Fix OOMing periodic shards (#95246)
Tests have been consistently failing with the error on the following shards with the error `RuntimeError: CUDA error: out of memory`
- `periodic / linux-bionic-cuda11.7-py3-gcc7-slow-gradcheck / test (default, 1, 2, linux.4xlarge.nvidia.gpu)`
- `periodic / linux-bionic-cuda11.7-py3-gcc7-slow-gradcheck / test (default, 2, 2, linux.4xlarge.nvidia.gpu)`
Seeing if serializing those test files makes the periodic jobs succeed again. This feels a bit off since there are so many different test files that have failed and need to be serialized, indicating a potential perf regression somewhere
Failures on hud: https://hud.pytorch.org/hud/pytorch/pytorch/master/1?per_page=100&name_filter=periodic%20%2F%20linux-bionic-cuda11.7-py3-gcc7-slow-gradcheck%20%2F%20test%20(default%2C%20
Pull Request resolved: https://github.com/pytorch/pytorch/pull/95246
Approved by: https://github.com/Skylion007, https://github.com/huydhn
