MWE differentiate callbacks

mwe.jl

# https://github.com/SciML/SciMLSensitivity.jl/issues/720
using Zygote
using StochasticDiffEq, SciMLSensitivity

function mwe1()
    x0 = [0.]
    drift(dx, x, p, t) = (dx .= p)
    noise(dx, x, p, t) = (dx .= 0.)
    n0 = zeros(1,1)
    T = 100.
    p0 = [1.]
    cb = ContinuousCallback((u,t,int)->(u[1]-1), terminate!, nothing)
    prob = SDEProblem(drift, noise, x0, T, p0, noise_rate_prototype = n0, callback=cb)

    sensealg = InterpolatingAdjoint(autojacvec=ReverseDiffVJP(), noisemixing=true)
    Zygote.gradient(p0) do ps
        solve(prob, EM(), dt=0.1, p=ps, sensealg=sensealg)[end][1]
    end
end

# fails on about 50%
function _mwe2()
    x0 = [0.]
    drift(dx, x, p, t) = (dx .= p)
    noise(dx, x, p, t) = (dx .= .1)
    n0 = zeros(1,1)
    T = 100.
    p0 = [1.]
    cb = ContinuousCallback((u,t,int)->(u[1]-1), terminate!)
    prob = SDEProblem(drift, noise, x0, T, p0, noise_rate_prototype = n0, callback=cb)

    sensealg = InterpolatingAdjoint(autojacvec=ReverseDiffVJP(), noisemixing=true)
    Zygote.gradient(p0) do ps
        solve(prob, EM(), dt=0.1, p=ps, sensealg=sensealg)[end][1]
    end
end

function mwe2()
    for i in 1:100
        _mwe2()
    end
end