GPT-J extra (pad) token

[attafosu]

During the finetuning of the GPT-J model, an extra token (pad="[PAD]") was inadvertently introduced as part of the model, which increased the lm_head dimension by 1 (from 50400 to 50401).

This extra token is now practically redundant (both in dataset preprocessing and post-processing if padding is used during batching).

Creating a thread to discuss making it optional to use 50401 vs 50400 in the lm_head.

[mrmhodak]

WG feedback is that this is okay. No code change is required, this is up to submitters.