statsui

A local dashboard for llama-server: live token throughput, time-to-first-token, prefill speed, and context usage, tracked per named session.

Read-only against llama-server, it polls /slots and never touches inference. Safe to run while you're actively chatting.

Install

npm install -g statsui

Usage

statsui

First run walks you through setup (port, llama-server URL), then starts the dashboard at http://127.0.0.1:8079 (or whatever port you chose). Later runs reuse the saved config, run statsui configure to change it.

Tip: start llama-server with --metrics for extra Prometheus-backed stats, statsui works fine without it too.

Tuning llama-server flags

statsui llama --model "/path/to/model.gguf" --c 32768

Reads the GGUF file's real metadata and tensor list (architecture, trained context length, layer count, and whether it has an MTP/speculative-decoding head), detects your GPU VRAM, CPU threads, and RAM, and prints a llama-server command tuned for low TTFT and high throughput, along with the reasoning behind every flag. Nothing launches until you confirm.

What it shows

• Input processed / cached, output generated, live and averaged
• Time to first token, tracked against context size
• Prefill (prompt-processing) speed, separated from generation speed
• Generation speed, tracked against context size
• Estimated API cost avoided, at prices you set
• Per-session history: rename, reset counters, or start a new one from the sidebar

License

MIT