TL;DR

Update the Qwen3.5 chat template to use tojson for all non-string argument values, ensuring JSON literals are produced instead of Python str() representations.

Guidance

• Identify the current Qwen3.5 chat template and update the line {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %} to {%- set args_value = args_value if args_value is string else args_value | tojson | safe %} to correctly handle all non-string arguments.
• Verify that the updated template produces the expected JSON literals for all argument types, including None which should be converted to "null".
• Test the updated template with the VLLM tool parser to ensure it can correctly parse the output.
• Review other chat templates for similar issues to prevent future problems.

Example

The corrected Qwen3.5 chat template should resemble the Qwen3 coder chat template:

{%- set args_value = args_value if args_value is string else args_value | tojson | safe %}

Notes

This fix assumes that the tojson filter is correctly configured to produce JSON literals for all argument types. If issues persist, further investigation into the tojson filter or the VLLM tool parser may be necessary.

Recommendation

Apply the workaround by updating the Qwen3.5 chat template to use the corrected tojson handling, as this should resolve the issue with the VLLM tool parser.