remove xpu_fused_moe weights handling

[mayuyuace]

align with https://github.com/intel-innersource/applications.ai.gpu.vllm-xpu/pull/145

[Copilot]

Pull request overview

This PR removes implicit weight-layout mutation/caching from xpu_fused_moe and updates tests to prepare weights in the expected layout externally, while also loosening activation-name matching for the SWIGLUOAI path.

Changes:

• Removed in-function transpose/caching for non-int4/non-mxfp4 weights and adjusted inter_size derivation based on expected layout.
• Updated fused MoE tests to transpose w13/w2 before calling xpu_fused_moe.
• Expanded SWIGLUOAI activation matching logic.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 4 comments.

File	Description
vllm_xpu_kernels/fused_moe_interface.py	Stops mutating weights for layout, changes inter_size inference, and adjusts SWIGLUOAI activation detection.
tests/fused_moe/test_fused_moe.py	Pre-transposes weights in tests to match the new expected layout.

---

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

[Copilot]

xpu_fused_moe now derives inter_size from w13.shape[-1] for non-int4/non-mxfp4, which implicitly requires w13/w2 to already be in the [E, K, N] layout (transposed vs the shapes documented above and previously handled in this function). To avoid silent wrong results/crashes when callers pass the old layout, please (1) update the parameter shape docs to match the new expectation and (2) add explicit shape/layout validation (e.g., check the dimension that should equal hidden_size, and that w13.shape[-1] is divisible by 2) before calling the GEMM op.

[Copilot]

The condition ("SWIGLUOAI" in str(activation)) is overly broad: it will also match unrelated values like "NOT_SWIGLUOAI" and route them to the swiglu kernel. Safer is to normalize once (e.g., str(activation).lower()) and compare against an explicit allowlist of supported names (including any enum/string variants you intend to support).

[Copilot]

Avoid assigning through .data here. It bypasses autograd and can lead to hard-to-debug issues if this test is adapted/reused; since these are local variables, prefer rebinding w13/w2 to the transposed contiguous tensors (or otherwise copying safely) instead of mutating .data.

Suggested change

	w13.data = w13.transpose(-1, -2).contiguous()
	w2.data = w2.transpose(-1, -2).contiguous()
	w13 = w13.transpose(-1, -2).contiguous()
	w2 = w2.transpose(-1, -2).contiguous()

[Copilot]

Avoid assigning through .data here. It bypasses autograd and can lead to hard-to-debug issues if this test is adapted/reused; since these are local variables, prefer rebinding w13/w2 to the transposed contiguous tensors (or otherwise copying safely) instead of mutating .data.

Suggested change

	w13.data = w13.transpose(-1, -2).contiguous()
	w2.data = w2.transpose(-1, -2).contiguous()
	w13 = w13.transpose(-1, -2).contiguous()
	w2 = w2.transpose(-1, -2).contiguous()

[Liangliang-Ma]

we probably should add some contiguous/stride check here to ensure correctness.

[jikunshang]

LGTM. while I want to on hold this at least this week, to not block weekly test.

[mayuyuace]

we probably should add some contiguous/stride check here to ensure correctness.

Sure. I can add check later.