TIMM cudagraphs_freezing inference regression

[zou3519]

https://hud.pytorch.org/benchmark/timm_models/inductor_with_cudagraphs_freezing?dashboard=torchinductor&startTime=Mon,%2016%20Dec%202024%2020:49:27%20GMT&stopTime=Wed,%2015%20Jan%202025%2020:49:27%20GMT&granularity=day&mode=inference&model=lcnet_050&dtype=bfloat16&deviceName=cuda%20(a100)&lBranch=main&lCommit=1dab79470dbecef79ba4c7d4308d8a181091e58e&rBranch=main&rCommit=297ce776363cc4802fa74d210fced2b4128960d5

This model used to pass sometime in the last year but is now failing with an accuracy issue

cc @ezyang @gchanan @kadeng @msaroufim @mcarilli @eellison @penguinwu @BoyuanFeng @chauhang

[IvanKobzarev]

Repro:

python benchmarks/dynamo/timm_models.py --only lcnet_050 --accuracy --no-translation-validation --inference --bfloat16 --backend inductor --device cuda --freezing

cuda eval  lcnet_050
E0117 06:11:29.783000 3580783 torch/_dynamo/utils.py:2734] RMSE (res-fp64): 0.12467, (ref-fp64): 0.01990 and shape=torch.Size([8, 1000]). res.dtype: torch.bfloat16, multiplier: 3.000000, tol: 0.001000, use_larger_multiplier_for_smaller_tensor: 0
fail_accuracy