add llama example

[prathikr]

This PR adds run_classification.py example with ONNX Runtime accelerator classes. The example was taken from Hugging Face's example and was modified to use ORTTrainer/ORTTrainingArguments. Additionally, the PR adds observed performance metrics for meta-llama/Llama-2-7b-hf.

[prathikr]

@JingyaHuang can you please review?

[JingyaHuang]

Hi @prathikr, thanks for adding the example, I put some minor nits. The classification example looks good to me, and I wonder if a causal language modeling fine-tuning example could also be beneficial to the community since llama2 is a generative model.

[prathikr]

Hi @prathikr, thanks for adding the example, I put some minor nits. The classification example looks good to me, and I wonder if a causal language modeling fine-tuning example could also be beneficial to the community since llama2 is a generative model.

@JingyaHuang run_clm.py is already available here, though it hasn't been updated in a while. Is there anyone specific whose duty it is to update the optimum examples to stay in step with transformers examples? All the examples are the same except for the usage of ORTTrainer/ORTTrainingArguments.

[JingyaHuang]

Hi @prathikr, it's true that the training examples of ort training last updated a while ago. I will update them this week, but in long term, it would be hard for the team to keep examples always updated with transformers releases, the best way to automate the process would be to build two CIs in optimum repo:

• 1st CI: clone transformers repo and replace training examples with ORT APIs(ORTTrainer and ORTTrainingArguments) and open a pull request in the optimum repo.
• 2nd CI: to valid all updated examples (probably wth test_examples.py which also need to be updated by adding all examples we have now)

Ping @fxmarty for some inputs, probably you have a better idea.

[prathikr]

Hi @prathikr, it's true that the training examples of ort training last updated a while ago. I will update them this week, but in long term, it would be hard for the team to keep examples always updated with transformers releases, the best way to automate the process would be to build two CIs in optimum repo:

• 1st CI: clone transformers repo and replace training examples with ORT APIs(ORTTrainer and ORTTrainingArguments) and open a pull request in the optimum repo.
• 2nd CI: to valid all updated examples (probably wth test_examples.py which also need to be updated by adding all examples we have now)

Ping @fxmarty for some inputs, probably you have a better idea.

Thanks @JingyaHuang for getting the ball rolling on this. We have internal pipelines that depends on optimum and transformers examples being the same, so this will be very useful. Can we follow up on this task in a separate issue/PR and merge this example for now?

[fxmarty]

LGTM!