How do I improve my SWEBench run on 3090?

I am running SWEBench locally with llama.cpp and unsloth's Qwen3.8-27B using the following server command: ./llama-server -m ~/gguf/Qwen3.8-27B-IQ4_NL.gguf --jinja -ngl 999 -fa 1 --load-mode mmap -np 1 --temp 1.0 --top_p 0.95 --top_k 20 --min_p 0.0 --presence-penalty 0.0 --repeat-penalty 1.0 -c 2293