Eval bug: How to use Gemma 31B Qat with mtp?

[ajcn02]

Name and Version

Hi, I'm trying to use Gemma 31B with the MTP assistant but I'm having trouble.

Operating systems

Windows

GGML backends

CUDA

Hardware

i5 13600KF and RTX 4090

Models

No response

Problem description & steps to reproduce

llama-server.exe ^
-m "C:\Users\ajcn2\Desktop\Lamma.cpp\Models\gemma-4-31B-it-qat-UD-Q4_K_XL.gguf" ^
--model-draft "C:\Users\ajcn2\Desktop\Lamma.cpp\Models\gemma-4-31b-it-qat-q4_0-assistant.gguf" ^
--spec-type draft-mtp ^
--spec-draft-n-max 2 ^
--mmproj "C:\Users\ajcn2\Desktop\Lamma.cpp\Models\mmproj-BF16.gguf" ^
--no-mmproj-offload ^
--chat-template-file "C:\Users\ajcn2\Desktop\Lamma.cpp\Models\custom_pub_chat_template_gemma4.jinja" ^
--port 5051 ^
-np 1 ^
-ngl 99 ^
-ngld 99 ^
-b 2048 -ub 512 ^
--ctx-size 32768 ^
--cache-type-k q8_0 --cache-type-v q8_0 ^
--flash-attn on ^
--jinja ^
--no-mmap --mlock ^
--no-host ^
--reasoning on ^
--temp 1.0 --top-k 64 --top-p 0.95 --min-p 0.0

First Bad Commit

No response

Relevant log output

Logs