chore: nightly sync main into dev (13_05_2026)

[svcnvidia-nemo-ci]

Summary

Nightly sync of main into dev for 13_05_2026.

• 23 commits from main merged into dev
• Python lines: +5956 / -1791 across 81 files

Files taken from main (overrides)

These files were explicitly overridden with main's version because dev's
versions reference args or APIs that main removed/renamed:

• megatron/training/training.py
• megatron/training/initialize.py
• megatron/training/utils.py
• megatron/training/datasets/data_samplers.py
• megatron/core/optimizer/layer_wise_optimizer.py
• tests/test_utils/recipes/h100/bert.yaml (restored from main — missing in dev)
• tests/test_utils/recipes/h100/t5.yaml (restored from main — missing in dev)

Files restored from dev (must not change in sync)

Per skill guidance:

• pyproject.toml (dev keeps fast-hadamard-transform + own transformer-engine / nvidia-resiliency-ext revisions)
• uv.lock
• docker/Dockerfile.ci.dev
• .github/CODEOWNERS

Surgical merges (kept main + dev features)

After -X theirs resolved conflicts in favor of main, the following files
had dev-only features re-injected to fix API breakage:

• megatron/core/extensions/transformer_engine.py — restored dev's
  TEFusedDenseMLP class (referenced by dev test + spec).
• megatron/core/models/gpt/experimental_attention_variant_module_specs.py
  — restored HyperConnectionTransformerLayer import (still referenced).
• megatron/core/models/gpt/gpt_layer_specs.py — restored
  dense_grouped_gemm parameter on get_mlp_module_spec_for_backend
  routing to TEFusedDenseMLP.as_mlp_submodule when set; preserves
  main's MlpBuilder pattern.
• megatron/core/transformer/moe/moe_layer.py — restored input_ids
  parameter on MoELayer.forward (still passed through custom_forward).
• megatron/core/transformer/transformer_layer.py — restored
  mlp_hyper_connection field on TransformerLayerSubmodules, restored
  **moe_kwargs and input_ids=input_ids plumbing on the MLP/router
  forward paths.
• tests/unit_tests/transformer/moe/test_routers.py — restored
  get_default_pg_collection and TopKRouter imports.
• tests/unit_tests/transformer/moe/test_token_dispatcher.py —
  restored MoETokenDispatcher import.

API mismatch audit

The merge took main's version in several "interface" files. After
re-injecting the dev-only features above:

• TEFusedDenseMLP.as_mlp_submodule is inherited from TEFusedMLP
  (since it's a subclass) and is invokable through main's MlpBuilder
  pattern.
• _forward_mlp_router now threads input_ids to self.mlp(...)
  again so MTP hash routing remains functional.

No deleted-from-dev files needed to be restored; the only file
genuinely deleted in this window (.github/workflows/multi-approval-bot.yml)
was intentionally removed on dev per PR #3576.

Conflict resolution overview

-X theirs was used. Files that had remerge conflicts:

megatron/core/extensions/transformer_engine.py
megatron/core/models/gpt/experimental_attention_variant_module_specs.py
megatron/core/models/gpt/gpt_layer_specs.py
megatron/core/transformer/moe/moe_layer.py
megatron/core/transformer/transformer_layer.py
megatron/training/arguments.py
megatron/training/checkpointing.py
megatron/training/config/container.py
megatron/training/training.py
pretrain_vlm.py
pyproject.toml
tests/unit_tests/transformer/moe/test_routers.py
tests/unit_tests/transformer/moe/test_token_dispatcher.py
uv.lock

See commit fd56123ce for the full merge content. The merge commit
preserves both parents (origin/dev and origin/main).

Generated by

Automated sync bot. CI fixes (if any) will land as a follow-up
commit on this branch.

[copy-pr-bot]

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.

[svcnvidia-nemo-ci]

/ok to test fd56123

[svcnvidia-nemo-ci]

/ok to test 187ce73

[svcnvidia-nemo-ci]

/ok to test d0efe4f

[svcnvidia-nemo-ci]

Superseded by today's nightly sync.