Tracking items to improve performance for long max-seqlen workloads, including training and generation performance at long context.