Fix LoRA contiguous tensor

[cuichenx]

What does this PR do ?

Fix contiguous tensor issue in LoRA with mbs>1

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The Jenkins CI system has been replaced by GitHub Actions self-hosted runners.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

[stevehuang52]

I'm seeing some strange warnings, are they normal?

1. not able to find val_loss, where does the model call self.log or self.log_dict?:

[NeMo W 2024-09-25 20:12:48 nemo_logging:349] /usr/local/lib/python3.10/dist-packages/pytorch_lightning/callbacks/model_checkpoint.py:384: `ModelCheckpoint(monitor='val_loss')` could not find the monitored key in the returned metrics: ['lr', 'consumed_samples', 'global_batch_size', 'global_step', 'step', 'reduced_train_loss', 'grad_norm', 'epoch']. HINT: Did you call `log('val_loss', value)` in the `LightningModule`?

2. torch/_dynamo/convert_frame.py:

Validation: iteration 4/5
[rank0]:W0925 20:09:59.218000 140276111091520 torch/_dynamo/convert_frame.py:744] [4/8] torch._dynamo hit config.cache_size_limit (8)
[rank0]:W0925 20:09:59.218000 140276111091520 torch/_dynamo/convert_frame.py:744] [4/8]    function: 'calculate_cross_entropy_loss' (/opt/megatron-lm/megatron/core/fusions/fused_cross_entropy.py:47)
[rank0]:W0925 20:09:59.218000 140276111091520 torch/_dynamo/convert_frame.py:744] [4/8]    last reason: tensor 'L['exp_logits']' size mismatch at index 0. expected 224, actual 144
[rank0]:W0925 20:09:59.218000 140276111091520 torch/_dynamo/convert_frame.py:744] [4/8] To log all recompilation reasons, use TORCH_LOGS="recompiles".
[rank0]:W0925 20:09:59.218000 140276111091520 torch/_dynamo/convert_frame.py:744] [4/8] To diagnose recompilation issues, see https://pytorch.org/docs/main/torch.compiler_troubleshooting.html.
[NeMo W 2024-09-25 20:09:59 nemo_logging:349] /opt/megatron-lm/megatron/core/tensor_parallel/layers.py:609: UserWarning: async_grad_allreduce is deprecated, not in use anymore and will be fully removed with 0.10.0. Please use allreduce_dgrad instead.
      warnings.warn(

[cuichenx]

These are not related to the PR so let's discuss elsewhere.

[stevehuang52]

LGTM, thanks!

[cuichenx]

adding CI tests in separate PR: #10632