[CI] Shard Kernels Attention Test to 7 shards, pin AMD mirror at 2 (<20 min) - #52340
Closed
khluu wants to merge 1 commit into
Closed
[CI] Shard Kernels Attention Test to 7 shards, pin AMD mirror at 2 (<20 min)#52340khluu wants to merge 1 commit into
khluu wants to merge 1 commit into
Conversation
…at 2 kernels-attention-test currently runs 2 NVIDIA shards at ~48-49m each (~96m logical, Phase 2 45+ min long pole). Raise NVIDIA parallelism 2 -> 7 (pytest-shard already in the command) to land each shard ~13-14m (<20m target). Pin mirror.amd.parallelism: 2 so the higher NVIDIA count does not multiply the AMD mirror; this relies on the ci-infra amd-mirror-parallelism-override (PR vllm-project#473). Per-shard timeout 65 -> 30.
Member
Author
|
Superseded by consolidated kernels PR #54754 on current main. |
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
What
kernels-attention-testruns 2 NVIDIA shards at ~48–49 min each (~96 min logical) — a Phase-2 long pole. Raise NVIDIAparallelism2 → 7 to land each shard ~13–14 min (<20 min target), while keeping the AMD mirror at its current 2 shards.How
parallelism: 2 → 7(thepytest -v -s kernels/attention --shard-id/--num-shardscommand already shards). Per-shardtimeout_in_minutes65 → 30. Coverage/selection unchanged.mirror.amd.parallelism: 2so the higher NVIDIA count does not multiply the AMD mirror. This depends on the ci-infra AMD-mirror parallelism override — Honor mirror.amd.parallelism so NVIDIA sharding cannot multiply AMD mirrors ci-infra#473 (amd-mirror-parallelism-override). Without indexSelectLargeIndex: block: [308,0,0], thread: [95,0,0] AssertionsrcIndex < srcSelectDimSizefailed. #473 the field is ignored (AMD would inherit 7); with indexSelectLargeIndex: block: [308,0,0], thread: [95,0,0] AssertionsrcIndex < srcSelectDimSizefailed. #473 it stays at 2. AMD command/coverage unchanged.Sizing
~96 min logical ÷ 7 ≈ 13.7 min/shard; leaves headroom under 20 min for pytest-shard's nodeid-hash imbalance. Provisional — will confirm from the targeted run and trim if there's excess headroom.
Validation — build #83915 (targeted
kernels-attention-teston #473, 7/7 green)srcIndex < srcSelectDimSizefailed. #473 generator, heade006ba64): NVIDIA renders 7, AMD mirror pinned to exactly 2 (un-pinned it would inherit 7); AMD command byte-identical.Part of the Phase-2 CI-overhaul sharding pass (kernels/basic-models workstream), one PR per key.