# Banc d'essai inférence GPU — DaniiT
Download: `mkdir -p ~/bench && cd ~/bench && for f in bench.py run_all.sh test1.jpg prompts/test1_vision.txt prompts/test2_static_site.txt prompts/test3_django.txt; do curl -fsSL --create-dirs -o $f https://inference.daniit-software.by/bench/$f; done && chmod +x run_all.sh`
Run: `./run_all.sh` (pulls the 4 models, ~33 GB). Result: `results/<hostname>-<date>.json`, raw outputs in `results/raw-*`.
Publish: `scp results/*.json lxc-139:/var/www/inference.daniit-software.by/results/` then add the file name to `results/index.json`.
Protocol: temperature 0, seed 42, num_ctx 8192, num_predict 6144; 1 cold run (after unload) + 3 warm runs, median of warm runs.