I thought that fine-tuning is no longer being done in the industry, instead transformer adapters like LoRA are being used? Having 1000 fine-tune models for each

I thought that fine-tuning is no longer being done in the industry, instead transformer adapters like LoRA are being used? Having 1000 fine-tune models for each customer seems too heavy when one can have instead 1000 transformer adapters and swap them during the inference for each batch. I mean there are tricks like Q-GaLore that allow training LLaMA-7B on a single 16GB GPU but LoRA still seems to be better for production to me.