Minor MPT-7B fixes and creation script update

[trias702]

What does this PR do ?

Updates the script used to create a Megatron MPT-7B checkpoint with fixes, and adds support to MegatronBaseModel for special tokens to be passed at runtime.

Collection: NLP

Changelog

• Updates the Megatron MPT-7B creation script with some bug fixes and a config file baseline
• Adds support for HuggingFace special tokens to be declared in Hydra/Yaml and passed to MegatronBaseModel

Usage

You can specify the new tokens in a config YAML like so:

...
model.tokenizer.special_tokens.pad_token="<PAD>"
model.tokenizer.special_tokens.mask_token="<MASK>"
...

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

The MegatronBaseModel changes are necessary because the HuggingFace tokenizer which is used by MPT-7B (and indeed all other HF tokenizers) does not have special tokens built in, tokens such as PAD and MASK. This leads to a crash when prompt-learning with MPT-7B, because there is no pad token, and no way to add one at runtime. With this fix to MegatronBaseModel, you can now pass a pad token at runtime, which allows prompt-learning with MPT-7B (or any model which uses a HF tokenizer).

[ericharper]

LGTM. Thanks!