This repository contains a bash script to run LLMs using llama.cpp.

Installation Instructions for MoE Model Script on llama.cpp Compiled with SYCL Support.

To install, clone the repository: git clone https://github.com/PochtiPlusovik/llama-scripts. Make the script executable: cd ~/llama-scripts && chmod a+x runner-MoE-Intel-Arc.sh. Change the model path on line 4 of the script using a text editor. You can view the list of installed models with the following command: sudo find / -name "*.gguf". Run the script: ./runner-MoE-Intel-Arc.sh.

Installation of a Script for Dataset Generation.

Attention: After generation, you will need to clean the dataset from the standard llama.cpp output. You can uncomment --reasoning-budget 0, but the generation quality may degrade. The --reasoning-mode none parameter must be commented out in the script when using LLMs that output Thinking Mode as a standard response, such as Gemma-4 26B A4B. To install, clone the repository: git clone https://github.com/PochtiPlusovik/llama-scripts. In the script, you are required to update the variables LLM, LANGUAGE, INPUTFILE, and OUTPUTFILE. You can view the list of installed models with the following command: sudo find / -name "*.gguf". The INPUTFILE file must include the topics for generation, each starting on a new line. For MoE models, uncomment --no-mmap. If you are using SYCL, enter: source /opt/intel/oneapi/setvars.sh. Make the script executable: cd ~/llama-scripts && chmod a+x generate-dataset.sh. Run the script in ~/llama-scripts: ./generate-dataset.sh.

The scripts were tested on Arch Linux and Intel Arc B580 with the Qwen 3.6-A3B model. These scripts should work on other backends and graphics cards, but testing has not been performed