# Deploying vLLM AMD

**URL:** <https://community.vastdata.com/t/deploying-vllm-amd/2066>\
**Category:** Education\
**Created:** [September 11, 2026, 4:00pm UTC](https://community.vastdata.com/t/deploying-vllm-amd/2066 "2026-09-11T16:00:36Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![jason.massae](https://sea2.discourse-cdn.com/flex022/user_avatar/community.vastdata.com/jason.massae/32/2233_2.png) [@jason.massae](https://community.vastdata.com/u/jason.massae)\
**Post date:** [September 11, 2026, 4:00pm UTC](https://community.vastdata.com/t/deploying-vllm-amd/2066/1 "2026-09-11T16:00:36Z")

</div>

As long-context LLM inference scales, GPU High Bandwidth Memory (HBM) becomes the primary bottleneck as the Key-Value (KV) cache grows linearly. Pairing vLLM’s native OffloadingConnector with host DRAM and VAST Data spills KV from GPU → CPU → NFS, bypassing GPU memory limits and cutting warm Time-To-First-Token (TTFT) when prefixes hit the CPU/FS tiers.

Read more: [Deploying vLLM AMD](https://kb.vastdata.com/docs/deploying-vllm-amd)

Let us know if you have any questions 🙂