FIx Llama3.1/Llama3.2 Exporter

[suiyoubi]

Important

The Update branch button must only be pressed in very rare occassions.
An outdated branch is never blocking the merge of a PR.
Please reach out to the automation team before pressing that button.

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The Jenkins CI system has been replaced by GitHub Actions self-hosted runners.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

[aflah02]

Hi @suiyoubi
I just saw this PR and I have also been facing issues when converting from .distcp checkpoints saved during pretraining to HF compatible ones. Does this PR intend to address that?

[suiyoubi]

Hi @aflah02 , this PR fixes the converted HF config only. Not the weight files.

Specifically, rope_scaling is not exported properly when converting the model to HF. The weight file is unchanged.

[aflah02]

ah gotcha, thanks!

[aflah02]

@suiyoubi Is there any way right now to convert a distcp llama 3.1 config based model's intermediate training checkpoint to nemo format? I can't see any way. Was wondering if you knew any since you're working on this issue

I have raised an issue here - #12381

[suiyoubi]

@suiyoubi Is there any way right now to convert a distcp llama 3.1 config based model's intermediate training checkpoint to nemo format? I can't see any way. Was wondering if you knew any since you're working on this issue

I have raised an issue here - #12381

What distcp do you have ? Currently we only officially support NeMo1/2 <=> HF checkpoint conversion

[aflah02]

Hi @suiyoubi

I basically used a variant of this script which saves intermediate checkpoints which saves in distcp format - https://github.com/NVIDIA/NeMo/blob/main/nemo/collections/llm/recipes/llama32_1b.py
These are the checkpoints from one of the runs - https://huggingface.co/aflah/llama32_1b_dclm-SL-2048-PGBS-16-GAS-4-NGPU-8-NNODES-1-TW-PERF/tree/main

I also tried to use the NeMoModelCheckpoint callback (https://github.com/NVIDIA/NeMo/blob/8d6b58c16e5568cb0755aab163d5d191cc23475a/nemo/utils/callbacks/nemo_model_checkpoint.py#L37) but it crashes due to an attribute error -

• Code: https://gist.github.com/aflah02/f5a8605dd4df2cf0c6abd1114f6cf5a7
• Logs: https://wandb.ai/aflah/Nemo_Testing/runs/spmoq2c3/logs?nw=nwuseraflah, https://wandb.ai/aflah/Nemo_Testing/runs/cx2ploh9/logs?nw=nwuseraflah

So essentially the checkpoints cannot be used at all making the training useless as the checkpoints are not shareable

[github-actions]

[🤖]: Hi @suiyoubi 👋,

We wanted to let you know that a CICD pipeline for this PR just finished successfully

So it might be time to merge this PR or get some approvals

I'm just a bot so I'll leave it you what to do next.

//cc @pablo-garay @ko3n1g

[cuichenx]

LGTM