Speeding up my logsumexp function

[mcabbott]

I wondered whether ReverseDiff had been quietly fixed, but it seems not, Use with ReverseDiff · Issue #21 · mcabbott/Tullio.jl · GitHub has a simple test. So I think you ought to get errors if Turing is trying to do this, at all. Is there a chance Turing isn’t using AD at all?

@vkv

For m = 600 and n = 1200

lsexp1 function fails with this double free or corruption (!prev) error.
lsexp2 functions works fine and the model starts sampling although the ETA is 38hours.

So lsexp1 which is purely broadcasting is giving you this corruption error? That might be worth isolating. Is A still sparse here?