Estimate the GPU memory needed to serve a 70B model, accounting for weights, KV cache, activations, and fragmentation.

PagedAttention maps logical KV blocks to noncontiguous physical pages, much like virtual memory. It reduces fragmentation, enables prefix sharing and copy-on-write, and lets the scheduler pack more sequences safely.

Make interfaces and ownership explicit; add versioning, access control, monitoring, canary rollout, rollback, and a plan for delayed labels or human review.