We might be able to increase the efficiency of the gradient of the loglikelihood computation in fitting.jl if we define a custom method that derives the full gradient (so every partial) simultaneously and builds a dense representation of (data ./ composite). As it is, when we calculate each partial separately it recomputes for every partial.
Our solves are still quite fast but I think this would be an improvement.
We might be able to increase the efficiency of the gradient of the loglikelihood computation in fitting.jl if we define a custom method that derives the full gradient (so every partial) simultaneously and builds a dense representation of (data ./ composite). As it is, when we calculate each partial separately it recomputes for every partial.
Our solves are still quite fast but I think this would be an improvement.