I am running GPT-J-6B inference on 13GB of VRAM. FP16 is a must at this scale. I've also loaded GPT-NeoX-20B using int8 quantization onto a single RTX 3090 before (24GB VRAM) however inference latency was quite high.