ERROR `no method matching increment_field!!(::Base.TwicePrecision{Float64}, ::Float64, ::Val{:lo})` #551

Moelf · 2025-04-13T03:30:27Z

function _ff(x, p)
    rx = x ./ 13600
    @. p[1] * (1 - rx)^p[2] * rx^(p[3])
end
DI.gradient(x -> sum(_ff(x, [1.0, 2.0, 3.0])), DI.AutoMooncake(; config=nothing), 100:1:105)

A different backend works:

DI.gradient(x -> sum(_ff(x, [1.0, 2.0, 3.0])), DI.AutoForwardDiff(), 100:1:105)

6-element Vector{Float64}:
 1.1693493343981998e-8
 1.1926169107999864e-8
 1.2161089728860294e-8
 1.2398253829652019e-8
 1.2637660033721705e-8
 1.2879306964673936e-8

Moelf · 2025-04-13T03:36:48Z

Notice, this is wrong:

DI.gradient(x -> sum(_ff(x, [1.0, 2.0, 3.0])), DI.AutoMooncake(; config=nothing), collect(100:1:105))

6-element Vector{Mooncake.NoTangent}:
 Mooncake.NoTangent()
 Mooncake.NoTangent()
 Mooncake.NoTangent()
 Mooncake.NoTangent()
 Mooncake.NoTangent()
 Mooncake.NoTangent()

You have to specifically do:

DI.gradient(x -> sum(_ff(x, [1.0, 2.0, 3.0])), DI.AutoMooncake(; config=nothing), collect(100:1.0:105))

6-element Vector{Float64}:
 1.1693493343982001e-8
 1.1926169107999864e-8
 1.2161089728860293e-8
 1.2398253829652017e-8
 1.2637660033721706e-8
 1.287930696467394e-8

willtebbutt · 2025-04-13T08:52:56Z

Thanks for opening this. I'll have a quick look at the first point -- the twice-precision arithmetic interacts in a slightly weird way with Mooncake, so I suspect that your work-around (collecting) the vector is the correct thing to do.

Notice, this is wrong:

This is actually intended behaviour. Mooncake draws a distinction between integers and reals -- gradients w.r.t. integers are always zero.

willtebbutt · 2025-04-13T09:03:50Z

I've had a quick look at this -- the source of the problem is some function that's grabbing the lo field of a TwicePrecision. I'll have to have a dig around at some point to figure out what's actually needed here. In the mean time, your collecting strategy seems sensible.

willtebbutt added the enhancement New feature or request label Apr 13, 2025

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

ERROR `no method matching increment_field!!(::Base.TwicePrecision{Float64}, ::Float64, ::Val{:lo})` #551

ERROR `no method matching increment_field!!(::Base.TwicePrecision{Float64}, ::Float64, ::Val{:lo})` #551

ERROR no method matching increment_field!!(::Base.TwicePrecision{Float64}, ::Float64, ::Val{:lo}) #551

ERROR no method matching increment_field!!(::Base.TwicePrecision{Float64}, ::Float64, ::Val{:lo}) #551

Comments

ERROR `no method matching increment_field!!(::Base.TwicePrecision{Float64}, ::Float64, ::Val{:lo})` #551

ERROR `no method matching increment_field!!(::Base.TwicePrecision{Float64}, ::Float64, ::Val{:lo})` #551