Back to glossary

Term

vLLM

vLLM is an open-source high-throughput inference server for LLMs — known for PagedAttention, continuous batching, and an OpenAI-compatible API endpoint, widely used in production self-hosting setups.

vLLM — explained

vLLM is built for server workloads: many concurrent requests, high tokens per second, GPU utilization as close to 100 % as possible. The core trick is PagedAttention — the KV cache is split into fixed blocks and managed like memory pages in an OS, which avoids fragmentation and makes continuous batching practical: running requests are merged token by token with newly incoming ones, instead of waiting for fixed batch boundaries. vLLM was originally developed at UC Berkeley.

Example / Practical context

On an A100 or H100, vLLM can run Llama 3.1 70B (via tensor parallelism across multiple GPUs) or Qwen 2.5 32B as a self-hosted backend for an internal ChatGPT-style assistant. The OpenAI-compatible endpoint (/v1/chat/completions) means existing clients (LangChain, LlamaIndex, custom SDKs) can switch over without code changes. Quantized models (AWQ, GPTQ) are supported out of the box.

Ollama and LM Studio target single users and easy setup; vLLM targets servers with many concurrent users and maximum throughput. Hugging Face TGI and NVIDIA TensorRT-LLM compete in the same league as vLLM, with different emphases on hardware support, batching strategy, and license. llama.cpp, by contrast, is an inference engine on the CPU/edge level — not a server stack.

Entdecke mehr

Themenuebersicht