[XPU] Gate FlashAttention-in-graph on oneAPI 2026.0+ runtime support

Purpose

FlashAttention SYCL kernels use sycl_ext_oneapi_work_group_scratch_memory,
which SYCL Graph cannot capture before oneAPI 2026.0
(intel/torch-xpu-ops#3142).
On older runtimes, enabling XPU graphs with FLASH_ATTN and a full cudagraph
mode crashes at warmup with:

work_group_scratch_memory ... not yet available for use with the SYCL Graph extension

and the practical guidance has been "FLASH_ATTN supports PIECEWISE only".

This PR makes full-graph FlashAttention capture available where the runtime
supports it, and safe everywhere else:

• supports_xpu_fa_in_graph(): torch.version.xpu >= 20260000.
• VLLM_XPU_GRAPH_FORCE_PIECEWISE (default on): clamps full graph modes
  to PIECEWISE, preserving today's behavior.
• Opt-out (FORCE_PIECEWISE=0) on a capable runtime keeps the requested
  FULL / FULL_AND_PIECEWISE mode and logs FlashAttention-in-graph enabled.
• Opt-out on an incapable runtime falls back to PIECEWISE with a warning
  instead of crashing at warmup (fail closed).

Required dependencies (do not merge this alone for production FA-in-graph)

This PR is gating only. The validated Arc Pro B70 / Ornith FA-in-graph
stack also requires the following companion PRs (please land / pin together
or treat as blockers for claiming full graphs support):

Context (not merge blockers for this gate PR):
intel/torch-xpu-ops#3142,
vllm#48946 (PVC MXFP4 MoE
graph notes; Arc datapoint below).

is_padding WA cycle is already handled by kernels
#481 + Intel WAs
in vLLM — do not re-open here.

Why this is not duplicating an existing PR

Duplicate-work checks (gh pr list / issue search on vllm-project/vllm and
vllm-xpu-kernels) found no open PR that adds supports_xpu_fa_in_graph,
the fail-closed FULL-mode clamp, or VLLM_XPU_GRAPH_FORCE_PIECEWISE. The
dependencies listed above are already open; this PR intentionally does not
re-bundle them.

Test Plan

• pytest tests/utils_/test_torch_utils.py -k supports_xpu_fa_in_graph
  (6 cases; no GPU required; pass in oneAPI 2026 container).
• Dense canary: Qwen2.5-0.5B-Instruct, Arc Pro B70, torch 2.13.0+xpu
  (torch.version.xpu=20260000), oneAPI 2026.0 — eager vs FA-in-graph FULL.
• Hybrid MoE canary: Ornith-1.0-35B-MXFP4 (MXFP4 MoE + GDN/Mamba), same
  device/runtime — eager vs PIECEWISE vs FA-in-graph FULL
  (auto → FULL_AND_PIECEWISE for UNIFORM_BATCH backends). Correctness
  smokes S1–S7 vs eager.
• Serving sweep (concurrency × length): graphs ON vs graphs OFF on the
  feature image (same serve profile).

Test Result

Unit tests

pytest tests/utils_/test_torch_utils.py -k supports_xpu_fa_in_graph
# 6 passed

No work_group_scratch_memory / SYCL Graph error in any serve arm. On
incapable runtimes the fail-closed path clamps to PIECEWISE as intended.

Correctness smokes — dense (Qwen2.5-0.5B-Instruct)

Correctness smokes — Ornith-1.0-35B-MXFP4 (S1–S7)

Arms: A eager, B PIECEWISE, C FA-in-graph FULL → FULL_AND_PIECEWISE.

Output agreement vs eager (byte-identical / similarity):

Single-stream perf (canary, greedy)

Dense — 8×128-token streamed completions after warmup:

Ornith hybrid MoE — single stream, 8×128-token greedy:

Serving sweep — concurrency × length (graphs ON vs OFF)

Hardware: Intel Arc Pro B70. Model: Ornith-1.0-35B-MXFP4. Profile:
max-model-len=131072, max-num-seqs=2, gpu-memory-utilization=0.9,
kv-cache-dtype=fp8, bf16. Grid: C ∈ {1,2,4,8} ×
(in,out) ∈ {(128,128),(2048,256),(8192,256)} (table shows C=1 and C=8).

Graphs ON (VLLM_XPU_ENABLE_XPU_GRAPH=1, FORCE_PIECEWISE=0,
FLASH_ATTN, -cc.cudagraph_mode=FULL):

Graphs OFF (same image; VLLM_XPU_ENABLE_XPU_GRAPH=0, --enforce-eager):

Deltas (graphs ON vs OFF) — same image/runtime:

Telemetry over sweep wall (xe hwmon energy counters):

Scope / non-goals

• Default behavior unchanged (FORCE_PIECEWISE=1). FA-in-graph is opt-in.
• Does not bump Dockerfile / torch pins — required companion: [XPU] upgrade to torch 2.13 #48677.
• Does not change kernels wheels or MoE oracle — required companions:
  [XPU] Prefer MXFP8 MoE XPU backend; forward softcap/ALIBI #49813 and vllm-xpu-kernels Flash Attention V2 #485–fix max seq len #489.
• Validated on Intel Arc Pro B70 + oneAPI 2026.0; PVC / other XPUs not
  re-run here.

AI assistance

AI assistance was used to develop and validate this change. Every changed
line has been reviewed by the human submitter.