support qwen3.5 input layout

[mayuyuace]

Qwen3 Next and Qwen 3.5 have different inputs layouts, modify the kernel to support qwen 3.5.
The UT also is updated.

vllm-xpu PR:
https://github.com/intel-innersource/applications.ai.gpu.vllm-xpu/pull/161

bash run-lm-eval-gsm-vllm-baseline.sh -m Qwen/Qwen3.5-9B,dtype=bfloat16 -b 20 -l 250 -f 5 -t 2

triton gdn:

sycl gdn:

[Copilot]

Pull request overview

Adds support for an alternate QKVZ/BA input layout (Qwen 3.5) to the XPU GDN attention path by threading a reorder_input flag through the Torch op binding, C++ interface, and SYCL kernels, and updating the unit test to validate both layouts.

Changes:

• Extend gdn_attention Torch op signature + C++ interface to accept reorder_input.
• Update causal-conv1d (native + XE2 chunk) kernels to interpret inputs in either legacy (Qwen Next) or reordered (Qwen 3.5) layout.
• Expand the GDN attention unit test to run with reorder_input=True/False and reorder the reference inputs accordingly.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
tests/gdn_attn/test_gdn_attn.py	Adds reorder_input parametrization and reference-side input reshaping to validate both layouts.
csrc/xpu/torch_bindings.cpp	Updates the Torch library schema for gdn_attention to include reorder_input.
csrc/xpu/ops.h	Extends the C++ op declaration with the reorder_input flag.
csrc/xpu/gdn_attn/gdn_attn_interface.cpp	Wires reorder_input through to causal-conv1d (native + XE2) launch paths.
csrc/xpu/gdn_attn/causal_conv1d.hpp	Adds compile-time specialization for reordered layout and dispatch based on reorder_input.
csrc/xpu/gdn_attn/xe_2/chunk_causal_conv1d_xe2.hpp	Adds reordered-layout specialization and runtime dispatch to the XE2 chunk kernels.

---

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

[jikunshang]

add below at the test case entry?

if num_actual_tokens == 8192: 
    pytest.skip("FIXME, skip because of random error")

[jikunshang]

Overall LGTM.