[benchmark] Shard the linear-attention benchmark across parallel jobs#3107
Open
tarinduj wants to merge 1 commit into
Open
[benchmark] Shard the linear-attention benchmark across parallel jobs#3107tarinduj wants to merge 1 commit into
tarinduj wants to merge 1 commit into
Conversation
meta-cla
Bot
added
the
CLA Signed
tarinduj
marked this pull request as ready for review
July 20, 2026 23:06
There was a problem hiding this comment.
Pull request overview
This PR fixes linear-attention (linattn) benchmark uploads timing out (hitting GitHub Actions’ 6-hour job limit) by sharding the linattn benchmark variants across a parallel matrix, aligning linattn with how other benchmark suites are dispatched and ensuring artifacts consistently upload for the dashboard.
Changes:
- Route
run-h100-linattnandrun-b200-linattnthroughcompute-benchmark-matrixand run each shard as its own job. - Update the
kernels_linattnworkflow input default to list only the 7 variants (not separate-bwdrows) to avoid splitting forward/backward expectations across shards. - Update dashboard expectation generation to expand linattn variants into both
<variant>and<variant>-bwdexpected rows.
Reviewed changes
Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.
| File | Description |
|---|---|
| .github/workflows/benchmark_dispatch.yml | Adds matrix-generation jobs and shards linattn runs; updates linattn kernel defaults to variant-only list. |
| .github/dashboard/build_dashboard_data.py | Expands kernels_linattn defaults into both forward and -bwd expected dashboard rows. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
The run-h100-linattn / run-b200-linattn jobs ran all variants in one job and hit GitHub's 6h job limit, so they were cancelled before uploading and never appeared on the dashboard. Route kernels_linattn through compute-benchmark-matrix like the other GPU suites so each variant runs in its own job. Feeding the fourteen forward/-bwd rows to the sharder would split a variant's forward and backward into different shards, running and autotuning the whole variant twice. kernels_linattn now lists the seven variant names so each runs once; run_linattn still emits both rows, and get_expected_kernels_per_platform expands the -bwd names so the dashboard still expects all fourteen.
tarinduj
force-pushed
the
tarindu/linear-bench-fix
branch
from
July 20, 2026 23:23
6212205 to
29eacde
Compare
AmesingFlank
approved these changes
Jul 21, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The linear-attention kernels did not appear on the dashboard. The
run-h100-linattn/run-b200-linattnnightly jobs ran all variants in one job and hit GitHub's 6h job limit, so they were cancelled before uploading. Every other benchmark suite is sharded into parallel jobs; this brings linattn in line by routingkernels_linattnthroughcompute-benchmark-matrixso each variant runs in its own job.kernels_linattnnow lists the seven variants instead of the fourteen dashboard rows (each variant has a forward and a forward+backward-bwdrow): sharding the fourteen would split a variant's two rows into different shards and run it twice.run_linattnstill emits both rows per variant, andget_expected_kernels_per_platformexpands the-bwdnames so the dashboard still expects every row.