[RMP] T4R - Support to multi-GPU training for binary classification and regression tasks

[gabrielspmoreira]

Problem:

Transformers4Rec supports multi-GPU training for the next-item prediction task because it uses the HF Trainer (RMP #522), which under-the-hood supports DataParallel / DistributedDataParallel.
The binary classification / regression tasks are currently not supported by HF Trainer, but rather trained with a custom model.fit() method we provided, that doesn't support DataParallel / DistributedDataParallel.

Goal:

• Change the implementation of binary classification / regression tasks so that they can be trained (with multi-GPU) by using HF Trainer.

Starting Point:

• [Task] Standardize the model output format.  Transformers4Rec#505
• Note 505 tracks the dev in 487 and 423
• [Task] Support multi-GPU training for all prediction tasks Transformers4Rec#487
• Binary classification - [QST] - Multi-GPU Support w/ Naive model.fit() Transformers4Rec#423 - This a customer issue that will be solved by above issues

[viswa-nvidia]

@rnyak needs to confirm with @sararb and confirm milestone

[rnyak]

We created a multi-gpu example in TF4Rec for next-item prediction task. For binary-classification task we shared a multi-gpu training example PoC with the customer, and I will add code snippet into the docs in TF4Rec repo.