Prompt Jinja2 not parsing
Thank you for making Qwen3 available to MLX community.
I was able to run Qwen3-30B-A3B-6bit on mlx_lm.server without problem. But when i tried to load it in LM Studio, I ran into some issues. I'm not sure why it just works in mlx_lm.server and with mlx server instance on LM Studio it complains about the Jinja2 prompt.
After reviewing the Jinja2 prompt with a Jinja2 parser, I found multiple parsing issues. I spent a few hours to debug it and fix the parsing, while trying to not modify the initial logic and intent. The change I had to do were significantly extensive, that I'm not 100% sure if I were able to preserve fully the initial logic and intent.
I would greatly appreciate if you could review this version that successfully parse and load in LM Studio and if you feel it pass your validation, I would invite to update it, so people can enjoy using it in LM Studio and other tools.
This version of the prompt parse succesfully and load successfully in LM Studio.
{%- if tools is defined and tools %}
{{- '<|im_start|>system\n' }}
{%- if messages is defined and messages|length > 0 and messages[0].role == 'system' %}
{{- messages[0].content + '\n\n' }}
{%- endif %}
{{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
{%- for tool in tools %}
{{- "\n" }}
{{- tool | tojson }}
{%- endfor %}
{{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
{%- else %}
{%- if messages is defined and messages|length > 0 and messages[0].role == 'system' %}
{{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
{%- endif %}
{%- endif %}
{# Set up namespace for tracking multi-step tools #}
{%- set ns = namespace(multi_step_tool=true, last_query_index=0) %}
{%- if messages is defined %}
{%- set ns.last_query_index = messages|length - 1 %}
{# First pass - find the last non-tool-response user query #}
{%- for i in range(messages|length - 1, -1, -1) %}
{%- set message = messages[i] %}
{%- if ns.multi_step_tool and message.role == "user" and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
{%- set ns.multi_step_tool = false %}
{%- set ns.last_query_index = i %}
{%- endif %}
{%- endfor %}
{# Main message processing loop #}
{%- for i in range(messages|length) %}
{%- set message = messages[i] %}
{%- if (message.role == "user") or (message.role == "system" and i > 0) %}
{{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
{%- elif message.role == "assistant" %}
{%- set content = message.content %}
{%- set reasoning_content = '' %}
{%- if message.reasoning_content is defined and message.reasoning_content is not none %}
{%- set reasoning_content = message.reasoning_content %}
{%- else %}
{%- if '</think>' in message.content %}
{%- set content_parts = message.content.split('</think>') %}
{%- if content_parts|length > 1 %}
{%- set content = content_parts[-1].lstrip('\n') %}
{%- set thinking_parts = content_parts[0].split('<think>') %}
{%- if thinking_parts|length > 1 %}
{%- set reasoning_content = thinking_parts[-1].lstrip('\n') %}
{%- endif %}
{%- endif %}
{%- endif %}
{%- endif %}
{%- if i > ns.last_query_index %}
{%- if loop.last or (not loop.last and reasoning_content) %}
{{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
{%- else %}
{{- '<|im_start|>' + message.role + '\n' + content }}
{%- endif %}
{%- else %}
{{- '<|im_start|>' + message.role + '\n' + content }}
{%- endif %}
{%- if message.tool_calls is defined and 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 is defined %}
{%- set function_call = tool_call.function %}
{%- else %}
{%- set function_call = tool_call %}
{%- endif %}
{{- '<tool_call>\n{\"name\": \"' }}
{{- function_call.name }}
{{- '\", \"arguments\": ' }}
{%- if function_call.arguments is string %}
{{- function_call.arguments }}
{%- else %}
{{- function_call.arguments | tojson }}
{%- endif %}
{{- '}\n</tool_call>' }}
{%- endfor %}
{%- endif %}
{{- '<|im_end|>\n' }}
{%- elif message.role == "tool" %}
{%- if i == 0 or (messages[i-1].role != "tool") %}
{{- '<|im_start|>user' }}
{%- endif %}
{{- '\n<tool_response>\n' }}
{{- message.content }}
{{- '\n</tool_response>' }}
{%- if i == messages|length - 1 or (messages[i+1].role != "tool") %}
{{- '<|im_end|>\n' }}
{%- endif %}
{%- endif %}
{%- endfor %}
{%- endif %}
{%- if add_generation_prompt is defined and add_generation_prompt %}
{{- '<|im_start|>assistant\n' }}
{%- if enable_thinking is defined and enable_thinking is false %}
{{- '<think>\n\n</think>\n\n' }}
{%- endif %}
{%- endif %}
I just found the other thread on the 8bit model.
It would seems those Jinja2 are generated and produced by LM Studio. For next update, I believe they will have fixed the Qwen Jinja template. But I still have numerous issue with other models from mlx-community. I will have to raise this issue with LM Studio directly.