Skip to content

[torch_checkpointing][test] use real barrier for test - #4574

Open
aelavender wants to merge 1 commit into
pytorch:mainfrom
aelavender:pr4573
Open

[torch_checkpointing][test] use real barrier for test#4574
aelavender wants to merge 1 commit into
pytorch:mainfrom
aelavender:pr4573

Conversation

@aelavender

@aelavender aelavender commented Sep 10, 2026

Copy link
Copy Markdown

This test does CheckpointWriterConfig(barrier_config=None), which I'm about to disallow without further config.

  • In meta-pytorch/torch_checkpointing@32f50c1, we change the default barrier from None to DefaultStoreBarrier. This is does not affect Titan; it currently explicitly sets up a TCPStoreBarrier.
  • Current torch_checkpointing behaviour is to write to a temp dir and atomically rename iff a barrier is configured. I am soon to make this an explicit flag (defaulting to True), which will make this test break. So fixing the test before I make the api change.

Tested with pytest ./tests/unit_tests/cpu/test_torch_checkpointing.py

@meta-cla meta-cla Bot added the CLA Signed This label is managed by the Meta Open Source bot. label Sep 10, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

CLA Signed This label is managed by the Meta Open Source bot.

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants