fix bug of attn backward in non-casual model with context parallel open.

[wplf]

This bug will cause bug[ERROR] failed (exitcode: -11) local_rank: 0 (pid: 1761020) of binary: ~/megatron/bin/python.

That is because we miss the rng_states that is required in attention recompute (for dropout) of backward, but no hint is provided.

It is very very very difficult to trace and cost me two weeks.

before the start of training step] datetime: 2024-07-22 18:26:45 
[2024-07-22 18:27:00,941] torch.distributed.elastic.multiprocessing.api: [ERROR] failed (exitcode: -11) local_rank: 0 (pid: 1761020) of binary: /home//miniconda3/envs/megatron/bin/python
Traceback (most recent call last):
  File "/home//miniconda3/envs/megatron/bin/torchrun", line 33, in <module>
    sys.exit(load_entry_point('torch==2.2.1+cu121', 'console_scripts', 'torchrun')())
  File "/home//miniconda3/envs/megatron/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 347, in wrapper
    return f(*args, **kwargs)
  File "/home//miniconda3/envs/megatron/lib/python3.10/site-packages/torch/distributed/run.py", line 812, in main
    run(args)
  File "/home//miniconda3/envs/megatron/lib/python3.10/site-packages/torch/distributed/run.py", line 803, in run
    elastic_launch(
  File "/home//miniconda3/envs/megatron/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 135, in __call__
    return launch_agent(self._config, self._entrypoint, list(args))
  File "/home//miniconda3/envs/megatron/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 268, in launch_agent
    raise ChildFailedError(
torch.distributed.elastic.multiprocessing.errors.ChildFailedError:

Description

Please include a brief summary of the changes, relevant motivation and context.

Fixes # (issue)

Type of change

• [ ] Documentation change (change only to the documentation, either a fix or a new content)
• [ ] Bug fix (non-breaking change which fixes an issue)
• [ ] New feature (non-breaking change which adds functionality)
• [x] Breaking change (fix or feature that would cause existing functionality to not work as expected)
• [ ] Infra/Build change
• [ ] Code refractor

Changes

Please list the changes introduced in this PR:

• Add the missing arguments rng_state in flash_attn_backward with CP open and no casual.

Checklist:

• [x] I have read and followed the contributing guidelines
• [x] The functionality is complete
• [x] I have commented my code, particularly in hard-to-understand areas
• [x] I have made corresponding changes to the documentation
• [x] My changes generate no new warnings
• [x] I have added tests that prove my fix is effective or that my feature works
• [x] New and existing unit tests pass locally with my changes

[ptrendx]

@xrennvidia @cyanguwa Could you both take a look?

[xrennvidia]

LGTM.

This is a good catch, I indeed missed it. Thanks for the fixing.

[ksivaman]

/te-ci pytorch

[cyanguwa]

LGTM as well.

[ptrendx]

Thank you @wplf for your contribution :-)!