[XPU] Fix causal_conv1d_update_xpu signature for GDN speculative decode

[dayanandav]

Motivation

Serving a hybrid GDN model (e.g. Qwen3.5) on XPU with speculative decoding crashes during warmup:

TypeError: causal_conv1d_update_xpu() got an unexpected keyword argument 'intermediate_conv_window'

The GDN target-verify path in sglang (python/sglang/srt/layers/attention/linear/gdn_backend.py) calls
causal_conv1d_update with five speculative-decoding tree arguments:

• intermediate_conv_window
• intermediate_state_indices
• retrieve_next_token
• retrieve_next_sibling
• retrieve_parent_token

causal_conv1d_update_xpu only accepted arguments up to pad_slot_id, so the call raises TypeError
before reaching the kernel. Only the TARGET_VERIFY path is affected; non-speculative decode is fine.

Fixes sgl-project/sglang#34720

Modifications

• python/sgl_kernel/mamba.py: causal_conv1d_update_xpu now accepts and ignores unknown keyword
  arguments via **kwargs, matching what causal_conv1d_fn_xpu in the same file already does. This
  keeps the XPU wrapper tolerant of upstream signature changes instead of hard-failing at warmup.
• tests/test_causal_conv1d.py: new test_causal_conv1d_update_ignores_spec_decode_kwargs regression
  test that reproduces the failing call site (all five kwargs) and asserts both the output and the
  in-place conv_state update still match the reference.

Note for reviewers

The tree-topology arguments are accepted but not honored by the SYCL kernel — no intermediate conv
window is checkpointed per draft token. This unblocks the crash and is correct for linear-chain draft
topologies, but tree-based drafting (topk > 1) would need the checkpointing implemented in
src/sycl/causal_conv1d.cpp to be numerically correct. Happy to follow up with either the kernel-side
support or an explicit guard/warning for the tree case if maintainers prefer failing loudly over
silently ignoring.

Accuracy Test

tests/test_causal_conv1d.py on Intel XPU — full file, including the new test:

761 passed in 16.33s

Benchmark & Profiling

Not applicable — signature-only change, no kernel or performance impact.

Checklist

• Format your code according to the Code Formatting with Pre-Commit.
• Add unit tests as outlined in the Running Unit Tests.
• Update documentation / docstrings / example tutorials as needed, per Writing Documentation.
• Provide throughput / latency benchmark results and accuracy evaluation results as needed, following Benchmark and Profiling and Accuracy Results.
• For reviewers: If you haven't made any contributions to this PR and are only assisting in merging the main branch, please remove yourself as a co-author when merging the PR.
• Please feel free to join our Slack channel at https://slack.sglang.ai to discuss your PR.

🤖 Generated with Claude Code

[dayanandav]

I recommend this solution on kernel side, @jiayisunx @Xia-Weiwen requested for your review.

[jiayisunx]

Currently, we don't need to optimize the SYCL conv1d kernels anymore. As Weiwen mentioned, there will be two backends for linear attention on XPU: Triton and intel_xpu. For the Triton backend, we always use Triton kernels, so I removed the SYCL conv1d kernels from SGLang in [GDN] remove XPU path of causal_conv1d_fn and causal_conv1d_update by jiayisunx · Pull Request #3577…. Therefore, the SYCL conv1d kernels are no longer used by SGLang.