Apocrypha

vllm

High-throughput, memory-efficient inference and serving engine for LLMs

vLLM is a fast and easy-to-use library for LLM inference and serving, with a Python API and an OpenAI-compatible HTTP server. USE=cpu/cuda/rocm pick a mutually-exclusive VLLM_TARGET_DEVICE; with none set the build is device-agnostic (API surface only).

Available in

OverlayNewestEbuildsLast activity
stuff GitHub ↗ 0.25.1 5 2 d details ›

Versions & arches

VersionOverlay amd64 Committed
0.25.1 stuff amd64 testing 5 d view · download · history ↗
0.25.0 stuff amd64 testing 15 d view · download · history ↗
0.24.0 stuff amd64 testing 22 d view · download · history ↗
0.23.0 stuff amd64 testing 22 d view · download · history ↗
0.22.1 stuff amd64 testing 22 d view · download · history ↗

Use flags of 0.25.1

  • cpu Build for CPU inference (VLLM_TARGET_DEVICE=cpu); pull torchaudio + numba
  • cuda Enable NVIDIA CUDA support (computation on GPU)
  • humming Pull dev-python/humming-kernels for the optional CUDA-only "humming" MXFP4 quantization backend (needs USE=cuda)
  • rocm Build for AMD ROCm inference (VLLM_TARGET_DEVICE=rocm); pull HIP libs + torch{audio,vision}
  • rust Build the optional vllm-rs Rust serving frontend (opt-in at runtime via VLLM_USE_RUST_FRONTEND=1)