[mehersai07]

Motivation

#40190 wired the fused QK-norm + RoPE kernel (sgl_kernel.fused_inplace_qknorm_rope) into Qwen3-MoE on XPU behind the existing opt-in --enable-fused-qk-norm-rope. Its A/B on Qwen3-30B-A3B (TP=4, 4x Arc Pro B60) showed +4.7-7.0% output throughput, -60% q/k-norm + RoPE stage device time, and no accuracy change (GSM8K 0.9414 vs 0.9460).
The flag is opt-in on CUDA for a reason that does not apply to XPU: on CUDA the unfused path already fuses RoPE with the KV-cache write (fused_set_kv_buffer), and the fused QK-norm path gives that up. enable_fused_set_kv_buffer is CUDA/HIP-only, so the XPU unfused path has nothing to lose. This PR makes the fused path the XPU default; everything else is unchanged.

Modifications

• arg_groups/fields/exec_.py: enable_fused_qk_norm_rope becomes a tri-state (Optional[bool] = None, BooleanOptionalAction, resolvable=True), the same shape as enable_dsa_fused_indexer. --enable-fused-qk-norm-rope keeps working; --no-enable-fused-qk-norm-rope is the new opt-out.
• arg_groups/platform_hook.py: handle_xpu_backends declares enable_fused_qk_norm_rope=True when the device is XPU and the user did not choose, next to the existing XPU sampling_backend default.
• Docs: the flag's rows in server_arguments.mdx and the Ascend support table.

On every other platform the field stays None. Its two readers (qwen3_moe.py, mellum.py) use it only in and chains and truthiness checks, so None behaves exactly like the old False. No model code changes.

One visible difference: off XPU, /get_server_info now reports enable_fused_qk_norm_rope: null instead of false, matching how enable_dsa_fused_indexer reports.

Accuracy Tests

No new numerics: the kernel and call site are the ones #40190 measured (GSM8K above). The nightly test/registered/xpu/llm_models/test_xpu_qwen3_30b_a3b.py (TP=4 GSM8K, threshold 0.90) now runs the fused path by default.

Resolution, checked through prepare_server_args + resolve_once():

device	flag	resolved
xpu	none	True
xpu	--enable-fused-qk-norm-rope	True
xpu	--no-enable-fused-qk-norm-rope	False
non-XPU	none	None (off)
non-XPU	--enable-fused-qk-norm-rope / --no-enable-fused-qk-norm-rope	True / False

Speed Tests and Profiling

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review and Merge Process

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

---

CI States

Latest PR Test (Base): 🚫 Run #38838489630
Latest PR Test (Extra): ❌ Run #38838489477
Latest PR Test (AMD ROCm 10): ❌ Run #38838489627

cc: @rahulvijayaraghavan @rbabukv

[polisettyvarma]

/tag-run-ci-label

[mingfeima]

@gaopengff could you please help review this one?

[gaopengff]

LGTM.

[mingfeima]

/rerun-failed-ci