Eval bug: Preset (models.ini) is slower then single model

[dejan-vasic]

Name and Version

./llama-cli --version
version: 10465 (fca3093)
built with GNU 11.4.0 for Linux x86_64

Operating systems

Linux

GGML backends

CUDA

Hardware

Quadro P4000 8GB

Models

Qwen3.6-35B-A3B-UD-Q4_K_M.gguf

Problem description & steps to reproduce

When I run llama turboquant with perset (models.ini) for model above I get 17 t/s but when I run it only for that specific model I get 24 t/s

First Bad Commit

No response

Relevant log output

Logs

[giveen]

Can you show us the output and logs? Same prompt?