[Store] Add Intel XPU host-staging path to the Store client, Python e2e and vLLM-XPU smoke tests - #4270

Description

This is PR3 of the Intel XPU enablement series (RFC #4031; PR1 #4030 added
the TENT XPU platform, PR2 #4157 added the XPU transport, routing and
host-staging in the Transfer Engine). It brings the Mooncake Store client up on
Intel XPU and adds end-to-end acceptance tests against an unmodified vLLM-XPU.

Store

• mooncake-store/src/device/xpu_accelerator_device.cpp (new, built only under
  USE_XPU): a SYCL-backed AcceleratorDevice (vendor kIntel). It
  classifies any USM pointer against the platform default context (the one
  PyTorch allocates from), so foreign device memory is recognised; copies run
  on queues bound to that context; pinned host memory comes from
  sycl::malloc_host. Linked in through EnsureXpuAcceleratorDeviceLinked().
• accelerator_device.h / accelerator_registry.cpp: kIntel vendor and the
  registration hook.
• mooncake-store/src/CMakeLists.txt: add the source and a USE_XPU branch in
  the D2H-staging block.

icpx compile fixes (behaviour-neutral for GCC)

• vendored fmt 8.0.1 core.h: two upstream-equivalent fixes so the
  consteval FMT_STRING checks compile under clang-based compilers.
• client_service.cpp: define Client::BatchUpsert after PutOperation is
  complete (std::vector<PutOperation> was instantiated on an incomplete type).

Engine selection is unchanged (per review on #4157): Intel XPU is only
supported by the TENT engine, so XPU deployments must set MC_USE_TENT=1. The
smoke scripts default it via os.environ.setdefault, the Python XPU test skips
unless it is set, and the MC_USE_TENT row in the store deployment guide notes
the requirement.

Tests

• pybind_client_test: XpuDeviceBufferPutGetRoundTrips (put_from /
  get_into on sycl::malloc_device buffers, byte-equal), compiled with
  MOONCAKE_TEST_XPU_H2D under USE_XPU.
• mooncake-wheel/tests/test_put_get_tensor.py: test_xpu_local_copy_paths
  (put_tensor / upsert_tensor from device="xpu", get_tensor_into_cuda
  into a registered XPU tensor, 4 MiB put_from / get_into round trip). The
  class now closes the store in tearDownClass: leaving it to the process-exit
  backstop tears the client down from a C atexit handler after the
  interpreter and the SYCL runtime have finalized, which on XPU intermittently
  aborted in free() while joining the client's worker threads.
• scripts/xpu/vllm_store_smoke.py: one vLLM instance on an Intel XPU with
  vLLM's V1 MooncakeStoreConnector (kv_role=kv_both) and local prefix
  caching disabled. Request 1 prefills and the connector saves the KV blocks
  into Mooncake Store (XPU KV cache → host staging → segment); request 2 is
  identical, cannot hit vLLM's own prefix cache, so the connector loads the
  blocks back into the XPU KV cache and decode resumes from them. PASS requires
  identical non-empty greedy tokens and connector metrics showing
  save_put / load_get keys with zero failures.
• scripts/xpu/vllm_store_pd_smoke.py: a vLLM kv_producer on one XPU
  (embedded Mooncake segment + master) and a kv_consumer on another
  (standalone-store, no segment), so every decode-side load is a real
  cross-device pull from the prefill segment. --role local spawns both on one
  node with ONEAPI_DEVICE_SELECTOR=level_zero:{0,1}; --role prefill /
  --role decode run the halves on separate nodes.
• Both scripts carry two in-process shims for vLLM ≤ 0.29 code that is
  CUDA-specific rather than Mooncake-specific, to be dropped once fixed
  upstream: the connector's wait_for_save uses torch.cuda.Event (a raising
  dummy on XPU torch → aliased to torch.xpu.Event), and
  ChunkedTokenDatabase.prepare_values computes KV addresses in np.int64,
  which overflows on Level Zero USM pointers (≥ 2^63) → uint64 math (the
  Mooncake bindings already take uintptr_t). Because of the shims the engine
  core runs in-process (VLLM_ENABLE_V1_MULTIPROCESSING=0).

Module

• Transfer Engine (mooncake-transfer-engine)
• Mooncake Store (mooncake-store)
• Mooncake Conductor (mooncake-conductor)
• Reshard (mooncake-reshard)
• Mooncake EP (mooncake-ep)
• Mooncake PG (mooncake-pg)
• Integration (mooncake-integration)
• P2P Store (mooncake-p2p-store)
• Python Wheel (mooncake-wheel)
• Common (mooncake-common)
• Mooncake RL (mooncake-rl)
• CI/CD
• Docs
• Other

Type of Change

• Bug fix
• New feature
• Refactor
• Breaking change
• Documentation update
• Performance improvement
• Other

How Has This Been Tested?

Test commands:

# XPU build (icpx, Intel Arc Pro B60)
. /opt/intel/oneapi/setvars.sh
cmake -G Ninja -S . -B build-xpu -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx \
  -DUSE_TENT=ON -DUSE_XPU=ON -DWITH_STORE=ON -DBUILD_UNIT_TESTS=ON \
  -DWITH_STORE_RUST=OFF -DWITH_P2P_STORE=OFF
cmake --build build-xpu
ctest --test-dir build-xpu -R 'pybind_client_test|runtime_accelerator_test' --output-on-failure

# Python e2e (against a live mooncake_master)
MC_USE_TENT=1 python -m unittest mooncake-wheel/tests/test_put_get_tensor.py -k xpu

# vLLM-XPU acceptance
python scripts/xpu/vllm_store_smoke.py --master-bin build-xpu/mooncake-store/src/mooncake_master
python scripts/xpu/vllm_store_pd_smoke.py --role local --master-bin build-xpu/mooncake-store/src/mooncake_master

# GCC CPU-only (USE_XPU=OFF, WITH_STORE=ON)
cmake -G Ninja -S . -B build -DUSE_TENT=ON -DUSE_XPU=OFF -DWITH_STORE=ON -DBUILD_UNIT_TESTS=ON
cmake --build build --target pybind_client_test runtime_accelerator_test
ctest --test-dir build -R 'pybind_client_test|runtime_accelerator_test'

Test results:

• Unit tests pass
  • XPU (icpx, B60): pybind_client_test incl. XpuDeviceBufferPutGetRoundTrips
    and the basic put/get cases; test_put_get_tensor.py incl.
    test_xpu_local_copy_paths against a live mooncake_master.
  • GCC (USE_XPU=OFF): store builds, runtime_accelerator_test 9/9,
    pybind_client_test pass — no behaviour change for non-XPU builds.
• Integration tests pass
  • vllm_store_smoke.py with vllm 0.29.0+xpu (torch 2.13.0+xpu),
    Qwen2.5-0.5B-Instruct: 525 prompt tokens, 8 keys / 6 MiB saved and loaded,
    0 failures, identical outputs.
  • vllm_store_pd_smoke.py on a single node (2 × B60) and across two nodes:
    8/8 KV blocks saved and loaded, decode tokens identical to the prefill
    reference.
• Manual testing done (above)

Checklist

• I have performed a self-review of my own code
• I have formatted my code using ./scripts/code_format.sh
• I have run pre-commit on the files changed in this PR and all hooks pass
• I have updated the documentation (if applicable)
• I have added tests to prove my changes are effective
• For changes >500 LOC: I have filed an RFC issue ([RFC]: Add Intel XPU support to Transfer Engine & Mooncake Store #4031)

AI Assistance Disclosure

• No AI tools were used
• AI tools were used (specify below)

GitHub Copilot (Claude) assisted with drafting the SYCL accelerator device,
the test scaffolding, the vLLM smoke scripts and this description. All changes
were reviewed, built and run on real Intel XPU hardware by the submitter.