fix: disable USM compression for any multi-device context

[F3zz1k]

Summary

Closes #921

The peer-USM compression guard in SVMAllocsManager::createUnifiedMemoryAllocation is gated on device->hasAnyPeerAccess().value_or(false). For the first USM device allocation in a multi-device L0 context that std::optional can be observed as Some(false) (peer-access cache not yet populated by initializePeerAccessForDevices). The disable arm is skipped, the BO is allocated compressed, and any later peer-import returns garbage through dma-buf because the importing device has no matching surface metadata.

This patch adds memoryProperties.rootDeviceIndices.size() > 1 as an additional, synchronously-known gate. Multi-device contexts conservatively disable compression, correct because a BO in a multi-device context can become peer-imported at any subsequent point in the application's lifetime.

Verification

• 2× and 4x Intel Arc Pro B70 (BMG-G31), Linux 7.1-rc4 + xe (with Leon Romanovsky's "dma-buf: Always build with DMABUF_MOVE_NOTIFY")
• Built locally with the patch loaded via LD_LIBRARY_PATH against compute-runtime 26.05/26.18, cross-device sycl::queue::memcpy(host, peer_usm_ptr) returned garbage before; returns source data correctly after
• Cause-and-effect confirmed by reverting the patch on the same build, returns to garbage
• The OpenCL UNRECOVERABLE_IF(!allocation) at enqueue_svm.h:308 symptom on the same workload is resolved by this fix
• llama.cpp tensor-split (--device SYCL0,SYCL1 --split-mode layer) regression-tested on both patched and stock, no behavioural change there (that workload's allocation timing happens to dodge the race)

Note: This patch remains necessary even on kernels with DRM_XE_VM_BIND_FLAG_DECOMPRESS (7.1-rc1+).
The kernel-side xe_bo_decompress() at drivers/gpu/drm/xe/xe_bo.c:3537 silently skips non-VRAM buffers, and
peer-imported BOs are ttm_bo_type_sg, so cross-device decompression is a no-op even when the flag is
accepted. Until that's addressed kernel-side, disabling compression at allocation time for multi-device
contexts is the correct behaviour.

Unit tests

Not added in this commit. The targeted code path requires a platform where releaseHelper->isUsmCompressionSupportedOnPeerAccess() returns false (release2001 / release2002 = BMG / LNL) AND usmCompressionSupported(hwInfo) returns true. The default MockReleaseHelper returns true for the former, so a test under shared/test/unit_test/memory_manager/ would skip on most configurations.

A natural home would be shared/test/unit_test/xe2_hpg_core/ which already uses the BMG release helper, or a new test using MockReleaseHelper overrides. If really needed, I could add it, would appreciate guidance on the preferred location and pattern (the existing givenLocalMemoryEnabledAndCompressionEnabledThenDeviceSideSharedUsmIsCompressed test in unified_memory_manager_tests.cpp:721 is the natural sibling but uses single-root setup).

Related

• [Bug]: XPU TP=2 on dual Intel Arc Pro B70 (Battlemage): GP fault + xe BCS engine reset reproduces in intel/vllm:0.17.0-xpu on Ubuntu 24.04 HWE 6.17 vllm-project/vllm#41663 (same hardware, downstream symptom on TP=2)
• [GSD-12641] USM device allocation fails with UR_RESULT_ERROR_OUT_OF_DEVICE_MEMORY when Level Zero context spans 2 discrete Arc A770 GPUs (regression from 25.40 to 26.09) #916 (USM device alloc fails on dual A770 L0 context)
• Multi-BMG Level Zero device enumeration broken by deferred internal-engine init (regression NEO 25.40→25.44, Arc Pro B50/B60/B70) #921 (multi-BMG L0 device enumeration)
• [Xe2/BMG-G31] multi-rank MPI + Level Zero: resource_info.cpp:15 abort after upgrading CR 26.05 → 26.14 (Ubuntu 26.04, xe driver) #922 (BMG multi-rank MPI + L0 crash)
• Kernel-side prerequisite (xe_dma_buf_pin -EINVAL on peer-imported BO) fixed upstream by Leon Romanovsky's "dma-buf: Always build with DMABUF_MOVE_NOTIFY", in mainline Linux 7.1.

cc @ivysochyn — this directly extends the guard you added in dc85f33 and refactored in a832f01. Race observed where hasAnyPeerAccess() is Some(false) for the very first multi-device alloc.