Machine: x86_64 Linux, 12 cores, 15920 MiB RAM | python 3.10.12, torch 2.13.0+cu130
Data: 128,000 rows x 768 float32 in 2,000 blocks of 64 rows | 5,000 random accesses per config

   max_cached_blocks  VmRSS delta  VMA delta
  (baseline, no access)      856.1MiB        636
                  16      +1.3MiB        +16
                 128      +7.3MiB       +112
             100,000    +203.0MiB      +1704
