Motivation
We want reliable, comparable performance numbers for the kernels under hf.co/kernels-community. Rather than building a full leaderboard, keep it deliberately simple: pick a handful of important kernels, always report their benchmark results in the system card, and track whether a new version caused a speed deviation.
Prior art for the "track deviations across versions" part: https://huggingface.co/datasets/diffusers/benchmarks
Proposal
- Select a small set of important kernels under
kernels-community to start with.
- Run a fixed benchmark per kernel and publish the results into the kernel's system card.
- Persist historical results (e.g. a dataset on the Hub) so version-over-version speed deviations are visible.
- Trigger: not nightly. Run after a PR is opened against an important kernel and
build-upload has completed.
- Expose it as
/kernel-bot benchmark, implemented as a GitHub Actions workflow with workflow_dispatch.
Notes
Discussed in Slack: @sayakpaul has a concrete idea for how to set this up; @danieldk agreed on the /kernel-bot benchmark entry point. @drbh previously worked on benchmarks (see also huggingface/kernels-benchmarks) and may have reusable pieces.
Requested by Sayak Paul - Slack thread - Agent trace
Motivation
We want reliable, comparable performance numbers for the kernels under hf.co/kernels-community. Rather than building a full leaderboard, keep it deliberately simple: pick a handful of important kernels, always report their benchmark results in the system card, and track whether a new version caused a speed deviation.
Prior art for the "track deviations across versions" part: https://huggingface.co/datasets/diffusers/benchmarks
Proposal
kernels-communityto start with.build-uploadhas completed./kernel-bot benchmark, implemented as a GitHub Actions workflow withworkflow_dispatch.Notes
Discussed in Slack: @sayakpaul has a concrete idea for how to set this up; @danieldk agreed on the
/kernel-bot benchmarkentry point. @drbh previously worked on benchmarks (see alsohuggingface/kernels-benchmarks) and may have reusable pieces.Requested by Sayak Paul - Slack thread - Agent trace