[BUG] Problem with chat template for new MTP models.

[nigelp]

System details / 系统信息

Windows 10/ Deepchat 1.0.3/Intel i9 14900

What's wrong? / 出了什么问题？

The chat function doesn't respond properly with local Qwen3.6 27B model with MTP. It responds with a broken answer. See image.

[zerob13]

Please enable logging, reproduce the issue, and then upload the log file to the issue. Thanks!

[nigelp]

Thanks very much. Here you go.

main.log

[zerob13]

From the logs, there’s no clear error message. It’s likely caused by the upstream model returning an invalid or non-standard format, so it’s currently quite difficult to pinpoint the exact issue.

[nigelp]

Looks like the MTP tech is breaking something.

[nigelp]

It would be really good if you could fix this because there are other harnesses which can do it, but I'd like to try using deepchat for most of my local coding and and agentic work now. But I need this MTP model because it's fast and superb.

https://old.reddit.com/r/LocalLLaMA/comments/1t57xuu/25x_faster_inference_with_qwen_36_27b_using_mtp/

[zerob13]

Is there anywhere I can try APIs for this kind of model? I currently don’t have an environment capable of running models like this.

Or alternatively, is there a way to get the raw streamed response data from this model via curl? I’d like to inspect what the actual text part looks like in the original stream output.

[nigelp]

Is this any good?

mtpstream.txt

[edit: and here's a longer one]

mtpstreamlong.txt

[zerob13]

Please check whether the custom provider type is set to OpenAI Completions. If you're using the default OpenAI provider, it will use the latest Responses API. From what I've researched, llama.cpp does not seem to be compatible with it yet.

[nigelp]

:) yep that fixed it. Thanks very much.