If you don't have access to the pipeline, you'll have to load the model and tokenizsr yourself, tokenize the text, call the model, and return the newly generated text
So if you have a model that you can already test giving inputs to, you likely have this all written already somewhere
