Switching from mmap to parallel pread. From 0.5tok/sec to almost 4tok/sec. Running GPU work while reading missed experts also helped a lot, 4.4 -> 4.7