Create nightly job for torch.compile benchmarks

[sachanub]

Description

Please read our CONTRIBUTING.md prior to creating your first pull request.

Creating PR to enable nightly benchmarks to compare torch.compile default mode and TorchScript for the following models:

• BERT Sequence Classification without Better Transformer. Please note that the BERT model artifacts will be updated again later to choose the one with the best performance (eg. with Better Transformer etc)
• VGG-16
• ResNet-50

Added new model artifacts for each of the above mentioned models:

• https://torchserve.pytorch.org/mar_files/bert-scripted.mar
• https://torchserve.pytorch.org/mar_files/bert-default.mar
• https://torchserve.pytorch.org/mar_files/vgg-16-scripted.mar
• https://torchserve.pytorch.org/mar_files/vgg-16-default.mar
• https://torchserve.pytorch.org/mar_files/resnet-50-scripted.mar
• https://torchserve.pytorch.org/mar_files/resnet-50-default.mar

Created Cloudwatch dashboard for the above mentioned models: https://cloudwatch.amazonaws.com/dashboard.html?dashboard=TorchServe-Benchmark-Torch-Compile-Nightly-GPU&context=eyJSIjoidXMtZWFzdC0xIiwiRCI6ImN3LWRiLTA4NDQ5NTcyODMxMSIsIlUiOiJ1cy1lYXN0LTFfdTlGeElpOXh1IiwiQyI6IjRjMm52MzFlZGN2MWFvdjY3Z2hoYmJjMTF2IiwiSSI6InVzLWVhc3QtMTozNzk2MDVjMC0xNDM5LTQzMTEtOWE4Ny03YTUwY2E0ODVmMjAiLCJPIjoiYXJuOmF3czppYW06OjA4NDQ5NTcyODMxMTpyb2xlL3NlcnZpY2Utcm9sZS9DV0RCU2hhcmluZy1QdWJsaWNSZWFkT25seUFjY2Vzcy1QQjRaVEo2QiIsIk0iOiJQdWJsaWMifQ==

[agunapal]

Looks good overall.
Few points

1. Where are we installing pytorch nightlies?
2. I think going forward, we should perhaps only measure BERT with BT. We should show the best possible numbers and not confuse customers.

[sachanub]

Looks good overall. Few points

1. Where are we installing pytorch nightlies?
2. I think going forward, we should perhaps only measure BERT with BT. We should show the best possible numbers and not confuse customers.

1. The pytorch nightlies are installed via the install_dependencies.py script. I have added an argument in auto_benchmark.py which enables the installation of nightly torch packages via the install_dependencies.py script.
2. For the scope of this PR, do we want to update the BERT model artifacts?

[agunapal]

LGTM