[Draft] Fix optimizer state_dict compatibility between Apex's fused_adam and distributed_fused_adam

[paul-gibbons]

Description

The use case I am supporting is the ability to take model checkpoints trained with fused_adam optimizer and then resume training with Apex's distributed_fused_adam while maintaining optimizer state.

There is mismatch in the optimizer state_dict where MoE-specific keys, which are ShardedTensorFactory objects and not ShardedTensor's, are not caught in the rename_fp32_params function, causing the moe keys to fall under the missing and unexpected keys when loading the optimizer state_dict.

Additionally, when resuming training from a model using fused_adam, state_dict['param_groups'] does not contain any keys aside from params. The param_groups need to be pulled from state_dict['optimizer'].

I override load_state_dict in NeMo's MegatronDistributedFusedAdam and manipulate the optimizer's state_dict if we see state_dict['param_groups'] only contains params key.

This behavior can be reproduced by training any MoE model with fused_adam and then attempt to resume training with distributed_fused_adam.

Additionally,

What does this PR do ?

Add a one line overview of what this PR aims to accomplish.

Collection: [Note which collection this PR will affect]

Changelog

• Add specific line by line info of high level changes in this PR.

Usage

• You can potentially add a usage example below

# Add a code snippet demonstrating how to use this 

GitHub Actions CI

The Jenkins CI system has been replaced by GitHub Actions self-hosted runners.

The GitHub Actions CI will run automatically when the "Run CICD" label is added to the PR.
To re-run CI remove and add the label again.
To run CI on an untrusted fork, a NeMo user with write access must first click "Approve and run".

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation?
• Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)
  • Reviewer: Does the PR have correct import guards for all optional libraries?

PR Type:

• New Feature
• Bugfix
• Documentation

If you haven't finished some of the above items you can still open "Draft" PR.

Who can review?

Anyone in the community is free to review the PR once the checks have passed.
Contributor guidelines contains specific people who can review PRs to various areas.

Additional Information

• Related to # (issue)