Fixes gpt mcore conversion to account for _extra_state that may be present

[terrykong]

What does this PR do ?

Fixes gpt mcore conversion to account for _extra_state that may be present

Collection: [Note which collection this PR will affect]

Changelog

In the 24.01 container, the following:

wget https://huggingface.co/nvidia/GPT-2B-001/resolve/main/GPT-2B-001_bf16_tp1.nemo
mkdir 2b_model_checkpoint && tar -xvf GPT-2B-001_bf16_tp1.nemo -C 2b_model_checkpoint
docker run -v $PWD/2b_model_checkpoint:/inputs -v $PWD:/outputs --rm -it nvcr.io/ea-bignlp/ga-participants/nemofw-training:24.01 python /opt/NeMo/scripts/nlp_language_modeling/convert_nemo_gpt_to_mcore.py --in-folder /inputs --out-file /outputs/2b_mcore_gpt.nemo --cpu-only

will result in this error:

[NeMo I 2024-03-08 23:08:38 nlp_overrides:1108] Model MegatronGPTModel was successfully restored from /outputs/2b_mcore_gpt.nemo.
[NeMo I 2024-03-08 23:08:38 convert_nemo_gpt_to_mcore:260] Sanity checks:
[NeMo I 2024-03-08 23:08:38 convert_nemo_gpt_to_mcore:265] ✅ Number of weights match
[NeMo W 2024-03-08 23:08:38 nemo_logging:349] /usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py:1876: UserWarning: Positional args are being deprecated, use kwargs instead. Refer to https://pytorch.org/docs/master/generated/torch.nn.Module.html#torch.nn.Module.state_dict for details.
      warnings.warn(
    
[NeMo I 2024-03-08 23:08:42 convert_nemo_gpt_to_mcore:292] ✅ Weights match
Traceback (most recent call last):
  File "/opt/NeMo/scripts/nlp_language_modeling/convert_nemo_gpt_to_mcore.py", line 326, in <module>
    run_sanity_checks(input_nemo_file, output_nemo_file, cpu_only=cpu_only, ignore_if_missing=ignore_if_missing)
  File "/opt/NeMo/scripts/nlp_language_modeling/convert_nemo_gpt_to_mcore.py", line 295, in run_sanity_checks
    assert len(nemo_state_dict) == 0, f"❌ unexpected items in nemo_state_dict: {nemo_state_dict}"
AssertionError: ❌ unexpected items in nemo_state_dict: OrderedDict([('model.language_model.encoder.layers.0.self_attention.query_key_value._extra_state', None), ('model.language_model.encoder.layers.0.self_attention.dense._extra_state', None), ('model.language_model.encoder.layers.0.mlp.dense_h_to_4h._extra_state', None), ('model.language_model.encoder.layers.0.mlp.dense_4h_to_h._extra_state', None), ('model.language_model.encoder.layers.1.self_attention.query_key_value._extra_state', None), ('model.language_model.encoder.layers.1.self_attention.dense._extra_state', None), 

This change just updates the check to ignore keys with "_extra_state". FWIW, the nemo_state_dict has None for these _extra_state keys and mcore_state_dict is an empty io.bytes object:

mcore_state_dict[ 'model.decoder.layers.15.mlp.linear_fc1._extra_state']
<_io.BytesIO object at 0x7fc01ea8a020>
(Pdb) mcore_state_dict[ 'model.decoder.layers.15.mlp.linear_fc1._extra_state'].read()
b''

Usage

• You can potentially add a usage example below

# Add a code snippet demonstrating how to use this 

Jenkins CI

To run Jenkins, a NeMo User with write access must comment jenkins on the PR.

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)

[cuichenx]

LGTM, thanks for fixing!