A high-throughput and memory-efficient inference and serving engine for LLMs

Easy, fast, and cheap LLM serving for everyone

| Documentation | Blog | Discussions |

Latest News 🔥

• [2023/06] We officially released vLLM! vLLM has powered LMSYS Vicuna and Chatbot Arena since mid April. Check out our blog post.

vLLM is a fast and easy-to-use library for LLM inference and serving.

vLLM is fast with:

• State-of-the-art serving throughput
• Efficient management of attention key and value memory with PagedAttention
• Dynamic batching of incoming requests
• Optimized CUDA kernels

vLLM is flexible and easy to use with:

• Seamless integration with popular HuggingFace models
• High-throughput serving with various decoding algorithms, including parallel sampling, beam search, and more
• Tensor parallelism support for distributed inference
• Streaming outputs
• OpenAI-compatible API server

vLLM seamlessly supports many Huggingface models, including the following architectures:

• GPT-2 (gpt2, gpt2-xl, etc.)
• GPT BigCode (bigcode/starcoder, bigcode/gpt_bigcode-santacoder, etc.)
• GPT-NeoX (EleutherAI/gpt-neox-20b, databricks/dolly-v2-12b, stabilityai/stablelm-tuned-alpha-7b, etc.)
• LLaMA (lmsys/vicuna-13b-v1.3, young-geng/koala, openlm-research/open_llama_13b, etc.)
• OPT (facebook/opt-66b, facebook/opt-iml-max-30b, etc.)

Install vLLM with pip or from source:

pip install vllm

Getting Started

Visit our documentation to get started.

• Installation
• Quickstart
• Supported Models

Performance

vLLM outperforms HuggingFace Transformers (HF) by up to 24x and Text Generation Inference (TGI) by up to 3.5x, in terms of throughput. For details, check out our blog post.

Serving throughput when each request asks for 1 output completion.

Serving throughput when each request asks for 3 output completions.

Contributing

We welcome and value any contributions and collaborations. Please check out CONTRIBUTING.md for how to get involved.