Eval bug: When I run Gemma 4e4b with MTP, it doesn't work and crashes to cmd #26017

Name and Version

I'm using llama.cpp b10090.

Operating systems

Windows

GGML backends

CUDA

Hardware

5 5600g and gtx 1660 super

Models

gemma-4-E4B-it-UD-Q4_K_XL and mtp-gemma-4-E4B-it

Problem description & steps to reproduce

it doesn't work with either CUDA 12 or 13 but vulkan work. I'm using this configuration: llama-cpp\llama-server.exe -m gemma-4-e4b\gemma-4-E4B-it-UD-Q4_K_XL.gguf -md gemma-4-e4b\mtp-gemma-4-E4B-it.gguf --fit off --reasoning on --n-gpu-layers all --ctx-size 131072 --spec-type draft-mtp --spec-draft-n-max 2. I've checked it several times, but there doesn't seem to be a problem with my configuration. I just downloaded those two models and applied that configuration.

First Bad Commit

I got an error when testing those models for the first time in version b10088

Relevant log output