[SW-191023][PyTorch][Optimum-Habana-fork]: Add flag to run inference with partial dataset

[pramodkumar-habanalabs]

What does this PR do?

While running the inference with a dataset (ex: Alpaca that has 52000 prompts), executing the full dataset takes longer time. For various experiment purposes running inference with the full dataset is not desirable. Add a command line argument to run the inference with the partial dataset.

Flag ' -- run_partial_dataset' that runs the inference with the dataset for specified --n_iterations(default:5)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

[ghost]

instead of creating a new argument --run_partial_dataset , can you use existing parameters by combining args.dataset_name & args.n_iterations ?

[ssarkar2]

regd: "can you use existing parameters by combining args.dataset_name & args.n_iterations "

Using dataset_name is a bit fragile. This change is to make Alpaca run less number of prompts, but tomorrow there might be some other dataset. Is it feasible/clean to maintain a list of datasets ?

[pramodkumar-habanalabs]

I agree that using 'args.dataset_name' doesn't look like a clean approach. And, '--n_iterations' alone can't be used for partial execution. @ssarkar2 why do we need to maintain a list of datasets and how does it help in the current problem statement.

Also, why using a new arg '--run_partial_dataset' seems problematic. (maintainability ?). BTW, it works well with any dataset, isn't it?

[ghost]

My point was to check if args.dataset_name is not Null and then use that condition with args.n_iterations

I didn't intend to check for alpaca dataset. We should not check for datasets name to any fixed list.