How to measure tok/s in a way someone else can use

A tok/s figure on its own is not evidence. To be usable, a measurement needs the conditions that produced it.

Record all of these:

• Model and quantisation, exactly as the file is named.
• Backend and version. A llama.cpp build from six months ago is a different measurement than today’s.
• Context length used, and whether the number is prompt processing (prefill) or generation (decode). Mixing these is the single most common way published figures become useless.
• Batch size. Batch 1 latency and batched throughput are different numbers answering different questions.
• Offload. Whether the whole model sat in fast memory, or spilled to system RAM, changes the result by an order of magnitude.
• The date.

Run the measurement more than once and report a steady-state figure, not the first run, which includes model load and cache warm-up.

State what you did not control. “Laptop on battery, thermally throttled after two minutes” is more useful than a clean-looking number that nobody can reproduce.

Then send it

A measurement recorded this way is one this site can publish. Post it somewhere public — a forum thread, a gist, an issue on the backend you used — and submit the link. It has to exist somewhere public first: every figure here cites a source, and a number that exists only in a submission form cannot be cited.

You keep the credit. Measurements are attributed to whoever made them, by name and date, and two independent people reporting the same run within 25% is what moves a record from single-source to verified.