Fine tuning a model like llama is super easy. You can fine tune models up to 33B parameters on a single consumer GPU at around 3 epochs per minute (less than 6 hours for 1000 epochs).

You can even fine tune on a free Google Colab GPU.

Start here: https://github.com/lxe/simple-llama-finetuner

Running it fully locally isn't easy because you need a complex setup with GPUs. Can you give some more color on exactly what you are trying to do (number of PDFs, types of queries, etc.)? With some more info we can give you some guidance.