[#2188] Align test_workflow_ops_xpu tests with pytorch.

[jmamzax]

Part of #2188 issue. Tests test_learnable_forward_per_channel_cuda_xpu, test_learnable_backward_per_channel_cuda_xpu were updated to match pytorch test cases.

[Copilot]

Pull request overview

Updates the XPU quantization workflow tests to match upstream PyTorch’s learnable per-channel fake-quant test cases, addressing failures tracked in #2188.

Changes:

• Removed Hypothesis-driven input generation for the two learnable per-channel CUDA/XPU tests and replaced it with fixed shapes/axes.
• Added dtype coverage for the learnable per-channel forward/backward tests (float32 and bfloat16).
• Dropped the unused to_tensor import after refactoring the backward test setup.

Comments suppressed due to low confidence (2)

test/xpu/quantization/core/test_workflow_ops_xpu.py:104

• shape = (2, 1, 2, 10) with axis = 1 makes channel_size = X_base.size(axis) equal to 1, so this “per-channel” test only exercises the single-channel case and won’t catch channel-dependent bugs. Consider using a shape/axis combination where the selected dimension is > 1 (while still matching the intended PyTorch reference).

    shape = (2, 1, 2, 10)
    axis = 1

    for dtype in [torch.float32, torch.bfloat16]:
        X_base = torch.randn(shape, device="xpu").to(dtype)
        channel_size = X_base.size(axis)

test/xpu/quantization/core/test_workflow_ops_xpu.py:106

• Inside the dtype loop, torch.randn(...).to(dtype) (and the subsequent .to(dtype) conversions) introduces extra allocations/copies on XPU. Prefer creating tensors with the target dtype directly (e.g., pass dtype= to the factory functions) to keep the test lighter and reduce overhead.

    for dtype in [torch.float32, torch.bfloat16]:
        X_base = torch.randn(shape, device="xpu").to(dtype)
        channel_size = X_base.size(axis)
        scale_base = (
            torch.normal(mean=0, std=1, size=(channel_size,)).clamp(1e-4, 100).to(dtype)

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated 1 comment.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

In _test_learnable_forward_per_channel_cuda, scale_base and zero_point_base are created on the CPU and only cast to dtype. Since X_base is on XPU, this introduces host→device transfers (or potential device-mismatch issues if the downstream helper doesn’t move them). Consider creating these tensors directly on device='xpu' (and with the target dtype at creation) to keep all inputs on the same device and avoid extra copies.

Suggested change

	scale_base = (
	torch.normal(mean=0, std=1, size=(channel_size,)).clamp(1e-4, 100).to(dtype)
	)
	zero_point_base = torch.normal(mean=0, std=128, size=(channel_size,)).to(dtype)
	scale_base = torch.normal(
	mean=0,
	std=1,
	size=(channel_size,),
	device="xpu",
	dtype=dtype,
	).clamp(1e-4, 100)
	zero_point_base = torch.normal(
	mean=0,
	std=128,
	size=(channel_size,),
	device="xpu",
	dtype=dtype,
	)