timm_models pnasnet5large train float16 eager_two_runs_differ

### 🐛 Describe the bug

There is `eager_two_runs_differ` for timm_models **pnasnet5large** train float16 on Windows with Intel XPU (Arc B580).

```
loading model: 0it [00:00, ?it/s]
loading model: 0it [00:15, ?it/s]
xpu train pnasnet5large
E0710 04:36:33.982000 9144 site-packages\torch\_dynamo\utils.py:3752] Accuracy failed: allclose not within tol=0
E0710 04:36:33.983000 9144 site-packages\torch\_dynamo\utils.py:3568] Accuracy failed for key name cell_0.comb_iter_0_left.bn_sep_1.bias
eager_two_runs_differ
```

### Versions

Click to expand traceback

PyTorch version: 2.14.0a0+git140e4d5
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A

OS: Microsoft Windows 11 Pro (10.0.26100 64-bit)
GCC version: Could not collect
Clang version: Could not collect
CMake version: version 3.31.6
Libc version: N/A

Python version: 3.12.13 | packaged by conda-forge | (main, Mar 5 2026, 16:36:12) [MSC v.1944 64 bit (AMD64)] (64-bit runtime)
Python platform: Windows-11-10.0.26100-SP0
Is CUDA available: False
CUDA runtime version: No CUDA
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA
Is XPU available: True
XPU used to build PyTorch: 20260000
Intel GPU driver version:
* 32.0.101.8826 (20260529000000.******+***)
Intel GPU models onboard:
* Intel(R) Arc(TM) B580 Graphics
Intel GPU models detected:
* [0] _XpuDeviceProperties(name='Intel(R) Arc(TM) B580 Graphics', platform_name='Intel(R) oneAPI Unified Runtime over Level-Zero V2', type='gpu', device_id=0xE20B, uuid=86800be2-0000-0000-0400-000000000000, driver_version='1.15.37858', total_memory=11875MB, local_mem_size=128KB, last_level_cache_size=18432KB, max_compute_units=160, memory_clock_rate=0MHz, memory_bus_width=64-bit, gpu_eu_count=160, gpu_subslice_count=20, max_work_group_size=1024, max_num_sub_groups=64, sub_group_sizes=[16 32], has_fp16=1, has_fp64=1, has_atomic64=1, is_integrated_gpu=0)
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: False
Caching allocator config: N/A

CPU:
Name: Intel(R) Core(TM) Ultra 5 245K
Manufacturer: GenuineIntel
Family: 773
Architecture: 9
ProcessorType: 3
DeviceID: CPU0
CurrentClockSpeed: 3071
MaxClockSpeed: 4200
L2CacheSize: 26624
L2CacheSpeed: None
Revision: None

Versions of relevant libraries:
[pip3] bert_pytorch==0.0.1a4
[pip3] functorch==1.14.0a0+b71aa0b
[pip3] intel-openmp==2026.1.0
[pip3] mkl-include==2026.0.0
[pip3] mkl-static==2026.0.0
[pip3] mypy==2.2.0
[pip3] mypy_extensions==1.1.0
[pip3] numpy==2.3.2
[pip3] onemkl-license==2026.0.0
[pip3] onnx==1.21.0
[pip3] onnx-ir==0.1.16
[pip3] onnxscript==0.6.2
[pip3] optree==0.13.0
[pip3] pytorch-labs-segment-anything-fast==0.2
[pip3] tbb==2023.0.0
[pip3] tbb-devel==2023.0.0
[pip3] tcmlib==1.5.0
[pip3] torch==2.14.0a0+git140e4d5
[pip3] torch_geometric==2.4.0
[pip3] torchao==0.17.0
[pip3] torchaudio==2.11.0a0+c0cbdb9
[pip3] torchbench==0.1
[pip3] torchmetrics==1.9.0
[pip3] torchmultimodal==0.1.0b0
[pip3] torchrec-nightly==2022.4.26
[pip3] torchvision==0.29.0a0+0bc41e6
[pip3] torchx-nightly==2026.7.9
[pip3] triton-xpu==3.7.2+git5fcc14d9
[conda] bert-pytorch 0.0.1a4 dev_0
[conda] functorch 1.14.0a0+b71aa0b pypi_0 pypi
[conda] intel-openmp 2026.1.0 pypi_0 pypi
[conda] mkl-include 2026.0.0 pypi_0 pypi
[conda] mkl-static 2026.0.0 pypi_0 pypi
[conda] numpy 2.3.2 pypi_0 pypi
[conda] onemkl-license 2026.0.0 pypi_0 pypi
[conda] optree 0.13.0 pypi_0 pypi
[conda] pytorch-labs-segment-anything-fast 0.2 pypi_0 pypi
[conda] tbb 2023.0.0 pypi_0 pypi
[conda] tbb-devel 2023.0.0 pypi_0 pypi
[conda] tcmlib 1.5.0 pypi_0 pypi
[conda] torch 2.14.0a0+git140e4d5 pypi_0 pypi
[conda] torch-geometric 2.4.0 pypi_0 pypi
[conda] torchao 0.17.0 pypi_0 pypi
[conda] torchaudio 2.11.0a0+c0cbdb9 pypi_0 pypi
[conda] torchbench 0.1 pypi_0 pypi
[conda] torchmetrics 1.9.0 pypi_0 pypi
[conda] torchmultimodal 0.1.0b0 pypi_0 pypi
[conda] torchrec-nightly 2022.4.26 pypi_0 pypi
[conda] torchvision 0.29.0a0+0bc41e6 pypi_0 pypi
[conda] torchx-nightly 2026.7.9 pypi_0 pypi
[conda] triton-xpu 3.7.2+git5fcc14d9 pypi_0 pypi

Start by reproducing eager_two_runs_differ for timm_models pnasnet5large in train float16 on Windows with Intel XPU, using the environment details in the report. Investigate the reported cell_0.comb_iter_0_left.bn_sep_1.bias mismatch and establish a reproducible fix or narrow the failure to the relevant XPU operation.

Written by the indexing model from the issue text.