[Fix] gptoss yarn parameter

[yiakwy-xpu-ml-framework-team]

What does this PR do ?

Fix gptoss GptOSS yarn default parameters when passing them to mcore where mcore does not accept None as valid vlaues.

Changelog

• Fix GptOss Yarn default parameters : mcore cannot read none values

GitHub Actions CI

See the CI sectionin the Contributing doc for how to trigger the CI. A Nvidia developer will need to approve and trigger the CI for external contributors.

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?

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

Additional Information

• Related to [Feature] add gptoss continue train bf16-fp8 (sft) example [part1 - mcore] NVIDIA/Megatron-LM#2383

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[yiakwy-xpu-ml-framework-team]

Doctor sergiopperez Could you have a look at this ?

This is essentail for support of GptOSS traininig D.

[yaoyu-33]

why is this change? please attach a note in the comment if the change is necessary

[yiakwy-xpu-ml-framework-team]

by default gloo will be used to create TCP connection in CPU side.

But we don't need it. It should be assumed to be Flase.

[yiakwy-xpu-ml-framework-team]

Will roll back to True

[maanug-nv]

updated version of this PR at #2413

[maanug-nv]

Sorry, I misinterpreted some settings in MCore , so #2413 is invalid. but yarn_mscale and yarn_mscale_all_dim are intentionally none (cc @cuichenx), and use_gloo_process_groups should stay as True to match initialize_model_parallel() defaults in MCore. So I don't think any of the changes in this PR are desired.