`remake` fails silently for vector-valued parameters

[ysfoo]

remake does not update parameters that are vector-valued. The following example is taken from the Catalyst documentation:

using OrdinaryDiffEq
using Catalyst

two_state_model = @reaction_network begin    
    @parameters k[1:2]
    @species X(t)[1:2]
    (k[1],k[2]), X[1] <--> X[2]
end

u0 = [:X => [0.0, 2.0]];
tspan = (0.0, 1.0);

ps = [:k => [1.0, 2.0]];
oprob = ODEProblem(two_state_model, u0, tspan, ps);
oprob_rmk = remake(oprob; p=[:k => [0.0, 0.0]]);

oprob_rmk does not have the updated parameters; both solve(oprob)(1.0) and solve(oprob_rmk)(1.0) produce the same output.

Interestingly, this behaviour is absent if the parameters are defined outside of the reaction_network macro:

@parameters k[1:2]
two_state_model = @reaction_network begin    
    @species X(t)[1:2]
    ($k[1],$k[2]), X[1] <--> X[2]
end

u0 = [:X => [0.0, 2.0]];
tspan = (0.0, 1.0);

ps = [k => [1.0, 2.0]];
oprob = ODEProblem(two_state_model, u0, tspan, ps);
oprob_rmk = remake(oprob; p=[k => [0.0, 0.0]]);

Here, solve(oprob_rmk)(1.0) produces the expected output, i.e. [0.0, 2.0].

[AayushSabharwal]

This doesn't work because Catalyst scalarizes vector parameters declared in the @reaction_network macro. As a result, MTK considers each of the elements as separate parameters. I can get it to recognize the vector version of k in limited cases, but @TorkelE would it be possible for Catalyst to not scalarize here?

[ChrisRackauckas]

Ahh that makes sense. The scalarizing can also potentially inhibit some of the codegen optimizations that are starting to happen as well, so it would be best to solve this by not scalarizing in Catalyst.

[TorkelE]

I think this might be fine on Catalyst master, but will need to check it tomorrow