[autorevert] dispatch untargeted signals without tests-to-include filter

[izaitsevfb]

Summary

Two related fixes to the autorevert lambda's tests-to-include dispatch path. Both surface the same anti-pattern: a signal that cannot be narrowed to a run_test.py-recognized module should be dispatched as a "re-run the whole job(s)" restart, not as a TEST-narrow restart.

Bug 1 — empty-file rows in tests.all_test_runs produce bogus tests-to-include entries

TestRow.test_id (signal_extraction_types.py:138-141) builds the test key as f"{file}::{name}" if file else name. When the CH row has empty file, the test key falls back to just the bare name (no ::).

signal_extraction.py then derived test_module from the key:

test_module = test_id.split("::")[0].replace(".py", "")

With no ::, test_module becomes the bare method name (e.g. test_partial_eval_graph_conv). The lambda forwards that into workflow_dispatch.tests-to-include, and pytorch/pytorch:test/run_test.py:144 TestChoices.__contains__ rejects it with:

argument -i/--include: invalid choice: 'test_partial_eval_graph_conv'

The dispatched shard fails before running any tests, surfacing as 15-25 spurious test-osdc failures per affected commit on HUD trunk. Tracking: pytorch/pytorch-gha-infra#1137.

Verified primary-source: tests.all_test_runs for name='test_partial_eval_graph_conv' over the last 2 days returns three distinct file values: 'test_jit.py', 'test_jit_legacy.py', ''.

Concrete instance — pytorch/pytorch run 25294380825 (workflow_dispatch by pytorch-auto-revert[bot], 2026-05-03 23:51 UTC):

jobs-to-include: linux-jammy-py3.10-clang18
tests-to-include: test_jit test_freeze_module_detach_gradient test_partial_eval_graph_conv test_freeze_interface_swapping_two_methods test_partial_eval_stitching test_returning_input_symbolic_shapes

test_jit is the actual file name; the other five are method names that leaked through the empty-file fallback. The default / crossref / dynamo_wrapped shards on this run failed at argparse; the openreg / einops shards were green only because their test.sh paths bypass $INCLUDE_CLAUSE entirely.

Bug 2 — JOB+TEST signals on the same (workflow, commit) silently lose JOB intent

signal_actions.py::group_actions keys the restart map on (workflow_name, commit_sha). ALL contributing signals — JOB-track and TEST-track — are merged into a single ActionGroup with the union of jobs_to_include and tests_to_include. A JOB-track signal (test_module=None, intent = "re-run the whole job") coexisting with a TEST-track signal in the same group becomes a TEST-narrow restart: only the test modules contributed by the TEST signals run, and the JOB signal's full-job-restart intent is silently thrown away.

This was a latent bug independent of Bug 1 — surfaced while auditing the tests-to-include path.

Fix

1. signal_extraction.py: when test_id lacks ::, set test_module=None. The signal stays alive but is marked untargeted.
2. signal_actions.py::group_actions: if any source in the group has test_module is None, dispatch with tests_to_include=frozenset() for the whole group.

JOB-track signals (always test_module=None) and TEST-track signals with no extractable module both reach the new branch, so they get identical "re-run the whole job(s)" treatment instead of being narrowed by sibling targetable TEST signals. workflow_checker.py:165 already drops the tests-to-include input when the frozenset is empty — no change needed there.

Test plan

• tests/test_signal_extraction.py (2 new): empty file → Signal with test_module=None; populated file → existing module path.
• tests/test_signal_actions.py (4 new in TestGroupActionsTestsToInclude): only-targeted-tests keeps filter; TEST-with-no-module drops filter; mixed targeted+untargeted TEST drops filter; JOB+TEST same-(wf,sha) drops filter.
• make test — all 146 tests pass (1 skipped due to no GITHUB_TOKEN).
• ruff format --check + ruff check clean.

Audit follow-up (not in this PR)

Worth checking which test uploader writes empty-file rows to tests.all_test_runs — likely a partial-export issue from a specific reporter — to fix the data leak at source. Open question carried in iz2-memory:core/knowledge/autorevert-bot-retry.md.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
torchci	Ignored		May 4, 2026 10:15pm

[huydhn]

LGTM!