[release/2.14] torchbench_training_float16 hf_T5_large eager perf regression

### 🐛 Describe the bug

Issue found during 2.14 RC test.

There is performance regression on Windows with Intel XPU (Arc B580). Batch size used: 2

Testcase | Model | Inductor status | Inductor diff | Inductor ref latency [ms] | Inductor cur latency [ms] | Eager status | Eager diff | Eager ref latency [ms] | Eager cur latency [ms]
-- | -- | -- | -- | -- | -- | -- | -- | -- | --
torchbench_training_float16 | hf_T5_large | PASS | 7.00% | 858.54 | 918.49 | FAIL | 10.70% | 1722.07 | 1907.13

### Versions

Click to expand versions
PyTorch version: 2.14.0+xpu
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A

OS: Microsoft Windows 11 Pro (10.0.26200 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, Aug 11 2026, 10:18:27) [MSC v.1944 64 bit (AMD64)] (64-bit runtime)
Python platform: Windows-11-10.0.26200-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: 20260100
Intel GPU driver version:
* 32.0.101.8864 (20260717000000.******+***)
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=86800be2-0000-0000-0300-000000000000, driver_version='1.15.38308', 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: 13th Gen Intel(R) Core(TM) i5-13400
Manufacturer: GenuineIntel
Family: 205
Architecture: 9
ProcessorType: 3
DeviceID: CPU0
CurrentClockSpeed: 2500
MaxClockSpeed: 2500
L2CacheSize: 9728
L2CacheSpeed: None
Revision: None

Versions of relevant libraries:
[pip3] dpcpp-cpp-rt==2026.1.0
[pip3] intel-cmplr-lib-rt==2026.1.0
[pip3] intel-cmplr-lib-ur==2026.1.0
[pip3] intel-cmplr-lic-rt==2026.1.0
[pip3] intel-opencl-rt==2026.1.0
[pip3] intel-openmp==2026.1.0
[pip3] intel-pti==1.0.1
[pip3] intel-sycl-rt==2026.1.0
[pip3] mkl==2026.1.0
[pip3] mkl-include==2024.2.0
[pip3] mkl-static==2024.2.0
[pip3] mypy_extensions==1.1.0
[pip3] numpy==1.26.4
[pip3] onemkl-license==2026.1.0
[pip3] onemkl-sycl-blas==2026.1.0
[pip3] onemkl-sycl-dft==2026.1.0
[pip3] onemkl-sycl-lapack==2026.1.0
[pip3] onemkl-sycl-rng==2026.1.0
[pip3] onemkl-sycl-sparse==2026.1.0
[pip3] onnx==1.21.0
[pip3] onnx-ir==0.1.16
[pip3] onnxscript==0.6.2
[pip3] optree==0.13.0
[pip3] tbb==2023.1.0
[pip3] tbb-devel==2021.13.1
[pip3] tcmlib==1.5.0
[pip3] torch==2.14.0+xpu
[pip3] torchaudio==2.11.0+xpu
[pip3] torchvision==0.29.0+xpu
[pip3] triton-xpu==3.8.0
[pip3] umf==1.1.0
[conda] dpcpp-cpp-rt 2026.1.0 pypi_0 pypi
[conda] intel-cmplr-lib-rt 2026.1.0 pypi_0 pypi
[conda] intel-cmplr-lib-ur 2026.1.0 pypi_0 pypi
[conda] intel-cmplr-lic-rt 2026.1.0 pypi_0 pypi
[conda] intel-opencl-rt 2026.1.0 pypi_0 pypi
[conda] intel-openmp 2026.1.0 pypi_0 pypi
[conda] intel-pti 1.0.1 pypi_0 pypi
[conda] intel-sycl-rt 2026.1.0 pypi_0 pypi
[conda] mkl 2026.1.0 pypi_0 pypi
[conda] mkl-include 2024.2.0 pypi_0 pypi
[conda] mkl-static 2024.2.0 pypi_0 pypi
[conda] numpy 1.26.4 pypi_0 pypi
[conda] onemkl-license 2026.1.0 pypi_0 pypi
[conda] onemkl-sycl-blas 2026.1.0 pypi_0 pypi
[conda] onemkl-sycl-dft 2026.1.0 pypi_0 pypi
[conda] onemkl-sycl-lapack 2026.1.0 pypi_0 pypi
[conda] onemkl-sycl-rng 2026.1.0 pypi_0 pypi
[conda] onemkl-sycl-sparse 2026.1.0 pypi_0 pypi
[conda] optree 0.13.0 pypi_0 pypi
[conda] tbb 2023.1.0 pypi_0 pypi
[conda] tbb-devel 2021.13.1 pypi_0 pypi
[conda] tcmlib 1.5.0 pypi_0 pypi
[conda] torch 2.14.0+xpu pypi_0 pypi
[conda] torchaudio 2.11.0+xpu pypi_0 pypi
[conda] torchvision 0.29.0+xpu pypi_0 pypi
[conda] triton-xpu 3.8.0 pypi_0 pypi
[conda] umf 1.1.0 pypi_0 pypi