Developers and researchers who need high throughput multi user LLM inference on their own GPUs.

A vLLM derived inference engine, renamed Sonar by dphnAI, formerly PygmalionAI, that serves large language models with continuous batching, paged attention, and broad quantization support.

Self hosted on local or cloud GPUs via pip install or Docker, exposing an OpenAI compatible API.

Originally released in 2023 as Aphrodite Engine and actively developed through 2026 under its new Sonar name.

Gives teams vLLM class throughput plus extra quantization and sampling options favored by the open model hosting community.

Built on PyTorch and CUDA, and shares much of its core architecture with the vLLM project.

Open source under AGPL 3.0 on GitHub, no account required, runs entirely on hardware the operator controls, the project renamed from Aphrodite Engine to Sonar and moved from the PygmalionAI to the dphnAI GitHub org.