Optimizing batch processing for transformer-based Word Embeddings on GPU

[maziyarpanahi]

The transformer-based word embedding models benefit massively from being computed on GPU devices (locally and in a cluster). However, there is one use case (worst-case scenario) where there is 1 sentence per row. In this scenario, the local GPU device suffers a big performance drawback compared to cluster mode or multiple sentences per row.

This PR follows the work that was done for BERT (word and sentence embeddings) annotators here which improves utilizing GPU device locally for the following model architecture: #6462

• Albert
• DeBERTa
• DistilBERT
• Longformer
• RoBERTa
• XLM-RoBERTa
• XLNet

Done previously:

• BertEmbeddings
• BertSentenceEmbeddings
• CamemBertEmbeddings

Description

Motivation and Context

How Has This Been Tested?

Screenshots (if appropriate):

Types of changes

• Bug fix (non-breaking change which fixes an issue)
• Code improvements with no or little impact
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to change)

Checklist:

• My code follows the code style of this project.
• My change requires a change to the documentation.
• I have updated the documentation accordingly.
• I have read the CONTRIBUTING page.
• I have added tests to cover my changes.
• All new and existing tests passed.