Serving Large Language Models (LLMs) requires careful management of GPU memory, especially for the KV-cache that stores past activations. The current standard, PagedAttention, dynamically allocates memory but creates non-contiguous memory blocks, forcing developers to rewrite attention kernels and manage complex block tables. This approach introduces performance bottlenecks and hinders the adoption of cutting-edge attention optimizations. Enter vAttention, a novel memory management system that leverages the operating system's virtual memory capabilities. By pre-reserving contiguous virtual memory and dynamically allocating physical memory only when needed, vAttention eliminates the need for block tables and allows the use of unmodified, highly optimized attention kernels like FlashAttention. This simplifies integration and boosts performance. Our experiments with models like Yi-6B, Llama-3-8B, and Yi-34B show vAttention significantly improves both prefill and decode throughput. For prefill, vAttention outperforms PagedAttention by up to 24%, while in decode operations, it matches the performance of the best PagedAttention implementations. This translates to up to a 29% improvement in end-to-end serving throughput, especially for workloads dominated by prefill operations. vAttention achieves this efficiency through several key optimizations: overlapping memory allocation with computation, deferring memory reclamation, and proactively allocating pages ahead of time. These techniques effectively hide the latency of memory allocation, ensuring smooth and responsive LLM serving. vAttention represents a significant step forward in LLM serving, simplifying development and unlocking the full potential of state-of-the-art attention kernels. By leveraging the OS's memory management capabilities, vAttention paves the way for more efficient and scalable LLM deployments.