Model weights (small tensors) must be in HBM for cuBLAS GEMM ops
which can't page-fault into managed memory. KV cache blocks are
large and numerous — prefetching them all fills HBM and causes
OOM. The 2 GiB threshold separates compute data from cache data.
Eager prefetching was filling HBM+EGM, causing subsequent
cudaMallocManaged calls to fail after model loading. On GH200
with EGM, pages should migrate on-demand via hardware page faults
over C2C NVLink. The cudaMemAdviseSetPreferredLocation(GPU) hint
is sufficient to prefer GPU placement with LPDDR fallback.
