Local LLM benchmarks

scripts/benchmark_local_llm.py benchmarks an already-installed Ollama model using a fictional duplicate-adjudication workload. It never downloads models, starts Ollama, changes persistent settings, or saves prompts/responses. Nothing contacts Ollama until --execute is supplied.

Plan a run without network activity:

python scripts/benchmark_local_llm.py --model existing-model

Run a first-request (observed cold) measurement plus one warm measurement:

python scripts/benchmark_local_llm.py --model existing-model --execute \
  --output /tmp/ollama-benchmark.json

The JSON report is aggregate-only: selected profile and its request options, model metadata, resident model memory when Ollama makes it available, benchmark process peak RSS when the platform exposes it, and per-request wall time, TTFT, Ollama load/total time, and prompt/completion token rates. It contains no prompt, generated response, or endpoint value. --output refuses paths inside this repository so reports cannot be accidentally committed.

cold means the first observed benchmark request by default; an already-loaded model may therefore still be warm at the Ollama level. To explicitly request an unload before that measurement, use --unload-before. This makes a local Ollama API call but does not download or remove the model. --unload-after requests the same unload once measurements finish and records whether it succeeded.

For operational behavior checks, --warm-runs N adds warm requests, --queue-depth N submits an additional concurrent batch and records each worker’s local queue delay, and --cancel-after-first-token closes each stream after its first generated token and records cancelled. --timeout-seconds N sets the HTTP timeout; a timed-out request remains in the JSON report with a timeout status. A missing endpoint or unavailable model exits with code 2 and a clear skip: message; the script will not fetch or configure anything.