Eca_halonext26ts AMP_BF16 training accuracy got failed

[mengfei25]

🐛 Describe the bug

python benchmarks/dynamo/timm_models.py --accuracy --amp --amp-dtype bfloat16 -d xpu -n10 --training --only eca_halonext26ts --backend=inductor

xpu  train eca_halonext26ts                   
E0104 12:46:08.845000 967154 site-packages/torch/_dynamo/utils.py:2353] RMSE (res-fp64): 0.01146, (ref-fp64): 0.00222 and shape=torch.Size([512]). res.dtype: torch.float32, multiplier: 3.000000, tol: 0.010000, use_larger_multiplier_for_smaller_tensor: 0
E0104 12:46:08.845000 967154 site-packages/torch/_dynamo/utils.py:2223] Accuracy failed for key name stages.3.0.post_attn.running_var
fail_accuracy

Versions

Envirnoments:
Device: PVC 1100
torch-xpu-ops: 18bcd9a
python: 3.10
TRITON_COMMIT_ID: e98b6fcb8df5b44eb0d0addb6767c573d37ba024
TORCH_COMMIT_ID: b9fbd65dfd5e703bacbc6c25258d1215108b4faf
TORCHBENCH_COMMIT_ID: 766a5e3a189384659fd35a68c3b17b88c761aaac
TORCHVISION_COMMIT_ID: d23a6e1664d20707c11781299611436e1f0c104f
TORCHAUDIO_COMMIT_ID: b6d4675c7aedc53ba04f3f55786aac1de32be6b4
DRIVER_VERSION: 1.23.10.49.231129.50 (803.61)
KERNEL_VERSION: 5.15.0-73-generic #80 SMP Mon May 15 15:18:26 UTC 2023
BUNDLE_VERSION: 2025.0.1.20241113 (DL-Essential 2025.0.1)
OS_PRETTY_NAME: Ubuntu 22.04.2 LTS
GCC_VERSION: 11

[mengfei25]

fbnetv3_b got same issue

xpu  train fbnetv3_b
E0104 13:41:14.249000 998978 site-packages/torch/_dynamo/utils.py:2353] RMSE (res-fp64): 0.30015, (ref-fp64): 0.05598 and shape=torch.Size([360]). res.dtype: torch.float32, multiplier: 3.000000, tol: 0.040000, use_larger_multiplier_for_smaller_tensor: 0
E0104 13:41:14.249000 998978 site-packages/torch/_dynamo/utils.py:2223] Accuracy failed for key name blocks.4.3.bn1.running_var
fail_accuracy