V3 Performance Signal Detected by TorchBench Userbenchmark "torch-nightly" on '871fc7bb76f05c3c487214404f687cf7a6a8e453'

[github-actions[bot]]

TorchBench CI has detected a performance signal or runtime regression.

Base PyTorch commit: 61239df555df02e8c60a2ad63363878a2a57c161

Affected PyTorch commit: 871fc7bb76f05c3c487214404f687cf7a6a8e453

Affected Tests:

• test_eval[DALLE2_pytorch-cuda-eager]_latency: +8.31993%
• test_eval[doctr_det_predictor-cuda-eager]_cmem: +10.12839%
• test_eval[timm_resnest-cuda-eager]_gmem: +15.90606%
• test_train[torchrec_dlrm-cuda-eager]_cmem: +32.42321%

Tests that were no longer run on affected commit:

Tests that were newly added on affected commit:

Runtime regressions found? No runtime errors were found in the new benchmarks run--you are all good there!

GitHub workflow that triggered this issue: https://github.com/pytorch/benchmark/actions/runs/5047970627

cc @xuzhao9

[xuzhao9]

2023-05-22