significant torchbench regression in cudagraphs configuration

[zou3519]

in last three days.

torch.compile has some serious regressions on some of these models

https://hud.pytorch.org/benchmark/torchbench/inductor_dynamic?dashboard=torchinductor&startTime=Sat,%2008%20Nov%202025%2018:23:46%20GMT&stopTime=Tue,%2011%20Nov%202025%2018:23:46%20GMT&granularity=hour&mode=inference&dtype=bfloat16&deviceName=cuda%20(h100)&lBranch=main&lCommit=8d5cceeb6ac1ef62f34ad86776f31046a250d295&rBranch=main&rCommit=5135ace3a3ca836338201a08404ac36c96b58b7c
https://hud.pytorch.org/benchmark/torchbench/inductor_with_cudagraphs_freezing?dashboard=torchinductor&startTime=Sat,%2008%20Nov%202025%2018:23:46%20GMT&stopTime=Tue,%2011%20Nov%202025%2018:23:46%20GMT&granularity=hour&mode=inference&dtype=bfloat16&deviceName=cuda%20(h100)&lBranch=main&lCommit=8d5cceeb6ac1ef62f34ad86776f31046a250d295&rBranch=main&rCommit=5135ace3a3ca836338201a08404ac36c96b58b7c
https://hud.pytorch.org/benchmark/torchbench/inductor_with_cudagraphs?dashboard=torchinductor&startTime=Sat,%2008%20Nov%202025%2018:22:14%20GMT&stopTime=Tue,%2011%20Nov%202025%2018:22:14%20GMT&granularity=hour&mode=inference&dtype=bfloat16&deviceName=cuda%20(h100)&lBranch=main&lCommit=8d5cceeb6ac1ef62f34ad86776f31046a250d295&rBranch=main&rCommit=5135ace3a3ca836338201a08404ac36c96b58b7c

cc @chauhang @penguinwu

[zou3519]

I did confirm #166342 caused this

[zou3519]

Since that PR was reverted I expect the benchmarks to go back to normal tomorrow. Though we're going to want to be careful with #167564

[colesbury]

@zou3519 can you help me ensure that:

• Rework PyObject preservation (v2) #167564

doesn't introduce a torchbench regression? How do I get torchbench to run on the PR?

[zou3519]

benchmarks are normal right now. Going to keep this issue open since we might need up to 24 hours to see the result of #167564 (I don't know what time the benchmarks are run)