Integrate sequence-packing scheduler into training loop and GPT pretraining - #6628
Closed
ilml wants to merge 2 commits into
Closed
Integrate sequence-packing scheduler into training loop and GPT pretraining#6628ilml wants to merge 2 commits into
ilml wants to merge 2 commits into
Conversation
…aining Split 7/10 from NVIDIA#3386 (sequence packing / THD E2E support). Wires wrap_data_iterator into train_step/evaluate with dynamic num_microbatches, adds global seqlen-stats plumbing for FLOPs accounting, gates the HybridCP sampler, and restructures pretrain_gpt.forward_step for packed (THD) batches with padding mask. Original changes by @xiaoyao0115 in NVIDIA#3386. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Signed-off-by: ilml <tolong@nvidia.com>
Follow-up to dropping the unused mirror field in NVIDIA#5903: the dataset config no longer carries this knob, so core_gpt_dataset_config_from_args must not set it. The scheduler is still selected from args.sequence_packing_scheduler via ModelParallelConfig. Original changes by @xiaoyao0115 in NVIDIA#3386. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: ilml <tolong@nvidia.com>
This was referenced Aug 18, 2026
Contributor
Author
|
/ok to test 6dc493f |
2 tasks
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Note
Replaces #5907, which was auto-closed when its review-time base (a copy-pr-bot mirror ref) was deleted after #5903 merged. Same branch, rebased onto main past #5903; prior review history lives in #5907.
What does this PR do?
Wires
wrap_data_iteratorintotrain_step/evaluatewith dynamicnum_microbatches, adds global seqlen-stats plumbing for variable-length FLOPs accounting, gates the HybridCP sampler on the new scheduler flag, and restructurespretrain_gpt.forward_stepfor packed (THD) batches with padding mask.Part 07/10 of splitting #3386 (Add E2E support for THD format; dev-branch PR #2924). Original changes by @xiaoyao0115 in #3386 — split into functionally self-contained PRs to ease review. Hard dependencies (must merge first): #5902, #6625, #6626.
Split series (#3386)
Branches are stacked linearly (each on the previous) so every PR shows a clean own-diff once its base is retargeted to the copy-pr-bot
pull-request/<parent>ref; until then the Files-changed view of a stacked PR includes its ancestors — its own change is the last commit.Stacked on #6627.
Review notes
data_samplers.pyidentity-collate or-chain includesuse_vanilla_collate_fn, which is defined nowhere in the series — candidate for deletion in review. The varlen flags in the same chain are getattr-guarded and land later in the series.Issue tracking
Linked issue: Related to #3386
Contribution process
Pre-checks
🤖 Generated with Claude Code