llvm

21ff7a50 - [CUDA] Use USM Device memory for multi-GPU tests instead of USM Shared

[CUDA] Use USM Device memory for multi-GPU tests instead of USM Shared CUDA Managed Memory (USM Shared) does not support explicit cross-device copies between separate per-device allocations. NVIDIA documentation shows Managed Memory as a single shared buffer with automatic migration. For multi-GPU tests on CUDA, use USM Device memory which supports cudaMemcpyPeer for peer-to-peer transfers, as documented in CUDA Programming Guide section 3.4.2.1.