chat template issue qwen3.8-27B

[amitpandey-io]

using latest gmlx
gguf is 'qwen3.8-27B-bf16-mtp.gguf' ( has mtp weights along with it)

2026-08-22 19:27:45 WARNING:  /v1/chat/completions: ignoring unsupported parameter(s): max_completion_tokens, store (see the parameter support table in docs/server-config.md)
Traceback (most recent call last):
  File "/Users/user/.local/share/uv/tools/gmlx/lib/python3.13/site-packages/mlx_vlm/server/openai.py", line 1564, in chat_completions_endpoint
    formatted_prompt = apply_chat_template(
        processor,
    ...<6 lines>...
        **gen_args.to_template_kwargs(),
    )
  File "/Users/user/.local/share/uv/tools/gmlx/lib/python3.13/site-packages/gmlx/server_patches/chat_behavior.py", line 335, in apply_chat_template
    out = fn(*a, **kw)
  File "/Users/user/.local/share/uv/tools/gmlx/lib/python3.13/site-packages/gmlx/server_patches/apc.py", line 102, in apply_chat_template
    out = orig_render(processor, config, prompt, *a, **kw)
  File "/Users/user/.local/share/uv/tools/gmlx/lib/python3.13/site-packages/gmlx/server_patches/render.py", line 123, in apply_chat_template
    return prompt_utils.get_chat_template(
           ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^
        processor, msgs, add_generation_prompt, **kwargs)
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/user/.local/share/uv/tools/gmlx/lib/python3.13/site-packages/mlx_vlm/prompt_utils.py", line 780, in get_chat_template
    return template_processor.apply_chat_template(
           ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^
        messages,
        ^^^^^^^^^
    ...<2 lines>...
        **template_kwargs,
        ^^^^^^^^^^^^^^^^^^
    )
    ^
  File "/Users/user/.local/share/uv/tools/gmlx/lib/python3.13/site-packages/transformers/processing_utils.py", line 2138, in apply_chat_template
    prompt, generation_indices = render_jinja_template(
                                 ~~~~~~~~~~~~~~~~~~~~~^
        conversations=conversations,
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    ...<6 lines>...
        **template_kwargs,
        ^^^^^^^^^^^^^^^^^^
    )
    ^
  File "/Users/user/.local/share/uv/tools/gmlx/lib/python3.13/site-packages/transformers/utils/chat_template_utils.py", line 581, in render_jinja_template
    rendered_chat = compiled_template.render(
        messages=chat,
    ...<3 lines>...
        **kwargs,
    )
  File "/Users/user/.local/share/uv/tools/gmlx/lib/python3.13/site-packages/jinja2/environment.py", line 1295, in render
    self.environment.handle_exception()
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
  File "/Users/user/.local/share/uv/tools/gmlx/lib/python3.13/site-packages/jinja2/environment.py", line 942, in handle_exception
    raise rewrite_traceback_stack(source=source)
  File "<template>", line 160, in top-level template code
  File "/Users/user/.local/share/uv/tools/gmlx/lib/python3.13/site-packages/jinja2/sandbox.py", line 401, in call
    return __context.call(__obj, *args, **kwargs)
           ~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/user/.local/share/uv/tools/gmlx/lib/python3.13/site-packages/transformers/utils/chat_template_utils.py", line 479, in raise_exception
    raise jinja2.exceptions.TemplateError(message)
jinja2.exceptions.TemplateError: Unexpected message role.

[asher]

Thanks for the report, can you still reproduce on the just released v0.4.0? If so, please point me to the gguf or chat template it uses (the two versions of qwen3.8-27B ggufs I have locally seem ok)

[asher]

Make that 0.4.1, which contains a fix that might cover your case. If a request contains parameters for the developer role but the chat template doesn't define one, it's mapped to system. Exceptions raised from a chat template now also make it to the server response, so if you're still hitting this, the error should be more helpful. If it still fails on 0.4.1, can you share which client/harness sent the request, the server response, and point me to the chat template or gguf?