Skip to main content

​
Performance Analysis Plotter

Visualize llama.cpp telemetry by turning one or more server logs into plots for prompt tokens, input tokens, output tokens, time-to-first-token (TTFT), tokens-per-second (TPS), and a prefill vs decode time split.

​
Requirements

  • Python 3.8+
  • matplotlib (pip install matplotlib)

​
Run the Plotter (GAIA CLI)

gaia perf-vis <log_file> [<log_file> ...]
  • Pass multiple log files to compare runs; each plot adds one line per log with a legend.

​
Collecting llama.cpp Logs

The script expects llama.cpp server logs. With Lemonade, you can capture telemetry like this: 
lemonade-server serve --ctx-size 32768 2>&1 | tee agent.log
gaia perf-vis agent.log

​
Outputs

Images are written to the directory where you run the script:
  • prompt_token_counts.png — prompt token totals per call
  • input_token_counts.png — input token counts
  • output_token_counts.png — output token counts
  • ttft_seconds.png — time to first token
  • tps.png — tokens per second
  • prefill_decode_split.png — one pie per log showing prefill (TTFT) vs decode (output tokens / TPS) time
When multiple logs are provided, every plot includes one line/pie per log plus legends mapping each series to its log filename.