QwenRTX 3090
RTX 3090 Single Card Doubles Qwen 27B Inference — Local LLM Playbook Rewritten
Solo dev L-Forster's CUDA megakernel lifts Qwen3-27B on a single RTX 3090 from 73 to 140 tok/s (~2x), KL divergence 0.0009. Open-sourced as open-jet.
Oct 12·2 min read
