Does Tapir support vector mode?

[gdalle]

By "vector mode" I mean the ability to propagate a chunk of seeds in the pullback, similar to ForwardDiff's Chunk or Enzyme's BatchDuplicated.
Thanks to JuliaDiff/DifferentiationInterface.jl#320, DifferentiationInterface will use vector mode to speed up (dense & sparse) Jacobians and Hessians.

[yebai]

I will leave this to @willtebbutt for an answer.

I have a question about vector mode in general. Given that Tapir supports input—and data-dependent control flows, what happens if the forward pass takes different paths for different inputs?

[gdalle]

In autodiff parlance, vector mode refers to propagating several pullback seeds dy1, ..., dyk with the same input x. In other words, the control flow (which is dictated by x) remains unchanged

[willtebbutt]

Currently no support for vector mode I'm afraid. You'll have to fall back to just running the forwards- and reverse-passes several times.