Set higher tolerance for squeezenet1_1 with bf16 (#156920)
Summary:
We need to increase the tolerance slightly to ensure that certain models pass the accuracy check on the XPU device.
This pull request preserves the original tolerance threshold for CUDA/CPU devices and introduces a new key, higher_bf16_xpu, which only affects the XPU device.
X-link: https://github.com/pytorch/pytorch/pull/156920
Approved by: https://github.com/soulitzer
Reviewed By: huydhn
Differential Revision: D77960762
fbshipit-source-id: e035122e3b090575b625c49c34f5a84d958462d8
Author
generatedunixname499836121