Repository navigation
Conversation
sunxxuns
requested review from
Fridge003,
HaiShaw,
Kangyan-Zhou,
bingxche,
ispobock and
merrymercy
as code owners
June 29, 2026 02:02
Contributor
|
Warning You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again! |
sunxxuns
added
run-ci
This was referenced Jun 30, 2026
sunxxuns
force-pushed
the
ci/amd-stage-b-shard-14to8
branch
2 times, most recently
from
July 7, 2026 23:35
89172b0 to
1a9185b
Compare
sunxxuns
changed the title
sunxxuns
force-pushed
the
ci/amd-stage-b-shard-14to8
branch
from
July 8, 2026 01:17
1a9185b to
c92fcd4
Compare
sunxxuns
force-pushed
the
ci/amd-stage-b-shard-14to8
branch
8 times, most recently
from
July 8, 2026 10:20
dc37c1e to
ff872bd
Compare
sunxxuns
force-pushed
the
ci/amd-stage-b-shard-14to8
branch
21 times, most recently
from
July 11, 2026 04:19
36cb3d3 to
cb8e306
Compare
…PU-min, -33% wall-clock) 6 shards = 1 wave at 6 concurrent MI325 runners (zero idle), vs 14 shards = 3 waves with 4 idle runners in the last wave. Saves 48 GPU-min/run of bring-up overhead (6min × 8 fewer shards) and 24min wall-clock. 8 shards (prior revision) was suboptimal: 2 waves (6+2) with 4 idle in wave 2, giving 76min wall-clock — slower than the 14-shard baseline (73min). 6 is optimal on both GPU-min (292 vs 340) and wall-clock (49min vs 73min). Longest shard at 6 = ~53min (LPT-balanced), under the 60min job timeout and far under the 480min wait-for-stage-b gate. Per-file 2400s (40min) timeout is generous (max single-test est = 4.35min). Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Collaborator
Author
|
Friendly nudge here — this one has been waiting for review for about three weeks. It's a CI-only change (consolidating stage-b shards 14→6) and all checks are green on the latest head. Is there anything else needed to move it forward? Happy to adjust if anything's unclear. — sunxxuns |
This was referenced Sep 1, 2026
This branch has not been deployed
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
3 participants
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.
Summary
Consolidates
stage-b-test-1-gpu-small-amdfrom 14 matrix shards to 6, saving ~14% GPU-minutes per PR run and ~33% wall-clock, with no test coverage change.Why 6 (not 8 or 14)
The suite has 108 tests, ~256min total
est_time. With ~6 concurrent MI325 1-GPU runners observed in practice:6 shards is optimal on both axes: 1 wave at 6 concurrent runners (zero idle), minimal bring-up overhead (6min × 6 = 36min vs 6min × 14 = 84min).
8 shards (a prior revision of this PR) was suboptimal: 2 waves (6+2) with 4 runners idle in wave 2, giving 76min wall-clock — slower than the 14-shard baseline (73min).
Timeout safety
Longest shard at 6 = ~53min (LPT-balanced), under the 60min job timeout and far under the 480min
wait-for-stage-bgate. Per-file 2400s (40min) timeout is generous (max single-test est = 4.35min).Changes
.github/workflows/pr-test-amd.yml: matrixpart14→6,--auto-partition-size14→6,expected_count14→6🤖 Generated with Claude Code
CI States
Latest PR Test (Base): ✅ Run #29337465584
Latest PR Test (Extra): ✅ Run #29337465077
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.