Oneiros: KV Cache Optimization through Parameter Remapping for Multi-tenant LLM Serving

TL;DR

Oneiros addresses the KV-cache memory bottleneck in multi-tenant LLM inference by introducing dynamic parameter remapping, which repurposes GPU-allocated model parameter memory as KV cache without back-and-forth CPU offloads. The system employs a Remapping Controller and Async Transfer Engine to adaptively remap layers at per-token granularity, overlapping parameter transfers with GPU computation and exploiting the GH200's high CPU-GPU bandwidth. The approach outperforms state-of-the-art baselines, achieving large reductions in tail latency (up to ~99% TTFT/TBT improvements) and substantial throughput gains, with robust performance under both temporal and spatial GPU sharing. Overall, Oneiros provides a scheduler-agnostic, hardware-aware memory engine that enhances multi-tenant LLM inference efficiency by dynamically reallocating GPU memory between model parameters and KV cache.

Abstract

KV cache accelerates LLM inference by avoiding redundant computation, at the expense of memory. To support larger KV caches, prior work extends GPU memory with CPU memory via CPU-offloading. This involves swapping KV cache between GPU and CPU memory. However, because the cache updates dynamically, such swapping incurs high CPU memory traffic. We make a key observation that model parameters remain constant during runtime, unlike the dynamically updated KV cache. Building on this, we introduce Oneiros, which avoids KV cache swapping by remapping, and thereby repurposing, the memory allocated to model parameters for KV cache. This parameter remapping is especially beneficial in multi-tenant environments, where the memory used for the parameters of the inactive models can be more aggressively reclaimed. Exploiting the high CPU-GPU bandwidth offered by the modern hardware, such as the NVIDIA Grace Hopper Superchip, we show that Oneiros significantly outperforms state-of-the-art solutions, achieving a reduction of 44.8%-82.5% in tail time-between-token latency, 20.7%-99.3% in tail time-to-first-token latency, and 6.6%-86.7% higher throughput compared to vLLM. Source code of Oneiros is available at https://github.com/UT-SysML/Oneiros/.

Figures (22)

• Figure 1: Existing LLM systems.
• Figure 2: GPU Trends.
• Figure 4: Both KV cache swapping and parameter remapping expand the effective KV cache capacity by utilizing CPU memory as an extension of GPU memory. KV cache swapping incurs overhead due to frequent synchronization, as it requires backing up KV cache to CPU memory after each token generation once swap begins. Instead, parameter remapping is unidirectional with negligible synchronization overhead.
• Figure 5: LLM inference serving on GPUs. GH200 features 450 GB/s CPU-GPU bandwidth, while H100 only has 64 GB/s. Swapping KV cache/parameter between CPUs and GPUs during runtime can fully utilize the increased bandwidth.
• Figure 6: CPU Compute.