`clip_model_path` should not be a mandatory parameter for `Qwen35ChatHandler`.

[Moon-404]

Is your feature request related to a problem? Please describe.
Qwen35ChatHandler currently requires the clip_model_path argument. This makes it impossible to use the handler for text-only use cases without providing a vision encoder. Although Qwen3.5 is a multimodal model, many real-world scenarios are strictly text-only, and in those cases loading a CLIP encoder is unnecessary overhead. In practice, even without specifying any chat_handler, Qwen3.5 can produce reasonable text outputs, but it appears to force “thinking mode” by default.

Describe the solution you'd like
clip_model_path should not be a mandatory parameter for Qwen35ChatHandler. It should be optional, and the handler should support a text-only mode where no vision encoder is loaded.

Describe alternatives you've considered
It should be possible to control whether the “think” template is enabled on a per-request basis (e.g., at create_chat_completion time), rather than requiring a full model reload to switch modes. Ideally, this would be exposed via something like chat_template_kwargs={"enable_thinking": False} (similar to the official behavior).

Additional context
No.

[JamePeng]

If you don't need to use multimodal processing, you don't need to use a chathandler to load qwen3.5; you can use llama directly without passing in Qwen35ChatHandler.

[JamePeng]

pseudo code：

model_path = "path\\to\\Qwen3.5-35B-A3B-MXFP4_MOE.gguf"
llm = Llama(model_path=model_path, n_ctx=40960, n_gpu_layers=66, verbose=True)
response = llm.create_chat_completion(
            messages=prompt,
            max_tokens=max_length,
            stop=["<|im_end|>"],
            temperature=temperature,
            top_k=top_k,
            top_p=top_p,
            repeat_penalty=repeat_penalty,
            stream=True
        )

[JamePeng]

It should be possible to control whether the “think” template is enabled on a per-request basis (e.g., at create_chat_completion time), rather than requiring a full model reload to switch modes. Ideally, this would be exposed via something like chat_template_kwargs={"enable_thinking": False} (similar to the official behavior).

This is a good request, but the Jinja variables for activating the thinking mode differ across models, and the methods also vary. Therefore, currently each multimodal model is specifically adapted for this purpose. I will also take a look at how llama.cpp uses this functionality.

[Moon-404]

I know I can use directly without passing in Qwen35ChatHandler. What matters is that in this condition thinking mode is forcely active. If I want to disable thinking mode, I have to use Qwen35ChatHandler now.

So if there is another way to disable thinking mode without using Qwen35ChatHandler, this issue is no longer needed.

[JamePeng]

So you mean to transform Qwen35ChatHandler into a text-only function? And then use enable_thinking to control whether the thought chain is output.

[Moon-404]

My opinion:

1. whether thinking should be decided in create_chat_completion, instead of chat_handler.
2. Maybe Qwen35ChatHandler can split into different handlers, or clip_model_path could be None for text-only.

[JamePeng]

My latest code removes the exception that forces the vision model, but it still requires loading the clip model to ensure that mtmd_ctx functions correctly. Now you can use the enable_thinking tag in chat_template normally without uploading images.