- "\n\n- Your final response should follow the JSON schema:\n```json\n" }}
        {{- response_format.json_schema.schema | tojson }}
        {{- "\n```\n- Please ensure your answers adhere to this format and do not contain any unnecessary text." }}
    {%- elif response_format.type == "json_object" %}
        {{- "\n\n- Your final response must be a valid JSON object." }}
        {{- "\n- Do not include any text outside the JSON object." }}
    {%- endif %}
{%- endmacro %}

{#-
Input convention for `tool_arguments`:
- Must be a mapping (dict), NOT a JSON-encoded string.
- The OpenAI API spec defines `tool_calls[].function.arguments` as a JSON string,
  but vLLM, HuggingFace transformers tool-use templates (Qwen, Hermes, Llama, ...),
  and this template all expect the value to be parsed into a dict before rendering.
- vLLM normalizes this in `_postprocess_messages()` (vllm/entrypoints/chat_utils.py).
- Callers (whale, training pipelines that call `tokenizer.apply_chat_template`
  directly, ...) are responsible for parsing the JSON string into a dict before
  invoking the template.
-#}
{%- macro render_tool_arguments(tool_arguments) %}
    {%- for args_name, args_value in tool_arguments|items %}
        {{- "<|tool_arg:start|>"+ args_name + "<|tool_arg:value|>"}}
        {%- set args_value = args_value if args_value is string else (args_value | tojson | safe) %}
        {{- args_value }}
        {{- "<|tool_arg:end|>\n" }}
    {%- endfor %}
{%- endmacro %}

{#- ======== Render system message ========  #}
{%- set ns = namespace(system_messages=[]) -%}
{%- for message in messages -%}
    {%- if message.role == 'system' -%}
        {%- set ns.system_messages = ns.system_messages + [message.content] -%}
    {%- endif -%}
{%- endfor -%}

{%- if ns.system_messages or provider_system_prompt or tools or response_format -%}
    {{- "<|im:start|>system<|im:content|>" }}
        {{- render_system_message(ns.system_messages) }}
        {%- if tools -%}
            {{- render_tool_instruction(tools, not response_format) }}
        {%- endif %}
        {%- if response_format -%}
            {{- render_json_response_format_instruction(response_format) }}
        {%- endif %}
    {{- "<|im:end|>\n" }}
{%- endif -%}

{#- ======== Render main messages ========  #}
{%- for message in messages -%}
    {%- if message.role == 'user' -%}
         {{- "<|im:start|>user<|im:content|>" + message.content + "<|im:end|>\n" }}
    {%- elif message.role == 'tool' -%}
        {%- set prev_is_tool = loop.index0 > 0 and messages[loop.index0 - 1].role == 'tool' -%}
        {#- Only render at the start of a contiguous tool block; subsequent tool messages -#}
        {#- are already emitted (reordered) within the first one. -#}
        {%- if not prev_is_tool -%}
            {%- set start_idx = loop.index0 -%}
            {#- Collect contiguous tool messages starting at start_idx -#}
            {%- set tool_ns = namespace(items=[], scanning=true) -%}
            {%- for j in range(start_idx, messages | length) -%}
                {%- if tool_ns.scanning -%}
                    {%- if messages[j].role == 'tool' -%}
                        {%- set tool_ns.items = tool_ns.items + [messages[j]] -%}
                    {%- else -%}
                        {%- set tool_ns.scanning = false -%}
                    {%- endif -%}
                {%- endif -%}
            {%- endfor -%}

            {{- "<|im:start|>tool<|im:content|>" }}

            {%- set prev_msg = messages[start_idx - 1] if start_idx > 0 else none -%}
            {%- if prev_msg and prev_msg.role == 'assistant' and prev_msg.tool_calls -%}
                {#- Render tool responses in the order of the preceding assistant's tool_calls ids. -#}
                {%- set sep_ns = namespace(first=true, rendered_ids=[]) -%}
                {%- for tool_call in prev_msg.tool_calls -%}
                    {%- for tool_msg in tool_ns.items -%}
                        {%- if tool_msg.tool_call_id == tool_call.id -%}
                            {%- if not sep_ns.first -%}{{- "\n" }}{%- endif -%}
                            {%- set sep_ns.first = false -%}
                            {%- set sep_ns.rendered_ids = sep_ns.rendered_ids + [tool_msg.tool_call_id] -%}
                            {{- "<|tool_response:start|>" + tool_msg.content + "<|tool_response:end|>" }}
                        {%- endif -%}
                    {%- endfor -%}
                {%- endfor -%}
                {#- Append any orphan tool messages (no matching id) in their original order. -#}
                {%- for tool_msg in tool_ns.items -%}
                    {%- if tool_msg.tool_call_id not in sep_ns.rendered_ids -%}
                        {%- if not sep_ns.first -%}{{- "\n" }}{%- endif -%}
                        {%- set sep_ns.first = false -%}
                        {{- "<|tool_response:start|>" + tool_msg.content + "<|tool_response:end|>" }}
                    {%- endif -%}
                {%- endfor -%}
            {%- else -%}
                {%- for tool_msg in tool_ns.items -%}
                    {%- if not loop.first -%}{{- "\n" }}{%- endif -%}
                    {{- "<|tool_response:start|>" + tool_msg.content + "<|tool_response:end|>" }}
                {%- endfor -%}
            {%- endif -%}

            {{- "\n<|im:end|>\n" }}
        {%- endif -%}
    {%- elif message.role == 'assistant' -%}
        {{- "<|im:start|>assistant<|im:content|>" }}

        {#- ======== Resolve reasoning field (reasoning or reasoning_content) ========  #}
        {%- if message.reasoning is defined -%}
            {%- set reasoning = message.reasoning -%}
        {%- elif message.reasoning_content is defined -%}
            {%- set reasoning = message.reasoning_content -%}
        {%- else -%}
            {%- set reasoning = none -%}
        {%- endif -%}

        {#- ======== Assistant Thinking ========  #}
        {%- if think_render_option == "preserved" -%}
            {%- if reasoning -%}
                {{- "<|think:start|>" + reasoning + "<|think:end|>" }}
            {%- else -%}
                {{- "<|think:start|><|think:end|>" }}
            {%- endif -%}
        {%- elif think_render_option == "interleaved" -%}
            {%- if reasoning and loop.index0 > last_user_idx.value -%}
                {{- "<|think:start|>" + reasoning + "<|think:end|>" }}
            {%- else -%}
                {{- "<|think:start|><|think:end|>" }}
            {%- endif -%}
        {%- endif -%}

        {#- ======== Assistant Messages ========  #}
        {%- if message.content -%}
            {{- message.content }}
        {%- endif -%}

        {#- ======== Assistant Tool calls ========  #}
        {%- if message.tool_calls -%}
            {%- for tool_call in message.tool_calls -%}
                {%- if not loop.first -%}{{- "\n" }}{%- endif -%}
                {{- "<|tool_call:start|>" + tool_call.function.name + "\n" }}
                {#- ======== FIXME: QWEN3.5 Method ========  #}
                {%- if tool_call.function.arguments is defined %}
                    {{- render_tool_arguments(tool_call.function.arguments) }}
                {%- endif %}
                {{- "<|tool_call:end|>" }}
            {%- endfor -%}
            {{- "\n" }}
        {%- endif -%}
        {{- "<|im:end|>\n" }}
    {%- endif -%}
{%- endfor -%}

{%- if add_generation_prompt -%}
    {%- if reasoning_effort in ["medium", "high"] -%}
        {{- "<|im:start|>assistant<|im:content|><|think:start|>" }}
    {%- else -%}
        {{- "<|im:start|>assistant<|im:content|><|think:start|><|think:end|>" }}
    {%- endif -%}
{%- endif -%}