feat: Nemotron Nano-v3 pipeline parallelism

[prestonfu]

What does this PR do?

Single-node pipeline parallelism for Nemotron NanoV3 30B.

Changelog

• parallelizer.py: Unpack ModuleList/ModuleDict in layer extraction
• functional.py:
  • Support backbone.* model structure (vs model.*).
  • Add stage_model.to_empty(device=device) to enable devicce storage for buffers such as e_score_correction_bias in MoE, which are otherwise on CPU.
• hf_utils.py: Support backbone and backbone.embeddings (vs embed_tokens)
• flops_utils.py: An (incorrect) attempt to calibrate Mamba2 SSM FLOPs
• train_ft.py:
  • Pass trust_remote_code to AutoConfig
  • Add checkpoint.enabled
  • Add MFU and nsys support

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?

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

Additional Information

• Related to # (issue)

[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.

[akoumpa]

@ZhiyuLi-Nvidia can you take a look? Thank you

[ZhiyuLi-Nvidia]

Hi, @prestonfu thanks a lot for contribution. I am just curious why you want to merge into this dev branch NVIDIA-NeMo:zhiyul/llm-optimization-workshop, which is for UCB homework only.
Are you interested in contributing into main branch instead?