port cuda specific case with hook

[daisyden]

Fix #2081

[Copilot]

Pull Request Overview

Ports a CUDA-specific large-index warp softmax test to XPU, adding large tensor coverage for 64-bit indexing correctness.

• Adds import for largeTensorTest and introduces a new large-shape log_softmax gradient correctness test.
• Registers the test dynamically on TestNNDeviceType with multiple large memory size decorators.

Comments suppressed due to low confidence (1)

test/xpu/test_nn_xpu.py:1

• [nitpick] The decorator iterates over dtype (torch.float, torch.half) but the input tensor is always created as torch.float16, causing the float case to test a half input with a float output cast instead of a true float input path. For clearer coverage of dtype-dependent kernel behavior, construct x with the selected dtype (e.g., x = torch.randn(shape, device=device, dtype=dtype, requires_grad=True)) unless there is a documented reason to force half.

# Owner(s): ["module: intel"]

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[Copilot]

The test receives a device argument (from the instantiated device-type tests and the second @largeTensorTest decorator specifying 'cpu'), but the tensor is hard-coded with device='xpu'. This prevents the CPU-annotated variant from actually exercising a CPU path and can cause mismatched expectations. Use the passed-in device variable for allocation: x = torch.randn(shape, device=device, dtype=torch.float16, requires_grad=True).

Suggested change

	x = torch.randn(shape, device="xpu", dtype=torch.float16, requires_grad=True)
	x = torch.randn(shape, device=device, dtype=torch.float16, requires_grad=True)