Fast, high-throughput inference engine for serving LLMs with PagedAttention, continuous batching, and OpenAI-compatible API.

Features

vLLM is a fast and easy-to-use library for LLM inference and serving, originally from UC Berkeley. It delivers state-of-the-art throughput through PagedAttention, continuous batching, speculative decoding, and optimized CUDA kernels. vLLM exposes an OpenAI-compatible API, supports NVIDIA and AMD GPUs, and integrates seamlessly with Hugging Face models. It’s widely used for production LLM serving, inference endpoints, and high-performance local deployment of open-source models.

Similar tools