vLLM: high-throughput LLM serving, OpenAI API, quantization.
vLLM: high-throughput LLM serving, OpenAI API, quantization.
# vLLM - High-Performance LLM Serving
## When to use
Use when deploying production LLM APIs, optimizing inference latency/throughput, or serving models with limited GPU memory. Supports OpenAI-compatible endpoints, quantization (GPTQ/AWQ/FP8), and tensor parallelism.
## Quick start
vLLM achieves 24x higher throughput than standard transformers through PagedAttention (block-based KV cache) and continuous batching (mixing prefill/decode requests).
**Installation**:
```bash
pip install vllm
``