chore(deps): track DLIO main @ PR #25 merge (fixes #455)#470
Merged
Conversation
Bumps the DLIO_local_changes pin from 1d11f982 to cbe20010, picking up PR #25 (sampler floor division — equal per-rank shards). Fixes the inter-epoch deadlock reported in #455 when num_samples is not an even multiple of comm_size. Also bumps mlpstorage version 3.0.12 -> 3.0.13 so users can tell from `mlpstorage --version` whether the fix is present. Regenerated uv.lock.
|
MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅ |
Contributor
Author
|
@idevasena Just Finishing off the DLIO work you helped fix up! |
idevasena
approved these changes
Jun 18, 2026
idevasena
left a comment
idevasena
left a comment
Contributor
There was a problem hiding this comment.
Looks good! Thank you @FileSystemGuy
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 subscribe to this conversation on GitHub.
Already have an account?
Sign in.
2 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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
Bumps the
dlio-benchmarkgit pin to pick up DLIO PR #25 (cbe20010), which fixes the inter-epoch sampler deadlock reported in #455.1d11f982→cbe200103.0.12→3.0.13uv.lockregenerated.Why
dlio_samplerwas usingmath.ceil(num_samples / comm_size)with a clamp on the last rank. Whennum_samples % comm_size != 0the last rank produced fewer batches per epoch and the per-step / end-of-epochMPI_Barriers inmain._train()matched across iterations — a silent CPU-spinning deadlock at the next epoch boundary with no diagnostic.PR #25 replaces
ceil + clampwith floor division at three call sites (torch_data_loader.dlio_sampler,config.build_sample_map_iter,config.get_global_map_index), emits a rank-0 warning when the floor drops samples, and fixes a pre-existingSampler.__len__contract violation (caught by @idevasena during review) solen(sampler) == len(list(iter(sampler))). Includes three regression tests.Closes #455.
Test plan
mlpstorage --versionreports3.0.13.pip install -e .resolves DLIO atcbe20010.num_samples % comm_size != 0(e.g.unet3dconfigured sonum_files_trainis not a multiple ofnum-accelerators) completes past epoch 1 without hanging.