vLLM Backport
A fork of vLLM focused on running recent frontier models on older NVIDIA GPUs such as the A6000, 3090, and A100, distributed as prebuilt Docker images.
vLLM Backport is a fork of the vLLM inference engine that targets older data-center and consumer NVIDIA cards (A6000, RTX 30-series, A100) rather than the newest hardware. It tracks support for recent model families such as DeepSeek-V4 Flash, Qwen3.8, and GLM-5.3, with quantization options including native FP4 and AWQ W4A16, and ships a paired LMCache fork for production KV-cache serving.
Prebuilt images are published to Docker Hub on every push, tagged by target GPU architecture (sm86, sm80, sm89) as well as by release version, and the entrypoint runs vllm serve with lmcache available in the same image. Images are single-architecture builds without Hopper/FA3 kernels, so users pick the tag matching their GPU.
The project also documents tuning recommendations, such as CUDA graph capture settings and NCCL configuration, for getting the best decode throughput on Ampere-class hardware.