Gradient of Predictor output with respect to input

[bradyneal]

I'm doing forecasting. How can I get the gradient of a forecasted output of a GluonTS Predictor (at a specific time in the future) with respect to specific inputs (e.g. a subset of feat_dynamic_real and feat_dynamic_cat at a specific point in time in the past)?