[zupengwang]

What does this PR do?

Adds a sync_dist_op argument to LightningModule.log and LightningModule.log_dict so the reduction across distributed ranks can be configured independently from the reduction across epoch steps.

When sync_dist_op is omitted, the distributed operation still falls back to reduce_fx, preserving existing behavior. The argument is appended to the public signatures to avoid shifting existing positional arguments. The API docstrings and PyTorch changelog are updated.

The regression test covers CPU, single-GPU, two-process CPU DDP, and two-GPU DDP. It logs rank-dependent values with sync_dist_op="sum" and reduce_fx="max", verifying the two reduction stages produce the expected metric.

Validation:

• python -m pytest -q tests/tests_pytorch/trainer/logging_/test_train_loop_logging.py::test_logging_sync_dist_true (4 passed)
• CUDA_VISIBLE_DEVICES= python -m pytest -q tests/tests_pytorch/core/test_results.py tests/tests_pytorch/trainer/logging_/test_train_loop_logging.py (31 passed, 4 skipped)
• PL_RUN_STANDALONE_TESTS=1 python -m pytest -q tests/tests_pytorch/strategies/test_ddp_integration.py::test_ddp_gradients_synced[...] (all four automatic/manual and static/non-static variants passed when run individually)
• ruff check and ruff format --check on the changed Python files
• mypy on the changed source files
• uv build --offline --wheel for the pytorch package

AI assistance disclosure: AI tools assisted with implementation and test preparation. The contributor reviewed the complete diff and validated the change on CPU and GPU.

Fixes #17831

No breaking changes are intended. The default distributed reduction remains unchanged.

Before submitting

• Was this discussed/agreed via a GitHub issue? (not for typos and docs) Yes, in self.log: strategy.reduce vs reduce_fx #17831.
• Did you read the contributor guideline, Pull Request section?
• Did you make sure your PR does only one thing, instead of bundling different changes together?
• Did you make sure to update the documentation with your changes? (if necessary) Yes, the public API docstrings are updated.
• Did you write any new necessary tests? (not for typos and docs) Yes.
• Did you verify new and existing tests pass locally with your changes?
• Did you list all the breaking changes introduced by this pull request? None.
• Did you update the CHANGELOG? (not for typos, docs, test updates, or minor internal changes/refactors) Yes.

PR review

Anyone in the community is welcome to review the PR.
Before you start reviewing, make sure you have read the review guidelines. In short, see the following bullet-list:

Reviewer checklist

• Is this pull request ready for review? (if not, please submit in draft mode)
• Check that all items from Before submitting are resolved
• Make sure the title is self-explanatory and the description concisely explains the PR
• Add labels and milestones (and optionally projects) to the PR so it can be classified

[zupengwang]

Hi @tchaton, I pushed the reviewed follow-up b03287f858655842d7621628d8751df48cf021bf for the sync_dist_op proposal accepted in #17831.

The epoch-mean path now applies an explicitly requested rank reduction to the local batch-weighted means. Previously the operation was applied separately to the weighted sums and sample counts: for local means of 12 and 3.5, explicit sum returned about 6.3333 instead of 15.5, and max returned 6 instead of 12. Omitting sync_dist_op still preserves the existing globally sample-weighted mean. The follow-up adds two-rank regressions, exercises both log and log_dict, and documents the explicit/default distinction without renaming existing parameters.

On a remote Linux host with Python 3.12.13 and torch 2.12.1+cpu using two-rank Gloo, the same added regressions give 9 failures / 2 passes / 2 skips with the previous head's unchanged production code and 11 passes / 2 skips with this commit. The related five-file logging/metric suite gives 74 passes / 8 GPU-dependent skips, and all 11 applicable pre-commit hooks pass. A separate conflict-free integration with master 84df182f50ab gives 78 passes / 8 skips in that suite; that integration was only a compatibility check and was not added to this commit. CUDA/NCCL, TPU, and multi-node behavior have not been validated. The submitting human reviewed the follow-up diff before publication.

Could you review the current commit and this explicit-operation/default-operation distinction? Read the Docs has passed on this head. The new Package, Probot, Code check, Docs builds, Check Schema, and Test PyTorch workflows all show action_required with zero jobs started (for example, https://github.com/Lightning-AI/pytorch-lightning/actions/runs/39784111539). Could a maintainer approve these workflows through the normal process and advise on any additional distributed coverage needed? The expired runs on the old head are historical; the remote CPU/Gloo results above do not replace CI on this commit. Thank you.