Using AR removes forecasts?

[M-Adrian-N]

Hi everyone,
I'm fairly new to forecasting. I'm using NP to forecast demand of bike models per country (Col A - Item, col B - date, col c- Qty).https://colab.research.google.com/drive/1spcVzuf9OKxjwGPjuoSpFOKXEFWCxh_3?usp=sharing

In my example I'm doing it just on a few bike models, on daily data starting 2019.
What I've noticed is at after training the model with live plots, overfitting gets better if I'm using AR parameter, (basically setting it to 1 greatly reduces the error term.
My issues are:

1. The thing is that now the model is not producing actual forecasts anymore, anytime n_lag is set to something different than 0 I'm not getting any forecasts (those 45 days I was looking for).

2. How do I optimize it so it's fast - I'll have to run this for over 500 bike models across multiple countries. I've played around with nr of epochs and batch_size and found 60/40 for my data set give a low error and no overfitting based on initial results - but having it done so many times across that many models vs countries takes ages. Would the Global model be a solution to this considering - each model has a different demand, and each country is different?

Thank you

[M-Adrian-N]

also the file I've used is here : https://we.tl/t-l0QWw2MWfj
Cause colab removes files after runtime is refreshed or I'm too novice to get around that