The inference side is fine, nowadays. llama.cpp has had a GPU-agnostic Vulkan backend for a while, it's the training side that tends to be a sticking point for consumer GPUs.