inf is returned by nn.TransformerEncoderLayer

[Stonepia]

inf is returned by nn.TransformerEncoderLayer

Fixes #2015

Root Cause: The test_transformerencoderlayer test added XPU to fast_path_device check (commit 792afdf), causing the test to expect NaN output on XPU when a fully-masked row is passed. However, XPU's TransformerEncoderLayer does not implement the fast path (unlike CUDA/CPU), so the attention implementation produces inf/NaN due to float16 overflow or softmax over -inf values rather than the expected NaN pattern. For float16, large input values ([20., 30., 40., 50.]) cause overflow in the softmax/attention computation, producing NaN instead of the expected reference values.

Failed Tests:

• test/test_nn.py::TestNNDeviceTypeXPU::test_transformerencoderlayer_gelu_xpu_float16
• test/test_nn.py::TestNNDeviceTypeXPU::test_transformerencoderlayer_xpu_float16
• test/test_nn.py::TestNNDeviceTypeXPU::test_transformerencoderlayer_xpu_float32
• test/test_nn.py::TestNNDeviceTypeXPU::test_transformerencoderlayer_xpu_float64

---

Diff stat:

test/xpu/test_nn_xpu.py | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

[chuanqi129]

@copilot The Python lint (flake8/ruff) check failed. Please fix the lint errors in this PR.

Lint job log: https://github.com/intel/torch-xpu-ops/actions/runs/25865859919

Instructions:

1. Review the lint failure log via the link above
2. Fix all lint errors in the relevant source files
3. Ensure both Python lint and Clang format checks pass

[Copilot]

@copilot The Python lint (flake8/ruff) check failed. Please fix the lint errors in this PR.

Lint job log: https://github.com/intel/torch-xpu-ops/actions/runs/25865859919...

Fixed in commit a5d61d7. The PYFMT formatter required collapsing the multi-line and (... or ...) condition into a single line: and ("cuda" in str(device) or "cpu" in str(device)).