[Connor-XY]

Depends on #8768, which contains #8767, #8766, #8765, #8747, #8761, #8728, #8746, #8037, #8049 and #6293. Please review the diff from the merge of #8768: scoped diff.

• Pass one process-group collection through the pretrain loop #8037 adds _build_pg_collection_from_parallel_state(), and Read the bootstrap process groups from the registry #8768 gives it the owner= keyword and adds the _register_data_parallel_groups test helper.
• The new checks extend tests/unit_tests/inference/test_inference_entry_point_groups.py from Deprecate use_mpu_process_groups outside core and close its silent paths #8767. The embedder tests extend tests/unit_tests/tools/test_tool_process_groups.py from Pass process groups explicitly in the standalone examples and tools #8746.

This PR does not depend on #8772 or #8773 and can land any time after #8768.

Why.

• finetune() in tasks/finetune_utils.py and the BertEmbedder of tools/bert_embedding/embed.py called setup_model_and_optimizer() without pg_collection. Every run took its deprecated fallback: a ProcessGroupFallbackWarning, and an error under --strict-process-group-fallbacks.
• get_model() calls the model provider with config= and pg_collection=. On main, the providers of the detoxification sampler, the multimodal run_text_generation.py and vision_model_tester.py take neither, so get_model() would fail on them with a TypeError. Deprecate use_mpu_process_groups outside core and close its silent paths #8767 and Use the model's process groups in ModelOpt, multimodal and MIMO examples #8761 fix them, but no test can run them: the first two import the removed megatron.inference.text_generation, and the third needs the GPU container.
• Nothing checked that an in-tree caller of setup_model_and_optimizer() passes its groups.

What.

• finetune() and BertEmbedder build the collection that pretrain() builds for a caller that passes none. _build_pg_collection_from_parallel_state(args, owner="setup_model_and_optimizer") reads the groups that initialize_megatron() registered and, with --create-all-gather-group, adds the all-gather groups. Those are the groups the fallback built, so the model and optimizer run on the same groups as before.
• On a grid set up without registering a collection, both still fall back. The warning is the same as on main and names setup_model_and_optimizer(), whose fallback it is.
• New static check: for every call of megatron.training's get_model() outside the tests, the guard finds the provider when the calling script defines it (a function, a nested function, a lambda or a functools.partial() of one) or imports it from a module next to the script or at the repository root. It resolves names in the scope of the call and follows assignments. The provider must accept config= and pg_collection=, by name or through **kwargs, and read the ones it names. The three scripts above must be among those checked. On main's versions, the check reports both keywords for each.
• New static check: every call of setup_model_and_optimizer() outside the tests passes pg_collection. Calls are matched under the names a script imports the function or megatron.training.training as, by the same matcher as get_model()'s calls, so an aliased call is checked and a method of the same name on another class is not. On the tree it finds three calls: pretrain(), the finetuning driver and the BERT embedding tool.

Tests.

• test_parallel_state_deprecation.py: finetune() builds on the registered collection, and on an unregistered grid it warns once with setup_model_and_optimizer as the owner.
• test_tool_process_groups.py: the same two cases for BertEmbedder. They need h5py, transformers and Transformer Engine, and skip otherwise.
• test_logger_ownership.py is adapted to the new call.

Before the 2026-11-01 sync, the branch ran these files and test_inference_entry_point_groups.py on 4 CPU ranks with Gloo: 155 passed, 6 skipped and 2 failed, the same per test as on the stacked version. There, 6 more tests passed than on its base. The 2 embedder tests skip. The 2 failures, test_core_metric_flags_keep_cli_defaults_and_owners[False|True], also failed on the stacked base without this PR.

On 4 GPUs against its base: test_inference_entry_point_groups.py: 18 passed here vs 14 on the base, 4 new tests pass. test_parallel_state_deprecation.py: 112 passed vs 110, 2 new tests pass. test_tool_process_groups.py: 2 passed and 3 skipped vs 2 passed and 1 skipped; the 2 extra skips are the new embedder tests. test_logger_ownership.py: 29 passed on both sides. inference/high_level_api/test_apis.py (29 passed) and test_example_process_groups.py (11 passed) give the same results on both sides. No test regresses against the base. In the integration run, this PR's test files show no regressions against main. On the earlier stacked version, test_inference_entry_point_groups.py also passed under -W error::...ProcessGroupFallbackWarning, and so did the embedder test on the registered grid. Nothing runs finetune() or BertEmbedder end to end, because no in-tree script calls either.

Compatibility.

• The groups are the ones the fallback built, including the all-gather groups. No public API changes.
• tasks/finetune_utils.py and tools/bert_embedding/embed.py now import the private training helper _build_pg_collection_from_parallel_state from megatron.training.training.
• On an unregistered grid the warning still says that setup_model_and_optimizer() "was called without pg_collection", as on main. finetune() and BertEmbedder take that fallback on its behalf and have no pg_collection argument of their own.
• The provider check reads syntax. Providers it cannot resolve are skipped, not failed: the two branch-local multimodal providers of vlm_dynamic_inference.py and simple_t5. OWN_PROVIDER_SCRIPTS is a hard-coded list. A provider that takes **kwargs is only checked to accept the keywords.
• The Testmon mandatory mapping does not list test_inference_entry_point_groups.py, so a selective PR run skips these static checks. That is left for a follow-up.
• Left as on main: the imports of removed modules in generate_samples_gpt.py, examples/multimodal/run_text_generation.py, tools/run_vlm_text_generation.py and tasks/eval_utils.py; finetune()'s training loop, which calls train_step() and training_log() with outdated arguments; and embed.py's top-level pretrain_bert import.
• determinism-exempt is needed only while the PRs below this one are unlanded. CI diffs against main, and the base contains registered kernel files. This PR's own diff passes the coverage check.

The branch was synced with main 77391e7 on 2026-11-01 by a merge commit that merges the updated #8768 without conflicts, so existing commits and review links stay valid.

Part of #6307.

🤖 Generated with Claude Code

[Connor-XY]

Superseded by #8907, which carries this change as commit bb2d6a4 under the same title, with "Originally #8775." in its message. Closing so review happens in one place. The branch stays on the fork.