Optimizing the locations of the Jacobians

Hi!

Thanks for the contributions!! I am facing a weird sutiation in which I am computing the jacobians of a model w.r.t to the parameters (which are fixed and cannot be tuned) and I want to optimize the input locations of those jacobians, lets say the x variables. Two different things are happening to me when trying to to this:
1. The locations retain a .grad value when using the backward call to compute the jacobians w.r.t to the parameters using BackPack.
2. The backward call of my loss function does not alter the .grad value of the inputs used to compute the jacobians,

Is there any way I can fix this?

No files or tests are named. Start by reproducing the two backward-call cases described for BackPACK Jacobians, using fixed parameters, optimized input locations, and the loss function; done means the intended input gradient behavior is consistent and documented by a regression test.

Written by the indexing model from the issue text.