[Intel][XPU] Device-agnostic fixes for XPU, scripted chunked-prefill, and DWDP

[dayanandav]

Fixes for #31995 #24922 #36478 and Device-agnostic device timer fix for Intel XPU backend

Verification

• pre-commit run --all-files clean (28 hooks) with the 36 changed files in the tree.
• The DWDP argparse cases live in test/registered/unit/server_args/test_server_args.py.
• Pending on my XPU box: Qwen3.5 before/after perf numbers, and a GDN decode run
  against the pre-PR fork.

Notes on three deliberate choices

DWDP platform gate (arg_groups/parallel_hook.py). The #31996 reference in an
earlier commit message is wrong: #31996 is closed, not merged, and
layers/moe/dwdp/page_pool.py:9 still carries a bare
from cuda.bindings import driver as cuda. The accurate justification is that the
import chain described in #31995 no longer exists on main --
model_runner.py:1331-1335 imports DwdpManager lazily inside maybe_init_dwdp
behind if get_parallel().dwdp_size <= 1: return, and
fused_moe_triton/layer.py:78 reaches the manager through
get_global_dwdp_manager, which pulls in nothing CUDA-specific. So the only
remaining route to ModuleNotFoundError: No module named 'cuda' is
--dwdp-size >= 2 on a platform without cuda-python, which is exactly what
handle_dwdp now rejects with a message. This PR changes nothing under
layers/moe/dwdp/.

ScriptedReqHandle.lock_refs contract (scripted_runtime/req_handle.py:56-59).
A root last_node now folds to 0 instead of returning the tree's permanent
sentinel. inc_lock_ref / dec_lock_ref both loop while node != self.root_node
(radix_cache.py:668,683), so the root's lock_ref is fixed at 1 for the process
lifetime and no req's lock can move it. The pre-change value on a root last_node
was therefore a constant unrelated to the req, and this is the fix for that, not a
weakening.

Deleted XPU GDN fork
(srt/hardware_backend/xpu/kernels/fla/fused_sigmoid_gating_recurrent.py). That
fork imported the shared fused_sigmoid_gating_delta_rule_update_kernel and
launched it by keyword without stride_h0_source, which has no default in the
kernel signature, so every XPU GDN decode on main raises at launch. Removing it
cannot regress a working path. XPU now takes the shared wrapper, whose only
XPU-specific behavior is the BV cap at 16.

🤖 Generated with Claude Code

---

CI States

Latest PR Test (Base): ❌ Run #37654107331
Latest PR Test (Extra): ❌ Run #37654106872
Latest PR Test (AMD ROCm 10): ❌ Run #37654107341

[dayanandav]

can you also provide perf numbers of qwen3.5 before and after this PR. need to verify no perf regression.

CUDA / ROCm: unchanged by construction. The only kernel change is XPU-gated:

BV, num_warps = _select_recurrent_launch_config(N, H, HV, K, V, is_kda)
if q.device.type == "xpu":
    BV = min(BV, 16)

_select_recurrent_launch_config (including the gfx950 tuning) is untouched, so BV / num_warps / grid are identical to main on CUDA and ROCm.

XPU: there is no "before" to measure. On main, gdn_triton.py dispatches XPU to a wrapper that never passed stride_h0_source, so every GDN decode raised:

TypeError: dynamic_func() missing 1 required positional argument: 'stride_h0_source'

After this PR — Qwen3.5-4B linear-attention shapes (H=16, HV=32, K=128, V=128), T=1 decode, bf16, 1x Intel XPU:

batch	1	8	32	64	128
us/call	55.1	55.8	375.0	718.8	1399.5

The launch config this PR uses on XPU is identical to the deleted wrapper's (BV=16, BK=128, num_warps=1, num_stages=3, grid=(NK, NV, N*HV)) — it de-duplicates the wrapper rather than retuning it, so there is no XPU tiling change to regress either.

[mingfeima]

fused_sigmoid_gating_recurrent.py: the request was to keep the XPU fork (sglang.srt.hardware_backend.xpu.kernels.fla.fused_sigmoid_gating_recurrent) and drop the shared-wrapper change.

The current head does the opposite — deletes the fork and keeps if q.device.type == "xpu": BV = min(BV, 16) in the shared file.

cc @Xia-Weiwen

[dayanandav]

Done — fork restored, shared-wrapper clamp removed, elif is_xpu() dispatch re-added in gdn_triton.py.

One line beyond a plain restore: the fork's kernel launch now passes stride_h0_source. The shared kernel made it a required positional, so main's fork raises TypeError on every XPU GDN decode without it.

[dayanandav]

Request for review