docs(skills): repair sequence-packing guidance and refresh signature by cuichenx · Pull Request #4636 · NVIDIA-NeMo/Megatron-Bridge

cuichenx · 2026-07-02T20:55:07Z

Summary

Repair the sequence-packing skill after recent packing refactors, update it against current main, and refresh its generated NVSkills evaluation, publication card, and OMS signature in the same PR.

The branch includes signed merge commits through 760be8f0, which incorporate main@572a8e9d without rewriting or overriding Yu Yao's changes.

The remediation, content, and main-sync commits after the original PR head were produced by Codex under user direction. Treat them as AI-authored changes requiring normal review; GPG signatures and DCO signoff establish commit identity/compliance, not human authorship.

Agent-authored skill changes

Replace stale embedded source ranges with anchors at the current SFT configuration, config validation, collate-time packing, and GPT THD metadata paths.
Clarify that offline pad_seq_to_mult is configured explicitly, while ConfigContainer computes the CP/SP LCM automatically only for in-batch packing.
Point VLM in-batch packing guidance at the collate-time implementation in data/sequence_batching.py while retaining the vlm_step.py consumer evidence.
Point long-context and model opt-out evidence at the actual recipes/*/h100/ implementation modules introduced by the recipe namespace refactor.
Document Yu Yao's mapped-padding loss-mask guarantee from fix(data): preserve loss masking for mapped padding #4675 and add its SFT and packed-parquet implementation/tests to the skill evidence and verification commands.
Replace brittle historical pass counts with focused verification commands that cover config mutual exclusion, direct collate-time packing, VLM metadata consumption, and mapped padding.
Correct the packed-plus-CP functional-test evidence path and update the validation card date and direct unit-test evidence.

No Megatron Bridge product/runtime source code is modified by this PR's agent-authored content commits; current main is merged verbatim, and the substantive changes are limited to skill instructions and metadata.

Generated NVSkills artifacts

NVSkills CI owns regeneration of:

BENCHMARK.md
skill-card.md
skill.oms.sig

The generated commit must be based on the latest non-NVSkills source/content commit before merge.

Validation

15 passed: tests/unit_tests/training/utils/test_packed_seq_utils.py
17 passed: focused packing/config/CP validation selection in tests/unit_tests/training/test_config.py
9 passed: tests/unit_tests/data/vlm_datasets/test_batching.py
2 passed: focused deferred-packing/packed-metadata selection in tests/unit_tests/training/test_vlm_step.py
2 passed: negative and mapped-negative loss-mask cases in tests/unit_tests/data/datasets/test_packed_parquet.py
2 passed: mapped-padding loss-mask cases for SFT and packed SFT in tests/unit_tests/data/datasets/test_sft.py
Every card.yaml evidence path exists in the merged tree, and all embedded source ranges match current line locations.
uv run --no-project pre-commit run --all-files in the NeMo 26.06 container
Full pre-commit was rerun after the final main@572a8e9d merge.
User-authorized, agent-authored merge/content commits are GPG-signed and include DCO signoff.

Merge gate

Do not merge until NVSkills CI regenerates the evaluation/card/signature artifacts from the latest non-NVSkills source/content commit and the signature verification reports the PR head as current.

copy-pr-bot · 2026-07-02T20:55:11Z

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

cuichenx · 2026-07-02T20:55:35Z

/nvskills-ci

cuichenx · 2026-07-02T21:23:51Z

/nvskills-ci

svcnvidia-nemo-ci · 2026-07-02T21:33:50Z

/nvskills-ci

claude · 2026-07-02T21:37:41Z

LGTM

Docs-only change: adds an HTML comment (NVSkills signature refresh request) after the frontmatter in skills/nemo-mbridge-perf-sequence-packing/SKILL.md, with no instruction changes. The comment sits after the closing frontmatter delimiter, so it does not interfere with YAML parsing. No code, logic, tests, or config are affected.

Suggested test cases:

No perf tests impacted.

svcnvidia-nemo-ci · 2026-07-02T21:45:53Z

/nvskills-ci

cuichenx · 2026-07-02T22:03:08Z

/nvskills-ci

cuichenx · 2026-07-02T22:20:57Z

/nvskills-ci

cuichenx · 2026-07-02T22:35:42Z

/nvskills-ci

cuichenx · 2026-07-02T22:51:07Z

/nvskills-ci

cuichenx · 2026-07-02T23:07:57Z

/nvskills-ci

cuichenx · 2026-07-02T23:22:46Z

/nvskills-ci

cuichenx · 2026-07-02T23:40:05Z

/nvskills-ci

cuichenx · 2026-07-02T23:56:45Z

/nvskills-ci

Signed-off-by: Chen Cui <chcui@nvidia.com>

cuichenx · 2026-07-03T00:11:27Z

/nvskills-ci

Signed-off-by: nvskills-svc-account <svc-nvskills-signing@nvidia.com>

copy-pr-bot · 2026-07-03T00:22:06Z

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

Signed-off-by: Chen Cui <chcui@nvidia.com>

cuichenx · 2026-07-06T18:27:07Z

/nvskills-ci

Signed-off-by: nvskills-svc-account <svc-nvskills-signing@nvidia.com>

cuichenx · 2026-07-06T20:57:50Z

/ok to test 8c260d1

claude · 2026-07-06T20:59:27Z

line1
line2

claude · 2026-07-06T20:59:50Z

LGTM - documentation-only update to the nemo-mbridge-perf-sequence-packing skill (BENCHMARK.md, SKILL.md, card.yaml, skill-card.md, sigstore bundle). I verified every code anchor and test reference in the revised SKILL.md against the current tree and they all resolve correctly:

finetune_utils.py:110-125, config.py:1219-1247 and 1399-1441 - line ranges and code match.
sequence_batching.py - prepare_padded_or_packed_sequence_batch (L397) and pack_right_padded_sequence_batch_to_mcore_thd (L290) exist as cited.
gpt_step.py packed-THD cu_seqlens constraint matches _cu_seqlens_for_cp_partition.
Verification -k filters map to real tests: offline_and_in_batch_packing_are_mutually_exclusive, deferred_in_batch_packing -> test_forward_step_rejects_deferred_in_batch_packing, packed_metadata -> test_get_batch_consumes_collated_packed_metadata.
New/renamed paths (tests/unit_tests/data/vlm_datasets/test_batching.py, tests/functional_tests/test_groups/training/test_seqpacking_cp_example.py) both exist.

No bugs, typos, or stale references found. The clarified CP-block comment (offline packing is not finalized by ConfigContainer; LCM is auto-computed for in-batch packing only) is consistent with config.py:1240-1247.

Suggested test cases

No perf tests impacted - this PR touches only skill markdown/metadata under skills/nemo-mbridge-perf-sequence-packing/ and no scripts/performance/configs/ files. The verification commands documented in the skill are the relevant coverage:

tests/unit_tests/training/utils/test_packed_seq_utils.py
tests/unit_tests/training/test_config.py (packed_sequence / enable_in_batch_packing / offline_and_in_batch_packing_are_mutually_exclusive / context_parallel_seq_length_divisibility / context_parallel_finetuning_validations)
tests/unit_tests/data/vlm_datasets/test_batching.py
tests/unit_tests/training/test_vlm_step.py (deferred_in_batch_packing / packed_metadata)

Signed-off-by: Chen Cui <chcui@nvidia.com>

svcnvidia-nemo-ci · 2026-07-06T21:14:20Z

/nvskills-ci

Signed-off-by: nvskills-svc-account <svc-nvskills-signing@nvidia.com>

Signed-off-by: Chen Cui <chcui@nvidia.com>

svcnvidia-nemo-ci · 2026-07-06T21:32:05Z

/nvskills-ci

Signed-off-by: nvskills-svc-account <svc-nvskills-signing@nvidia.com>

cuichenx · 2026-07-06T21:53:43Z

/ok to test a7d8b57

…VIDIA-NeMo#4636) Signed-off-by: Chen Cui <chcui@nvidia.com> Signed-off-by: nvskills-svc-account <svc-nvskills-signing@nvidia.com> Co-authored-by: nvskills-svc-account <svc-nvskills-signing@nvidia.com>

cuichenx marked this pull request as ready for review July 2, 2026 21:25

copy-pr-bot Bot temporarily deployed to public July 2, 2026 21:36 Inactive

copy-pr-bot Bot had a problem deploying to test July 2, 2026 21:37 Error

copy-pr-bot Bot temporarily deployed to test July 2, 2026 21:46 Inactive

copy-pr-bot Bot temporarily deployed to public July 2, 2026 21:47 Inactive

cuichenx marked this pull request as draft July 2, 2026 21:55

copy-pr-bot Bot temporarily deployed to public July 2, 2026 21:56 Inactive

docs(skills): repair sequence-packing references

d3ea0be

Signed-off-by: Chen Cui <chcui@nvidia.com>

cuichenx force-pushed the automation/nvskills-signature-refresh-20260702 branch from 12967f2 to d3ea0be Compare July 3, 2026 00:11

Attach NVSkills validation signatures

e661e34

Signed-off-by: nvskills-svc-account <svc-nvskills-signing@nvidia.com>

docs(skills): correct sequence-packing guidance

a9f966c

Signed-off-by: Chen Cui <chcui@nvidia.com>

cuichenx changed the title ~~[automation] chore: refresh stale NVSkills signatures (2026-07-02)~~ docs(skills): repair sequence-packing guidance and refresh signature Jul 6, 2026

Attach NVSkills validation signatures

8c260d1

Signed-off-by: nvskills-svc-account <svc-nvskills-signing@nvidia.com>

cuichenx added the docs-only With great power comes great responsibility. label Jul 6, 2026

cuichenx marked this pull request as ready for review July 6, 2026 20:57

cuichenx added 2 commits July 6, 2026 14:09

merge(main): update sequence-packing PR

49aea08

Signed-off-by: Chen Cui <chcui@nvidia.com>

docs(skills): sync sequence-packing guidance with main

7595f59

Signed-off-by: Chen Cui <chcui@nvidia.com>

svc-nvskills-signing and others added 2 commits July 6, 2026 21:26

Attach NVSkills validation signatures

6a4ae79

Signed-off-by: nvskills-svc-account <svc-nvskills-signing@nvidia.com>

merge(main): refresh sequence-packing PR

760be8f

Signed-off-by: Chen Cui <chcui@nvidia.com>

Attach NVSkills validation signatures

a7d8b57

Signed-off-by: nvskills-svc-account <svc-nvskills-signing@nvidia.com>

cuichenx merged commit 9c4b997 into main Jul 6, 2026
40 checks passed

cuichenx deleted the automation/nvskills-signature-refresh-20260702 branch July 6, 2026 22:42

Conversation

cuichenx commented Jul 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Agent-authored skill changes

Generated NVSkills artifacts

Validation

Merge gate

Uh oh!

copy-pr-bot Bot commented Jul 2, 2026

Uh oh!

cuichenx commented Jul 2, 2026

Uh oh!

cuichenx commented Jul 2, 2026

Uh oh!

svcnvidia-nemo-ci commented Jul 2, 2026

Uh oh!

claude Bot commented Jul 2, 2026

Uh oh!

svcnvidia-nemo-ci commented Jul 2, 2026

Uh oh!

cuichenx commented Jul 2, 2026

Uh oh!

cuichenx commented Jul 2, 2026

Uh oh!

cuichenx commented Jul 2, 2026

Uh oh!

cuichenx commented Jul 2, 2026

Uh oh!

cuichenx commented Jul 2, 2026

Uh oh!

cuichenx commented Jul 2, 2026

Uh oh!

cuichenx commented Jul 2, 2026

Uh oh!

cuichenx commented Jul 2, 2026

Uh oh!

cuichenx commented Jul 3, 2026

Uh oh!

copy-pr-bot Bot commented Jul 3, 2026

Uh oh!

cuichenx commented Jul 6, 2026

Uh oh!

cuichenx commented Jul 6, 2026

Uh oh!

claude Bot commented Jul 6, 2026

Uh oh!

claude Bot commented Jul 6, 2026

Uh oh!

svcnvidia-nemo-ci commented Jul 6, 2026

Uh oh!

svcnvidia-nemo-ci commented Jul 6, 2026

Uh oh!

cuichenx commented Jul 6, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

cuichenx commented Jul 2, 2026 •

edited

Loading