Add terminate_iteration to skip batch handlers

[nightcityblade]

Fixes #996

Description:

Add Engine.terminate_iteration() so a process function can ignore the current batch without firing ITERATION_COMPLETED handlers that consume state.output. Apply the signal consistently to both the generator and legacy execution paths, and cover both modes with a regression test.

Tests:

• python -m pytest tests/ignite/engine/test_engine.py -q (185 passed, 10 skipped)
• ruff check ignite/engine/engine.py tests/ignite/engine/test_engine.py

Check list:

• New tests are added (if a new feature is added)
• New doc strings: description and/or example code are in RST format
• Documentation is updated (if required)

[Copilot]

Pull request overview

Adds a new Engine control signal to allow a process function to skip ITERATION_COMPLETED handlers for the current batch, preventing downstream handlers from consuming state.output when a batch should be ignored (addressing #996).

Changes:

• Add Engine.terminate_iteration() and a new should_terminate_single_iteration flag.
• Apply the skip logic to both generator (_run_once_on_dataset_as_gen) and legacy (_run_once_on_dataset_legacy) execution paths.
• Add a regression test that runs in both modes via the existing interrupt_resume_enabled parameterization.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
ignite/engine/engine.py	Introduces terminate_iteration() and conditionally skips Events.ITERATION_COMPLETED in both run paths.
tests/ignite/engine/test_engine.py	Adds a regression test verifying skipped batches do not trigger ITERATION_COMPLETED handlers in both execution modes.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[nightcityblade]

Follow-up validation for 122cee3:

• python -m pytest tests/ignite/engine/test_engine.py -q — 185 passed, 10 skipped
• ruff check ignite/engine/engine.py tests/ignite/engine/test_engine.py — passed
• git diff --check — passed

[aaishwarymishra]

Also terminate_epoch has extra parameter I notiched skip_epoch_completed, I think terminate_iteration should be consistent with that

[nightcityblade]

Addressed in 54f5a5e. terminate_iteration now accepts skip_iteration_completed, defaulting to False consistently with terminate_epoch, and both execution paths honor it. The focused parametrized tests pass in generator and legacy modes (4 passed), and Ruff check/format pass.

[aaishwarymishra]

@vfdev-5 LGTM

[nightcityblade]

CI note: the only failing check is linkcheck, caused by the pre-existing Polyaxon URL redirect in generated/ignite.handlers.polyaxon_logger.rst (https://polyaxon.com/docs/experimentation/tracking/client/ redirects to /docs/references/python-library/tracking/client/). The PR does not touch that documentation; all code-style, CPU, doctest, MPS, TPU, Horovod, and pyrefly jobs pass.

[vfdev-5]

@nightcityblade thanks for your effort working on the issue. I remember it was not a simple issue as we can not simply skip iterations especially with metrics attached to engines.
Before merging this PR we should check if this work with 1) a trainer engine and a running metric attached and ensure that computed running metric is correct 2) an evaluator engine with metrics we should ensure that computed metrics are correct

[nightcityblade]

Addressed in 2466d55. The terminate_iteration regression in tests/ignite/engine/test_engine.py now verifies both requested cases: skipped batches are excluded from a trainer's RunningAverage and from an evaluator's epoch Average, with the expected values asserted in both engine execution modes. Validation: 185 passed, 10 skipped for the full engine test file; Ruff check and format check passed on the changed test.

[aaishwarymishra]

sorry, I missed the updates :)

[nightcityblade]

Updated the test in bba1810 to use the default RunningAverage decay while retaining the required identity output transform. The expected value is now 1.04 (only completed iterations 1 and 3 contribute with alpha=0.98). Focused result: 2 passed.

[aaishwarymishra]

@vfdev-5 dont you think output_transform=lambda output: output should be default here like in Average so we dont have to explicitly pass it when declaring running average?