/ The Arena

Problem Overview

In modern LLM serving engines like vLLM, memory fragmentation is eliminated by managing the Key-Value (KV) Cache as a virtual memory paging system.

In this landmark capstone, you will implement a complete PagedAttention Virtual Memory Block Allocator that handles:

1. Block Table Allocation: Mapping logical sequence token blocks to non-contiguous physical GPU VRAM blocks.
2. Dynamic Token Appending: Automatically allocating new physical pages when sequence boundaries are crossed.
3. Copy-on-Write (CoW): Forking sequences with shared physical pages and duplicating pages only on mutation.
4. Sequence Deallocation: Returning freed physical blocks back to the free pool with reference counting.