the trend step artifact used to keep track the time of the data so we know the expected trend of the input data. For example, on the first data which is trend_step = 1 the trend value is 10, then if the trend_step = 10 (the tenth data) our regressor will predict the trend value of the selected trend_step. this method is still in research to make it more efficient so it doesn't need to upload artifact every request

Make sense! I would suggest you add a GitHub issue with feature request for fast key/value storage that supports multi instance, I think this usage pattern will be greatly appreciated
BTW as an optimization I would use Task scalars (they are send in the background, you can relativity easily get the latest value etc.) do you also need it to be atomic ?

X is the sequence generated from df. df contains 2 columns (date and value). Size of X for this example is (1,60,1) with type np.array, X is sequence with size 1(number of data), 60(timestep), 1(value from 'value' column of df).

So in theory 1,60,-1 should work as a size for Triton, Are you getting an error?
(BTW: if you were to manually run the model inference I'm assuming you would have created a 3d matrix where the dims are 1,60,<batch_size>, is that correct?)