TorchLean kernels

One Kernel.lean of exact tensor kernels, proved equal to TorchLean's binary32 specs, then timed on every model-family board (overall and one per model) against that board's reference: its current champion, or the kernel the competition shipped with. An entry at or below the promotion bar, significantly faster (the bootstrap's lower bound on the gain is above zero) and within the per-cell guard takes the board. Each board pays its two newest champions by rank, out of its share of the competition. Scores, admissions and ranks are the coordinator's stored decisions; nothing here is re-scored.

Each dot is one benchmark job on one worker environment, timed from its start to its result as the worker reported them; the line holds the median of that environment's last five jobs until its next one finishes. A failed job has no result time, so it is counted but not timed: it is a cross along the bottom at its start, and the other environment's run of the same evaluation is cancelled (not shown). A board evaluation runs from planning to decision, so it includes the wait for a free worker; an entry's turnaround runs from submission to every active board's decision. Timings come from each entry's report, read again only when its evaluations change, over the newest 100 entries.

Click a board for its challengers and champions. Scores are time ÷ the board's reference, the geometric mean over its cells (model × batch × input kind) and environments; lower is faster. A champion's score is against the reference it took over from. "Pay" is each paid rank's fraction of TorchLean's allocation: rank share × the board's reward share (zero while the board is paused under the treasury policy).

Each cell is the entry's latest decision on that board: its score, time ÷ the reference it was timed against (blue faster, red slower, gray about even; the colour saturates at 1.5× either way). Entries were timed against whichever reference held the board then, so scores in one column compare only within the same reference. Click a cell for the entry's report. The newest 100 entries; rejected ones are in the submission list below.

Challengers are evaluations against this reference under the active benchmark configuration; a superseded evaluation (re-planned) is left out, a pending one is still being timed. Champions are every promotion on the board, newest first; an operator disqualification hands the board back to the newest eligible earlier champion. A former champion's source is published once it holds no board.

Newest first; click an entry for its report. Processing: queued and proving before the proof check, waiting proved but not yet planned on every board, benchmarking while an evaluation is timed, completed once every active board has a decision (which does not mean promoted). Board pills carry the latest evaluation's score, coloured by its admission.