Module bench 

What a model does on this machine: one run’s measurement, the summary over several, and the order and bar widths a table of them draws with.

Nothing here runs a model or reads a clock. The driver takes the marks and hands them over; this decides what the figures mean, which of them the backend actually reported, and how they rank.

Structs

Figures

What a finished model’s runs came to.

Measure

A figure over several runs: the middle one, and the two ends it varied between.

Row

One row of the bench: the model, what runs it, and how it is doing.

Sample

One run of one model.

WallClock

The marks the driver takes around one run, in milliseconds: when the first token arrived and when the stream ended, both measured from the request.

Enums

ColdStart

What the cold run measured, or why it did not.

Phase

Which phase of a model’s turn is under way.

Status

Where a model is in the bench.

TimingSource

Where a run’s timing came from.

Functions

estimate_tokens

Tokens a reply of text is worth, when nothing counted them: its characters over [CHARS_PER_TOKEN], and never zero for text that exists.

fastest

The fastest rate any row measured, which the bars are drawn against.

filled_cells

How many of a bar’s width cells a rate of rate lights, against the fastest row. A rate that measured anything at all keeps one cell, so a slow model reads as slow rather than as nothing measured.

rank

rows in the order the finished table draws them: measured rows fastest first, then everything that produced no figure, in the order it came.