Distributed optimizer aligns FP8 transposes over pipeline-parallel group

[timmoon10]

What does this PR do ?

When training GPT in FP8 with interleaved pipeline parallelism, we have experienced some GPU idling due to load imbalances across the pipeline parallel group. The first microbatch in a training step requires computing an FP8 weight transpose, which is reused in later microbatches. The fix is to align the FP8 weight transposes across the pipeline parallel group, similar to how we align param all-gathers and grad reduce-scatters (see https://github.com/NVIDIA/apex/pull/1611 and https://github.com/NVIDIA/NeMo/pull/6183).

Collection: NLP

Changelog

• Distributed optimizer aligns FP8 transposes over pipeline-parallel group

Usage

Run GPT, e.g. with the config at https://github.com/NVIDIA/NeMo/blob/main/examples/nlp/language_modeling/conf/megatron_gpt_config.yaml.

The following settings are required for this optimization:

• model.virtual_pipeline_model_parallel_size=<layers per virtual pipeline stage>
• model.transformer_engine=True
• model.fp8=True
• model.fp8_params=True
• model.optim.name=distributed_fused_adam
• model.optim.overlap_param_sync=True

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The Jenkins CI system has been replaced by GitHub Actions self-hosted runners.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

• [x] 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:

• [x] 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

[github-actions[bot]]

This PR is stale because it has been open for 14 days with no activity. Remove stale label or comment or update or this will be closed in 7 days.