chore: nightly sync main into dev (28_04_2026)

[svcnvidia-nemo-ci]

Summary

Nightly sync of main into dev for 2026-04-28. Squashes 42 commits from main into the dev branch.

Python lines: +11647 / -2593 across 129 files

Files where main's version was taken

Per the nightly-sync skill, the following files were overridden with main's version because dev's version referenced args / APIs that main has removed or renamed:

• megatron/training/training.py — references dev-only args; main has the more recent control-flow
• megatron/training/utils.py — references dev-only args
• megatron/training/datasets/data_samplers.py — main's MegatronFullValidationSampler plus
  HybridCPMegatronPretrainingSampler are the canonical implementations

Beyond the explicit override list, megatron/training/initialize.py was kept at dev's version
because main's call passes hybrid_context_parallel= to initialize_model_parallel, which dev's
function signature no longer accepts (renamed/refactored to dynamic_context_parallel +
min_dynamic_context_parallel_size).

Files restored from main

• megatron/core/pipeline_parallel/hybrid_cp_schedule.py — dev had renamed/removed it during the
  Dynamic CP refactor; main's data_schedule.py::HybridCPDataLoaderWrapper (used by
  training.py for the deprecated hybrid_context_parallel feature) still imports
  BalancedCPScheduler from it.

Special-case merges

• megatron/core/datasets/data_schedule.py — kept dev's file (Dynamic-CP schedulers) and
  appended main's HybridCPDataLoaderWrapper class verbatim. Imports updated to add
  BalancedCPScheduler and the typing helpers (Any, List).

Files kept at dev's version (dependency-management triple + CODEOWNERS)

• pyproject.toml (dev-only nvtx~=0.2, fast-hadamard-transform, dev-pinned
  nvidia-resiliency-ext revision)
• uv.lock (machine-generated; matches pyproject.toml)
• docker/Dockerfile.ci.dev
• .github/CODEOWNERS (intentionally divergent from main; sync bot must never touch this)

No new git sources were present in main but missing from dev, so no [tool.uv.sources]
reconciliation was needed.

Formatting

All changed Python files were formatted with black --skip-magic-trailing-comma --skip-string-normalization (24.10) followed by isort (5.13.2).
pylint reports 10.00/10 on the changed megatron/core/ files.

Remerge-diff stat (only shows where conflict resolution differed from a clean three-way merge)

 examples/mimo/train.py                            |   3 -
 megatron/core/transformer/moe/fused_a2a.py        |  13 -
 megatron/core/transformer/moe/token_dispatcher.py |   4 -
 megatron/core/transformer/transformer_config.py   |  27 -
 megatron/training/activation_logging.py           |   7 -
 megatron/training/arguments.py                    | 280 ----------
 megatron/training/checkpointing.py                |  20 -
 megatron/training/training.py                     |  72 ---
 pretrain_bert.py                                  |  13 -
 pretrain_mamba.py                                 | 363 -------------
 pretrain_vlm.py                                   |   5 -
 tests/unit_tests/models/test_hybrid_moe_model.py  |   4 -
 tools/checkpoint/checkpoint_inspector.py          |   5 -
 tools/checkpoint/remap_gpt_dsa_to_mamba.py        |   7 -
 train_rl.py                                       |   7 -
 uv.lock                                           | 635 ----------------------

Test plan

• linting
• cicd-container-build (dev + lts)
• Coverage (unit-test) and unit-test workers
• Internal GitLab functional tests (H100 + GB200)
• Nemo_CICD_Test

🤖 Generated with Claude Code

[svcnvidia-nemo-ci]

/ok to test cd2001d

[svcnvidia-nemo-ci]

/ok to test 54cbb38

[svcnvidia-nemo-ci]

/ok to test 0e9f938

[svcnvidia-nemo-ci]

/ok to test 3d25b74

[Phlip79]

/ok to test 42f1108

[svcnvidia-nemo-ci]

Superseded by today's nightly sync.

[FDecaYed]

@xiaoyao0115 to fix the dynamic context parallel related code. I think claude is taking main over dev but in this case main is lagging behind since the mirror PR has not been merged.

[svcnvidia-nemo-ci]

Superseded by today's nightly sync.