vLLM 2080 Ti Definitive Edition
A hardware-focused fork of vLLM tuned for dual RTX 2080 Ti and other Turing (SM75) GPUs, adding launcher profiles and validated inference routes for running 27B to 35B class models locally.
vLLM 2080 Ti Definitive Edition is a fork of vLLM that preserves SM75-specific source changes, launcher profiles and validation evidence needed to reproduce inference on Turing-generation cards: dual RTX 2080 Ti GPUs, Tesla T10/T40/T4, TITAN RTX and Quadro RTX 6000/8000. It targets serving 27B and 35B class large language models locally rather than only small models, using two 22 GB RTX 2080 Ti cards joined by NVLink for 44 GB of combined VRAM.
The fork adds Marlin, FlashInfer/FlashQLA, TurboQuant/INT8 KV cache, MTP/DFlash2 speculative decoding and CUDA Graph support on top of the upstream vLLM baseline, and supports tensor-parallel, pipeline-parallel and mixed TP/PP inference layouts across validated multi-GPU configurations.
The current release line targets Ubuntu 26.04, Linux kernel 7, GCC/G++ 15, CUDA 13.0 and PyTorch 2.13; an older, no-longer-maintained line covers CUDA 12.8 and PyTorch 2.11.