chore: nightly sync main into dev (12_05_2026)

[svcnvidia-nemo-ci]

Summary

Nightly main → dev sync bringing 17 commits from main into dev.

Python-only line-change stats: Python lines: +3923 / -496 across 58 files

Files Taken Wholesale

These files were kept from one side because no surgical reconciliation was needed.

• From dev: pyproject.toml, uv.lock, docker/Dockerfile.ci.dev — preserved per
  the merge policy (dev has fast-hadamard-transform, dev-pinned TE rev, omegaconf
  dependencies). No new git sources in main vs. dev's [tool.uv.sources] needed
  porting.
• From dev: .github/CODEOWNERS — intentionally divergent from main; never
  modified.
• From dev: megatron/training/training.py, initialize.py, utils.py,
  datasets/data_samplers.py, core/optimizer/layer_wise_optimizer.py — no recent
  main commits touched these files. Dev's evolution (e.g. args.hybrid_context_parallel
  → args.dynamic_context_parallel rename in c3dbea746) is preserved.

Surgical Conflict Resolutions

git merge origin/main -X theirs --no-edit left these "Frankenstein" hybrids that
required manual reconciliation:

1. megatron/core/extensions/transformer_engine.py — restored dev's
   TEFusedDenseMLP class (Dense+Grouped GEMM fusion on SM100+, ~184 lines)
   that -X theirs dropped when accepting main's TEFusedMLP.as_mlp_submodule
   addition. Both additions now coexist.

2. megatron/core/models/gpt/gpt_layer_specs.py — three repairs:

   • Restored HyperConnectionTransformerLayer import (used at L362, L489, L791).
   • Restored dense_grouped_gemm: bool = False parameter on
     get_mlp_module_spec_for_backend (callers in get_gpt_layer_with_transformer_engine_spec
     pass it on L243).
   • Restored the dense_grouped_gemm branch using main's new .as_mlp_submodule
     classmethod pattern (not_none(TEFusedDenseMLP).as_mlp_submodule).

3. megatron/core/models/gpt/experimental_attention_variant_module_specs.py —
   restored HyperConnectionTransformerLayer import (used at L304).

4. megatron/core/transformer/moe/moe_layer.py — restored
   input_ids: Optional[torch.Tensor] = None parameter on MoELayer.forward;
   the body references input_ids at L702 (self.route(... input_ids)).

5. megatron/core/transformer/transformer_layer.py — four repairs:

   • Restored mlp_hyper_connection field on TransformerLayerSubmodules
     (used by HyperConnectionTransformerLayer at L1531-L1604, L1749, L1872).
   • Re-added **moe_kwargs to tensor_parallel.checkpoint(functools.partial(apply_module(self.mlp)...)).
   • Re-added **moe_kwargs to direct apply_module(self.mlp)(...) call.
   • Restored input_ids=input_ids on MoETransformerLayer's router_outputs call.

6. tests/unit_tests/transformer/moe/test_routers.py — restored imports of
   TopKRouter and get_default_pg_collection (used at L621-L681, L620+L657
   respectively). Without these the suite fails with
   NameError: name 'TopKRouter' is not defined (the failure that closed the
   prior bot run).

7. tests/unit_tests/transformer/moe/test_token_dispatcher.py — restored
   MoETokenDispatcher import (used at L37 for _NestedAttrTestDispatcher).

Files Kept from Main (Conflict Resolution Verified)

• megatron/training/config/container.py — main's lazy-import of omegaconf
  (PR Guard omegaconf imports #4685 "Guard omegaconf imports") was the correct resolution; dev's
  unconditional from omegaconf import OmegaConf was the regression main fixed.
• pretrain_vlm.py — kept main's import set (mpu not used in file;
  get_submodules used at L186).

Files Deleted in Dev (Not Restored)

• tests/test_utils/recipes/h100/bert.yaml, tests/test_utils/recipes/h100/t5.yaml
  — both intentionally deleted on dev (commit 40d30942d: "fix: remove files
  that dev intentionally deleted"). Nothing in the merged tree references them.
• .github/workflows/multi-approval-bot.yml — intentionally deleted on dev
  (PR ci: Remove multi-approval action from dev branch #3576). Dev's version respected.

Formatting

• black --config pyproject.toml clean on all 58 changed Python files
• isort clean (applied to examples/multimodal/radio/radio_g.py)

Test plan

• Unit tests pass
• Integration tests pass (or are skipped per maintainer policy)
• Internal GitLab functional tests pass

🤖 Generated with Claude Code

[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 4e241c8

[svcnvidia-nemo-ci]

/ok to test b408019

[svcnvidia-nemo-ci]

/ok to test fb8317b

[svcnvidia-nemo-ci]

Superseded by today's nightly sync.