Compétences
pytorch-pytorch

deprecate-feature

Deprecate, freeze, or remove PyTorch APIs; choose a removal version, handle won't-fix frozen components, or audit overdue removals.
étoiles105007
forks32477
observateurs1803
mise à jour2026-10-23 18:25:03

Deprecations

This process is provisional; we are evaluating it in practice. Human judgment wins: skip steps that do not apply. New APIs that are technically public but barely used may be deprecated or removed faster at the human's discretion.

The BC policy is the source of truth:

  • Read release N from version.txt.
  • Warn for at least two releases (180 days for nightly users); earliest removal is N+2.

Deprecate or freeze

  • State the start version, removal target, replacement (or how to retain current behavior), and tracking issue URL in the warning.

  • Warn once per process. Python: FutureWarning behind a module-level flag:

    _warned = False

def foo():
    global _warned
    if not _warned or torch.is_warn_always_enabled():
        warnings.warn(message, FutureWarning, stacklevel=2)
        _warned = True

  • Use @deprecated only with category=None for type-checker diagnostics.

  • C++: prefer TORCH_WARN_ONCE; TORCH_WARN_DEPRECATION repeats and maps to hidden DeprecationWarning (#72948).

  • Add docstring .. deprecated:: <N>, migrate internal callers, test the warning (under set_warn_always_context(True)), and label the PR topic: deprecation.

  • Frozen components also need tracking issues that record the freeze decision; link it when closing issues against the component. Use Kind: Frozen component, with Removal target and Next step: none if no removal is planned.

Track

  • Reuse or create one issue per item via .github/ISSUE_TEMPLATE/deprecation.yml. Keep its ### <field> headings.

  • Get user approval of the exact text before creating or editing any issue (AI_POLICY.md). Label it module: deprecation; use slowroll only for staged opt-in rollouts (#191236).

  • Attach it as a sub-issue of #201298 using its internal id, not number:

    id=$(gh api repos/pytorch/pytorch/issues/<N> --jq .id)
gh api -X POST repos/pytorch/pytorch/issues/201298/sub_issues -F sub_issue_id="$id"

  • Keep Next step current (e.g. remove in 2.18).

  • Link PRs with Part of #N, or Fixes #N for the final step.

Remove or audit

  • Check that two warning releases shipped and the removal target arrived.

  • Remove code, warnings, tests, and references, including .pyi.in, docs, and torch/overrides.py. Use .. versionchanged:: <version> with migration guidance where docs remain; label the PR topic: bc breaking.

  • Audit targets against version.txt; remove overdue items or retarget with a reason, tracking each under #201298:

    git grep -nIE "(removed|deleted|dropped|become an error) (in|by|after) (PyTorch |version |release |v)?[0-9]+\.[0-9]+" -- torch aten c10