Torch Elastic DDP DeadLock bug fix

[tchaton]

What does this PR do?

This PR introduces some changes to better support multi node dead lock mechanism prevention when processes are created by the cluster.

This works.

MASTER_ADDR="localhost" MASTER_PORT=1234 WORLD_SIZE=2 LOCAL_RANK=0 python tests/plugins/environments/torch_elastic_deadlock.py
MASTER_ADDR="localhost" MASTER_PORT=1234 WORLD_SIZE=2 LOCAL_RANK=1 python tests/plugins/environments/torch_elastic_deadlock.p

TorchElastic detects rank 1 as Terminated as expected. Therefore, rank 0 has succeeded and rank 1 failed. Therefore the run is declared as failed. Depending on use-case, this can be a wrong behaviour.

python -m torch.distributed.run --nproc_per_node=2 --max_restarts 0 tests/plugins/environments/torch_elastic_deadlock.py

TorchElastic should expose a kill function, so it registers that we manual kill a process.

from torch.distributed.elastic.utils import elatic_kill

Does your PR introduce any breaking changes? If yes, please list them.

Before submitting

• Was this discussed/approved via a GitHub issue? (not for typos and docs)
• Did you read the contributor guideline, Pull Request section?
• Did you make sure your PR does only one thing, instead of bundling different changes together?
• Did you make sure to update the documentation with your changes? (if necessary)
• Did you write any new necessary tests? (not for typos and docs)
• Did you verify new and existing tests pass locally with your changes?
• Did you list all the breaking changes introduced by this pull request?
• Did you update the CHANGELOG? (not for typos, docs, test updates, or internal minor changes/refactorings)

PR review

Anyone in the community is welcome to review the PR.
Before you start reviewing make sure you have read Review guidelines. In short, see the following bullet-list:

• Is this pull request ready for review? (if not, please submit in draft mode)
• Check that all items from Before submitting are resolved
• Make sure the title is self-explanatory and the description concisely explains the PR
• Add labels and milestones (and optionally projects) to the PR so it can be classified

Did you have fun?

Make sure you had fun coding 🙃

[pep8speaks]

Hello @tchaton! Thanks for updating this PR. We checked the lines you've touched for PEP 8 issues, and found:

There are currently no PEP 8 issues detected in this Pull Request. Cheers! 🍻

Comment last updated at 2021-08-02 16:38:39 UTC

[codecov]

Codecov Report

Merging #8655 (fd52122) into master (85bba06) will decrease coverage by 4%.
The diff coverage is 73%.

@@           Coverage Diff           @@
##           master   #8655    +/-   ##
=======================================
- Coverage      93%     89%    -4%     
=======================================
  Files         168     167     -1     
  Lines       13984   13970    -14     
=======================================
- Hits        12948   12397   -551     
- Misses       1036    1573   +537     

[tchaton]

Hey @ananthsub.

I will add a Multi Node test running on Grid using Torch Elastic soon, so our support for Torch Elastic is higher.

Best,
T.C

[SeanNaren]

More important to get this in even if there is no test atm I think