vLLM air-gapped deployment
Running vLLM with no internet access: staging model weights and images inside the boundary, offline API docs, turning off usage stats, private cache directories and the internal ports that must stay closed.
vLLM features
Pages on deployment and hardening features of self-hosted vLLM, from container images and hardware targets to tenant isolation and FIPS hosts.
vLLM Docker deployment
Running vLLM from the official vllm/vllm-openai image on a single GPU host: shared memory settings, model cache volumes, gated model tokens, custom images for extra dependencies and version pinning.
vLLM on AMD ROCm GPUs
Self-hosting vLLM on AMD Instinct or Radeon GPUs: supported accelerators and ROCm versions, the Python 3.12 wheel requirement, the ROCm container image and when a source build is needed.
vLLM multi-tenant isolation and cache salting
Serving several teams or customers from vLLM safely: why one process is not an isolation boundary, per-tenant instances versus a gateway, and per-tenant cache_salt values against the prefix cache timing side channel.
vLLM on FIPS-enabled hosts
Configuring vLLM for FIPS-enabled Linux hosts: hash algorithm settings, TLS cipher restrictions, dependencies that bundle their own crypto, and why unencrypted inter-node traffic needs an external FIPS-validated layer.
vLLM dynamic LoRA adapter loading in production
Operating runtime LoRA adapter loading on a shared vLLM server: the flags that enable it, restricting the load and unload endpoints to administrators at the proxy, and when static adapter configuration is safer.
