Adjoint crash to desktop on v1.7-beta3 without autodiff #479
Comments
|
Yeah, it seems that if the compiler even sees that branch it will crash on v1.7-beta3. This is documented in SciML/DiffEqFlux.jl#586 (comment) but seems to be known in Base as JuliaLang/julia#41503 . It was most likely fixed in JuliaLang/julia#41516, but it would be good to check on the latest v1.7 (i.e. build from source) to make sure. |
|
It works on the latest nightly. |
|
Nightly still crashes on steady state adjoints for me |
|
What does the crash look like? Post the segfault and let's get a Julia Base issue open. |
|
If I remove res2a to res2e it does not crash. using DiffEqSensitivity, SteadyStateDiffEq, OrdinaryDiffEq
function f!(du, u, p, t)
du[1] = p[1] + p[2] * u[1]
du[2] = p[3] * u[1] + p[4] * u[2]
end
function jac!(J, u, p, t) # df/dx
J[1,1] = p[2]
J[2,1] = p[3]
J[1,2] = 0
J[2,2] = p[4]
nothing
end
function paramjac!(fp, u, p, t) # df/dp
fp[1,1] = 1
fp[2,1] = 0
fp[1,2] = u[1]
fp[2,2] = 0
fp[1,3] = 0
fp[2,3] = u[1]
fp[1,4] = 0
fp[2,4] = u[2]
nothing
end
function dg!(out, u, p, t, i)
(out .= -2.0 .+ u)
end
function g(u, p, t)
sum((2.0 .- u).^2) / 2 + sum(p.^2) / 2
end
u0 = zeros(2)
p = [2.0,-2.0,1.0,-4.0]
f2 = ODEFunction(f!;jac=jac!)
prob2 = SteadyStateProblem(f2,u0,p)
sol2 = solve(prob2,DynamicSS(Rodas5()))
res2a = adjoint_sensitivities(sol2,DynamicSS(Rodas5()),sensealg=SteadyStateAdjoint(),g,dg!)
res2b = adjoint_sensitivities(sol2,DynamicSS(Rodas5()),sensealg=SteadyStateAdjoint(),g,nothing)
res2c = adjoint_sensitivities(sol2,DynamicSS(Rodas5()),sensealg=SteadyStateAdjoint(autodiff=false),g,nothing)
res2d = adjoint_sensitivities(sol2,DynamicSS(Rodas5()),sensealg=SteadyStateAdjoint(autojacvec=TrackerVJP()),g,nothing)
res2e = adjoint_sensitivities(sol2,DynamicSS(Rodas5()),sensealg=SteadyStateAdjoint(autojacvec=ReverseDiffVJP()),g,nothing)
res2f = adjoint_sensitivities(sol2,DynamicSS(Rodas5()),sensealg=SteadyStateAdjoint(autojacvec=ZygoteVJP()),g,nothing)(@v1.8) pkg> status
Status `C:\Users\arno\.julia\environments\v1.8\Project.toml`
[41bf760c] DiffEqSensitivity v6.58.0 `https://github.com/SciML/DiffEqSensitivity.jl.git#master`
[1dea7af3] OrdinaryDiffEq v5.63.3
[9672c7b4] SteadyStateDiffEq v1.6.4julia> versioninfo()
Julia Version 1.8.0-DEV.394
Commit 3d058b205e (2021-08-24 12:35 UTC)
Platform Info:
OS: Windows (x86_64-w64-mingw32)
CPU: AMD Ryzen 9 5900X 12-Core Processor
WORD_SIZE: 64
LIBM: libopenlibm
LLVM: libLLVM-12.0.1 (ORCJIT, znver3) |
|
@DhairyaLGandhi @simeonschaub is this failure known? Looks like a core Zygote/IRTools kind of failure? |
|
This no longer crashes to desktop on 1.7.1 (did not test on 1.7). |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
When I step through it with a debugger it works fine.
It then does not go into this branch as Chris suggested
https://github.com/SciML/DiffEqSensitivity.jl/blob/master/src/adjoint_common.jl#L125
The text was updated successfully, but these errors were encountered: