@jmschreiber91 I think that might be an intentional difference; if delta_in < eps, then query ~= ref, in which case the elementwise nonlinearity derivatives are the same, but the max pool derivative wouldn’t necessarily be the same; they want to use the query’s max pool derivative
@jmschreiber91 i dont think that's the bug? I think that `inputs` is a tensor whose first dimension is length 2, with the [0] entry corresponding to the query input and the [1] entry corresponding to the reference input, and that indexing might be deliberate