Skip to content

test(dist): assert reload replays new_group timeout and options - #2284

Open
EazyReal wants to merge 1 commit into
THUDM:mainfrom
EazyReal:test/reload-preserves-new-group-options
Open

test(dist): assert reload replays new_group timeout and options#2284
EazyReal wants to merge 1 commit into
THUDM:mainfrom
EazyReal:test/reload-preserves-new-group-options

Conversation

@EazyReal

Copy link
Copy Markdown
Contributor

What

CPU regression test that reload_process_groups() calls new_group with the original timeout, pg_options, and group_desc, not {ranks, backend="nccl"}.

Why

Follow-up to #2095. The constructor replay landed in #2208; this is the assertion #2095 carried that did not.

#2208's multiprocess test checks that a reloaded group can still barrier. That still passes if reload drops a 2-hour NCCL timeout and falls back to PyTorch's default — the offload/heal watchdog bug. This test spies on the constructor.

Verified locally: passes on current main; fails if reload is patched back to old_new_group(ranks=..., backend="nccl").

Test

pytest tests/test_reloadable_process_group_world.py::test_reload_replays_original_new_group_timeout_and_options -q

Made with Cursor

The WORLD reload tests only check that a reloaded group can still barrier. That still passes if reload drops timeout/pg_options and falls back to {ranks, backend=nccl}, which is the offload/heal watchdog bug.

Co-authored-by: Cursor <cursoragent@cursor.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant