bench-console

Simple TUI for benchmarking local GGUF models: edit all llama-server and llama-benchy parameters in one form, press Run, and the tool:

1. Automatically stops llama.service if it's running, since the GPU only has room for one model at a time.
2. Starts llama-server with your parameters.
3. Waits until the server responds on /health.
4. Writes a header with all the llama.cpp parameters at the top of the log window.
5. Automatically starts llama-benchy against it, with results scrolling down in the same log window.
6. Saves the full transcript (header + server log + benchy results) to results/<preset>_<timestamp>.txt for easy comparison later.

llama.service is not automatically restarted afterward — run systemctl --user start llama.service when you're done benchmarking.

Running

cd ~/bench-console
.venv/bin/bench-console
# or: uv run bench-console

Presets

Settings are saved as JSON files under presets/. Pick a preset from the dropdown at the top — selecting it loads it automatically — or save the form you've edited by typing a name in "Save as" and pressing Ctrl+S. presets/example.json is a starting point.

Shortcuts

• Ctrl+R: Run
• Ctrl+X: Stop (kills any running processes)
• Ctrl+S: Save preset
• q: Quit

Environment variables

• LLAMA_SERVER_BIN: path to llama-server (default: ~/ai/llama.cpp/build/bin/llama-server)
• BENCHY_BIN: path to llama-benchy (default: ~/ai/llama-benchy/.venv/bin/llama-benchy)
• BENCH_CONSOLE_READY_TIMEOUT: seconds to wait for the server to become ready (default 180)