Port pytorch/benchmarks/dynamo to the dynamo userbenchmark

[xuzhao9]

Copy the code from https://github.com/pytorch/pytorch/tree/main/benchmarks/dynamo to make it as a torchbench userbenchmark.
Need to refactor the code to work on pytorch nightly release.

Test plan:

$ python run_benchmark.py dynamo --performance --only resnet50 --training
WARNING:userbenchmark.dynamo.common:torch.cuda.is_available() == False, using CPU
loading model: 0it [00:18, ?it/s]
cpu  train resnet50
 15/ 16 +1 frames  -2s 12 graphs  3 graph calls  231/ 581 =  40% ops  32% time

$ python run_benchmark.py dynamo --performance --only resnet50 --training
loading model: 0it [00:06, ?it/s]
cuda train resnet50
 22/ 22 +0 frames   2s  4 graphs  4 graph calls  233/ 586 =  40% ops  51% time

[FindHao]

For now, it only works with --only model_name, right? How can we run all models with this userbenchmark?

[xuzhao9]

@FindHao
Yes, it could iterate through all models, just remove the --only argument.

[facebook-github-bot]

@xuzhao9 has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

[FindHao]

Only serveral models failed the testing for training. Here is the log. run_log.txt
LGTM!

[xuzhao9]

It is fine, we can fix them when dynamo team is really using this.

[facebook-github-bot]

@xuzhao9 merged this pull request in ee5266b.