Latest release broke DistributionsAD #252

devmotion · 2023-05-18T23:34:08Z

Hej!

The latest release of FillArrays broke DistributionsAD. More precisely, some tests of ReverseDiff and Tracker started to fail. Due to the large number of changes it's not immediately clear to me what change exactly is the problem and hence I was wondering if maybe for some of the contributors in this repo it might be more obvious.

Two days ago tests passed without problems: https://github.com/TuringLang/DistributionsAD.jl/actions/runs/4992157009/jobs/8939577198 (the single failing test is unrelated, this specific job has been terminated in all recent runs)
Whereas since yesterday many tests fail: https://github.com/TuringLang/DistributionsAD.jl/actions/runs/4992157009

I compared the installed packages in https://github.com/TuringLang/DistributionsAD.jl/actions/runs/4992157009/jobs/8988760248 (fails) and https://github.com/TuringLang/DistributionsAD.jl/actions/runs/4992157009/jobs/8939577198 (succeeds). It turns out the diff is simple, namely

19c19
<   [31c24e10] Distributions v0.25.92
---
>   [31c24e10] Distributions v0.25.93
23c23
<   [1a297f60] FillArrays v1.0.0
---
>   [1a297f60] FillArrays v1.0.1

The changes in Distributions are completely unrelated: JuliaStats/Distributions.jl@v0.25.92...v0.25.93 only fixed some type inference issues of Truncated{<:Normal} but the failing tests are caused by MvNormal (no truncated distributions).
Thus FillArrays 1.0.1 must have broken it.

The text was updated successfully, but these errors were encountered:

jishnub · 2023-05-19T05:36:46Z

Hi, I looked at the error in https://github.com/TuringLang/DistributionsAD.jl/actions/runs/4992157009/jobs/8988760248

Matrixvariate distributions: Error During Test at /home/runner/work/DistributionsAD.jl/DistributionsAD.jl/test/ad/utils.jl:357
  Test threw exception
  Expression: ≈(Tracker.data((Tracker.gradient(f, x))[1]), finitediff, rtol = rtol, atol = atol)
  MethodError: no method matching Float64(::ForwardDiff.Dual{Nothing, Float64, 1})
  Closest candidates are:
    (::Type{T})(::Real, ::RoundingMode) where T<:AbstractFloat at rounding.jl:200
    (::Type{T})(::T) where T<:Number at boot.jl:760
    (::Type{T})(::AbstractChar) where T<:Union{AbstractChar, Number} at char.jl:50
    ...
  Stacktrace:
    [1] (::ComposedFunction{Type{Float64}, typeof(-)})(x::ForwardDiff.Dual{Nothing, Float64, 1})
      @ Base ./operators.jl:938
    [2] #8
      @ ./broadcast.jl:368 [inlined]
    [3] #8
      @ ./broadcast.jl:366 [inlined]
    [4] #2
      @ ./broadcast.jl:314 [inlined]
    [5] partial(::Base.Broadcast.var"#2#4"{Base.Broadcast.var"#8#10"{Base.Broadcast.var"#1#3", Base.Broadcast.var"#8#10"{Base.Broadcast.var"#5#6"{Base.Broadcast.var"#7#9"}, Base.Broadcast.var"#5#6"{Base.Broadcast.var"#7#9"}, Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#18"}, Base.Broadcast.var"#11#12"{Base.Broadcast.var"#13#14"}, ComposedFunction{Type{Float64}, typeof(-)}}, Base.Broadcast.var"#15#16"{Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#18"}}, Base.Broadcast.var"#11#12"{Base.Broadcast.var"#11#12"{Base.Broadcast.var"#13#14"}}, typeof(/)}, typeof(abs2)}, ::Float64, ::Int64, ::Float64, ::Float64)
      @ Tracker ~/.julia/packages/Tracker/xe6h6/src/lib/array.jl:596
    [6] _broadcast_getindex_evalf
      @ ./broadcast.jl:648 [inlined]
    [7] _broadcast_getindex
      @ ./broadcast.jl:621 [inlined]
    [8] getindex
      @ ./broadcast.jl:575 [inlined]
    [9] copy
      @ ./broadcast.jl:922 [inlined]
   [10] materialize(bc::Base.Broadcast.Broadcasted{Base.Broadcast.DefaultArrayStyle{1}, Nothing, typeof(Tracker.partial), Tuple{Base.RefValue{Base.Broadcast.var"#2#4"{Base.Broadcast.var"#8#10"{Base.Broadcast.var"#1#3", Base.Broadcast.var"#8#10"{Base.Broadcast.var"#5#6"{Base.Broadcast.var"#7#9"}, Base.Broadcast.var"#5#6"{Base.Broadcast.var"#7#9"}, Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#18"}, Base.Broadcast.var"#11#12"{Base.Broadcast.var"#13#14"}, ComposedFunction{Type{Float64}, typeof(-)}}, Base.Broadcast.var"#15#16"{Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#18"}}, Base.Broadcast.var"#11#12"{Base.Broadcast.var"#11#12"{Base.Broadcast.var"#13#14"}}, typeof(/)}, typeof(abs2)}}, Vector{Float64}, Int64, SubArray{Float64, 1, Matrix{Float64}, Tuple{Base.Slice{Base.OneTo{Int64}}, Int64}, true}, Vector{Float64}}})
      @ Base.Broadcast ./broadcast.jl:883
   [11] broadcast(::typeof(Tracker.partial), ::Base.RefValue{Base.Broadcast.var"#2#4"{Base.Broadcast.var"#8#10"{Base.Broadcast.var"#1#3", Base.Broadcast.var"#8#10"{Base.Broadcast.var"#5#6"{Base.Broadcast.var"#7#9"}, Base.Broadcast.var"#5#6"{Base.Broadcast.var"#7#9"}, Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#18"}, Base.Broadcast.var"#11#12"{Base.Broadcast.var"#13#14"}, ComposedFunction{Type{Float64}, typeof(-)}}, Base.Broadcast.var"#15#16"{Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#18"}}, Base.Broadcast.var"#11#12"{Base.Broadcast.var"#11#12"{Base.Broadcast.var"#13#14"}}, typeof(/)}, typeof(abs2)}}, ::Vector{Float64}, ::Int64, ::Vararg{Any, N} where N)
      @ Base.Broadcast ./broadcast.jl:821
   [12] ∇broadcast
      @ ~/.julia/packages/Tracker/xe6h6/src/lib/array.jl:600 [inlined]
   [13] copy(bc::Base.Broadcast.Broadcasted{Tracker.TrackedStyle, Tuple{Base.OneTo{Int64}}, typeof(Tracker.partial), Tuple{Base.RefValue{Base.Broadcast.var"#2#4"{Base.Broadcast.var"#8#10"{Base.Broadcast.var"#1#3", Base.Broadcast.var"#8#10"{Base.Broadcast.var"#5#6"{Base.Broadcast.var"#7#9"}, Base.Broadcast.var"#5#6"{Base.Broadcast.var"#7#9"}, Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#18"}, Base.Broadcast.var"#11#12"{Base.Broadcast.var"#13#14"}, ComposedFunction{Type{Float64}, typeof(-)}}, Base.Broadcast.var"#15#16"{Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#18"}}, Base.Broadcast.var"#11#12"{Base.Broadcast.var"#11#12"{Base.Broadcast.var"#13#14"}}, typeof(/)}, typeof(abs2)}}, Vector{Float64}, Int64, SubArray{Float64, 1, Matrix{Float64}, Tuple{Base.Slice{Base.OneTo{Int64}}, Int64}, true}, TrackedArray{…,Vector{Float64}}}})
      @ Tracker ~/.julia/packages/Tracker/xe6h6/src/lib/array.jl:630
   [14] materialize
      @ ./broadcast.jl:883 [inlined]
   [15] (::Tracker.var"#665#667"{Vector{Float64}, Base.Broadcast.var"#2#4"{Base.Broadcast.var"#8#10"{Base.Broadcast.var"#1#3", Base.Broadcast.var"#8#10"{Base.Broadcast.var"#5#6"{Base.Broadcast.var"#7#9"}, Base.Broadcast.var"#5#6"{Base.Broadcast.var"#7#9"}, Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#18"}, Base.Broadcast.var"#11#12"{Base.Broadcast.var"#13#14"}, ComposedFunction{Type{Float64}, typeof(-)}}, Base.Broadcast.var"#15#16"{Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#18"}}, Base.Broadcast.var"#11#12"{Base.Broadcast.var"#11#12"{Base.Broadcast.var"#13#14"}}, typeof(/)}, typeof(abs2)}, Tuple{SubArray{Float64, 1, Matrix{Float64}, Tuple{Base.Slice{Base.OneTo{Int64}}, Int64}, true}, TrackedArray{…,Vector{Float64}}}})(i::Int64)
      @ Tracker ~/.julia/packages/Tracker/xe6h6/src/lib/array.jl:603
   [16] ntuple
      @ ./ntuple.jl:49 [inlined]
   [17] (::Tracker.var"#back#666"{2, Base.Broadcast.var"#2#4"{Base.Broadcast.var"#8#10"{Base.Broadcast.var"#1#3", Base.Broadcast.var"#8#10"{Base.Broadcast.var"#5#6"{Base.Broadcast.var"#7#9"}, Base.Broadcast.var"#5#6"{Base.Broadcast.var"#7#9"}, Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#18"}, Base.Broadcast.var"#11#12"{Base.Broadcast.var"#13#14"}, ComposedFunction{Type{Float64}, typeof(-)}}, Base.Broadcast.var"#15#16"{Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#18"}}, Base.Broadcast.var"#11#12"{Base.Broadcast.var"#11#12"{Base.Broadcast.var"#13#14"}}, typeof(/)}, typeof(abs2)}, Tuple{SubArray{Float64, 1, Matrix{Float64}, Tuple{Base.Slice{Base.OneTo{Int64}}, Int64}, true}, TrackedArray{…,Vector{Float64}}}})(Δ::Vector{Float64})
      @ Tracker ~/.julia/packages/Tracker/xe6h6/src/lib/array.jl:603
   [18] back_(c::Tracker.Call{Tracker.var"#back#666"{2, Base.Broadcast.var"#2#4"{Base.Broadcast.var"#8#10"{Base.Broadcast.var"#1#3", Base.Broadcast.var"#8#10"{Base.Broadcast.var"#5#6"{Base.Broadcast.var"#7#9"}, Base.Broadcast.var"#5#6"{Base.Broadcast.var"#7#9"}, Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#18"}, Base.Broadcast.var"#11#12"{Base.Broadcast.var"#13#14"}, ComposedFunction{Type{Float64}, typeof(-)}}, Base.Broadcast.var"#15#16"{Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#18"}}, Base.Broadcast.var"#11#12"{Base.Broadcast.var"#11#12"{Base.Broadcast.var"#13#14"}}, typeof(/)}, typeof(abs2)}, Tuple{SubArray{Float64, 1, Matrix{Float64}, Tuple{Base.Slice{Base.OneTo{Int64}}, Int64}, true}, TrackedArray{…,Vector{Float64}}}}, Tuple{Nothing, Tracker.Tracked{Vector{Float64}}}}, Δ::Vector{Float64}, once::Bool)
      @ Tracker ~/.julia/packages/Tracker/xe6h6/src/back.jl:35
   [19] back(x::Tracker.Tracked{Vector{Float64}}, Δ::Vector{Float64}, once::Bool)
      @ Tracker ~/.julia/packages/Tracker/xe6h6/src/back.jl:58
   [20] #670
      @ ~/.julia/packages/Tracker/xe6h6/src/back.jl:38 [inlined]
   [21] foreach
      @ ./abstractarray.jl:2142 [inlined]
   [22] back_(c::Tracker.Call{Tracker.var"#591#592"{TrackedArray{…,Vector{Float64}}}, Tuple{Tracker.Tracked{Vector{Float64}}}}, Δ::Float64, once::Bool)
      @ Tracker ~/.julia/packages/Tracker/xe6h6/src/back.jl:38
  --- the last 4 lines are repeated 4 more times ---
   [39] back(x::Tracker.Tracked{Float64}, Δ::Int64, once::Bool)
      @ Tracker ~/.julia/packages/Tracker/xe6h6/src/back.jl:58
   [40] #back!#672
      @ ~/.julia/packages/Tracker/xe6h6/src/back.jl:77 [inlined]
   [41] #back!#13
      @ ~/.julia/packages/Tracker/xe6h6/src/lib/real.jl:18 [inlined]
   [42] back!(x::Tracker.TrackedReal{Float64})
      @ Tracker ~/.julia/packages/Tracker/xe6h6/src/lib/real.jl:16
   [43] gradient_(f::Function, xs::Vector{Float64})
      @ Tracker ~/.julia/packages/Tracker/xe6h6/src/back.jl:4
   [44] #gradient#681
      @ ~/.julia/packages/Tracker/xe6h6/src/back.jl:164 [inlined]
   [45] gradient
      @ ~/.julia/packages/Tracker/xe6h6/src/back.jl:164 [inlined]
   [46] test_ad(f::Function, x::Vector{Float64}, broken::Tuple{}; rtol::Float64, atol::Float64)
      @ Main ~/work/DistributionsAD.jl/DistributionsAD.jl/test/ad/utils.jl:357
   [47] test_ad(f::Function, x::Vector{Float64}, broken::Tuple{})
      @ Main ~/work/DistributionsAD.jl/DistributionsAD.jl/test/ad/utils.jl:351
   [48] test_ad(dist::DistSpec{DistributionsAD.VectorOfMultivariate{Continuous, ZeroMeanDiagNormal{Tuple{Base.OneTo{Int64}}}, Vector{ZeroMeanDiagNormal{Tuple{Base.OneTo{Int64}}}}}, var"#143#239"{var"#69#165"}, Tuple{Vector{Float64}}, Matrix{Float64}, Nothing, Tuple{}}; kwargs::Base.Iterators.Pairs{Union{}, Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
      @ Main ~/work/DistributionsAD.jl/DistributionsAD.jl/test/ad/utils.jl:343
   [49] test_ad(dist::DistSpec{DistributionsAD.VectorOfMultivariate{Continuous, ZeroMeanDiagNormal{Tuple{Base.OneTo{Int64}}}, Vector{ZeroMeanDiagNormal{Tuple{Base.OneTo{Int64}}}}}, var"#143#239"{var"#69#165"}, Tuple{Vector{Float64}}, Matrix{Float64}, Nothing, Tuple{}})
      @ Main ~/work/DistributionsAD.jl/DistributionsAD.jl/test/ad/utils.jl:284
   [50] macro expansion
      @ ~/work/DistributionsAD.jl/DistributionsAD.jl/test/ad/distributions.jl:591 [inlined]
   [51] macro expansion
      @ /buildworker/worker/package_linux64/build/usr/share/julia/stdlib/v1.6/Test/src/Test.jl:1151 [inlined]
   [52] macro expansion
      @ ~/work/DistributionsAD.jl/DistributionsAD.jl/test/ad/distributions.jl:472 [inlined]
   [53] macro expansion
      @ /buildworker/worker/package_linux64/build/usr/share/julia/stdlib/v1.6/Test/src/Test.jl:1151 [inlined]

I don't really understand the FillArrays issue from the stacktrace. Could you perhaps narrow it down to a MWE?

Some of the other runs were simply cancelled, perhaps after this run failed. In any case, perhaps we may test FillArrays PRs against certain dependent packages to avoid downstream breakages, so if you could suggest appropriate packages, that would be great.

jishnub · 2023-06-05T10:49:31Z

Hi, would you mind verifying if this is fixed with FillArrays v1.1.1?

devmotion · 2023-06-05T12:16:30Z

Unfortunately the issues persist even with FillArrays 1.1.1: https://github.com/TuringLang/DistributionsAD.jl/actions/runs/4992157009/jobs/9325223758 Unfortunately, I haven't had time to debug this further. Maybe there's another, similar issue still present in FillArrays?

torfjelde · 2023-06-16T21:45:39Z

From the trace above it seems like

FillArrays.jl/src/fillbroadcast.jl

Line 199 in 61a1cf9

for op in (:+, :-)

, i.e. adding proper suppport for -, is a likel candidate?

jishnub · 2023-06-17T07:36:06Z

You're likely right, the latest failure seems to point at the fixed line.

Matrixvariate distributions: Error During Test at /home/runner/work/DistributionsAD.jl/DistributionsAD.jl/test/ad/utils.jl:357
  Test threw exception
  Expression: ≈(Tracker.data((Tracker.gradient(f, x))[1]), finitediff, rtol = rtol, atol = atol)
  MethodError: no method matching Float64(::ForwardDiff.Dual{Nothing, Float64, 1})
  Closest candidates are:
    (::Type{T})(::Real, ::RoundingMode) where T<:AbstractFloat at rounding.jl:200
    (::Type{T})(::T) where T<:Number at boot.jl:760
    (::Type{T})(::AbstractChar) where T<:Union{AbstractChar, Number} at char.jl:50
    ...
  Stacktrace:
    [1] #8
      @ ./broadcast.jl:368 [inlined]
    [2] #8
      @ ./broadcast.jl:366 [inlined]
    [3] #2
      @ ./broadcast.jl:314 [inlined]
    [4] partial(::Base.Broadcast.var"#2#4"{Base.Broadcast.var"#8#10"{Base.Broadcast.var"#1#3", Base.Broadcast.var"#8#10"{Base.Broadcast.var"#5#6"{Base.Broadcast.var"#7#9"}, Base.Broadcast.var"#5#6"{Base.Broadcast.var"#7#9"}, Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#18"}, Base.Broadcast.var"#11#12"{Base.Broadcast.var"#13#14"}, DataType}, Base.Broadcast.var"#15#16"{Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#18"}}, Base.Broadcast.var"#11#12"{Base.Broadcast.var"#11#12"{Base.Broadcast.var"#13#14"}}, typeof(/)}, typeof(abs2)}, ::Float64, ::Int64, ::Float64, ::Float64)
      @ Tracker ~/.julia/packages/Tracker/xe6h6/src/lib/array.jl:596
    [5] _broadcast_getindex_evalf
      @ ./broadcast.jl:648 [inlined]
    [6] _broadcast_getindex
      @ ./broadcast.jl:621 [inlined]
    [7] getindex
      @ ./broadcast.jl:575 [inlined]
    [8] copy
      @ ./broadcast.jl:922 [inlined]
    [9] materialize(bc::Base.Broadcast.Broadcasted{Base.Broadcast.DefaultArrayStyle{1}, Nothing, typeof(Tracker.partial), Tuple{Base.RefValue{Base.Broadcast.var"#2#4"{Base.Broadcast.var"#8#10"{Base.Broadcast.var"#1#3", Base.Broadcast.var"#8#10"{Base.Broadcast.var"#5#6"{Base.Broadcast.var"#7#9"}, Base.Broadcast.var"#5#6"{Base.Broadcast.var"#7#9"}, Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#18"}, Base.Broadcast.var"#11#12"{Base.Broadcast.var"#13#14"}, DataType}, Base.Broadcast.var"#15#16"{Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#18"}}, Base.Broadcast.var"#11#12"{Base.Broadcast.var"#11#12"{Base.Broadcast.var"#13#14"}}, typeof(/)}, typeof(abs2)}}, Vector{Float64}, Int64, Vector{Float64}, Vector{Float64}}})
      @ Base.Broadcast ./broadcast.jl:883
[...]

devmotion · 2023-07-07T23:32:49Z

I finally managed to condense the remaining (or, as you'll see, actually new) issues to a MWE 🎉

The following example works as expected in FillArrays 1.0.0, 1.0.1, 1.0.2, and 1.1.0 (so it was not a problem when opening this issue):

julia> using FillArrays, ReverseDiff

julia> ReverseDiff.gradient(x -> sum(abs2.((Zeros(5) .- zeros(5)) ./ x)), rand(5))
5-element Vector{Float64}:
 0.0
 0.0
 0.0
 0.0
 0.0

julia> ReverseDiff.gradient(x -> sum(abs2.((zeros(5) .- Zeros(5)) ./ x)), rand(5))
5-element Vector{Float64}:
 0.0
 0.0
 0.0
 0.0
 0.0

However, with FillArrays >= 1.1.1:

julia> using FillArrays, ReverseDiff

julia> ReverseDiff.gradient(x -> sum(abs2.((Zeros(5) .- zeros(5)) ./ x)), rand(5))
5-element Vector{Float64}:
 0.0
 0.0
 0.0
 0.0
 0.0

julia> ReverseDiff.gradient(x -> sum(abs2.((zeros(5) .- Zeros(5)) ./ x)), rand(5))
ERROR: MethodError: no method matching Float64(::ForwardDiff.Dual{ForwardDiff.Tag{ReverseDiff.var"#109#111"{Base.Broadcast.var"#12#14"{Base.Broadcast.var"#18#20"{Base.Broadcast.var"#11#13", Base.Broadcast.var"#18#20"{Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#19"}, Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#19"}, Base.Broadcast.var"#25#26"{Base.Broadcast.var"#27#28"}, Base.Broadcast.var"#21#22"{Base.Broadcast.var"#23#24"}, DataType}, Base.Broadcast.var"#25#26"{Base.Broadcast.var"#25#26"{Base.Broadcast.var"#27#28"}}, Base.Broadcast.var"#21#22"{Base.Broadcast.var"#21#22"{Base.Broadcast.var"#23#24"}}, typeof(/)}, typeof(abs2)}, Tuple{}, Val{(1, 2)}}, Float64}, Float64, 2})

Closest candidates are:
  (::Type{T})(::Real, ::RoundingMode) where T<:AbstractFloat
   @ Base rounding.jl:207
  (::Type{T})(::T) where T<:Number
   @ Core boot.jl:792
  (::Type{T})(::AbstractChar) where T<:Union{AbstractChar, Number}
   @ Base char.jl:50
  ...

Stacktrace:
  [1] #18
    @ ./broadcast.jl:396 [inlined]
  [2] #18
    @ ./broadcast.jl:394 [inlined]
  [3] #12
    @ ./broadcast.jl:342 [inlined]
  [4] macro expansion
    @ ~/.julia/packages/ReverseDiff/wIfrd/src/derivatives/broadcast.jl:126 [inlined]
  [5] splatcall
    @ ~/.julia/packages/ReverseDiff/wIfrd/src/derivatives/broadcast.jl:111 [inlined]
  [6] (::ReverseDiff.var"#109#111"{Base.Broadcast.var"#12#14"{Base.Broadcast.var"#18#20"{Base.Broadcast.var"#11#13", Base.Broadcast.var"#18#20"{Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#19"}, Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#19"}, Base.Broadcast.var"#25#26"{Base.Broadcast.var"#27#28"}, Base.Broadcast.var"#21#22"{Base.Broadcast.var"#23#24"}, DataType}, Base.Broadcast.var"#25#26"{Base.Broadcast.var"#25#26"{Base.Broadcast.var"#27#28"}}, Base.Broadcast.var"#21#22"{Base.Broadcast.var"#21#22"{Base.Broadcast.var"#23#24"}}, typeof(/)}, typeof(abs2)}, Tuple{}, Val{(1, 2)}})(s::StaticArraysCore.SVector{2, ForwardDiff.Dual{ForwardDiff.Tag{ReverseDiff.var"#109#111"{Base.Broadcast.var"#12#14"{Base.Broadcast.var"#18#20"{Base.Broadcast.var"#11#13", Base.Broadcast.var"#18#20"{Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#19"}, Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#19"}, Base.Broadcast.var"#25#26"{Base.Broadcast.var"#27#28"}, Base.Broadcast.var"#21#22"{Base.Broadcast.var"#23#24"}, DataType}, Base.Broadcast.var"#25#26"{Base.Broadcast.var"#25#26"{Base.Broadcast.var"#27#28"}}, Base.Broadcast.var"#21#22"{Base.Broadcast.var"#21#22"{Base.Broadcast.var"#23#24"}}, typeof(/)}, typeof(abs2)}, Tuple{}, Val{(1, 2)}}, Float64}, Float64, 2}})
    @ ReverseDiff ~/.julia/packages/ReverseDiff/wIfrd/src/derivatives/broadcast.jl:150
  [7] static_dual_eval
    @ ~/.julia/packages/ForwardDiff/vXysl/ext/ForwardDiffStaticArraysExt.jl:24 [inlined]
  [8] vector_mode_gradient!(result::DiffResults.ImmutableDiffResult{1, Float64, Tuple{StaticArraysCore.SVector{2, Float64}}}, f::Function, x::StaticArraysCore.SVector{2, Float64})
    @ ForwardDiffStaticArraysExt ~/.julia/packages/ForwardDiff/vXysl/ext/ForwardDiffStaticArraysExt.jl:64
  [9] gradient!(result::DiffResults.ImmutableDiffResult{1, Float64, Tuple{StaticArraysCore.SVector{2, Float64}}}, f::Function, x::StaticArraysCore.SVector{2, Float64})
    @ ForwardDiffStaticArraysExt ~/.julia/packages/ForwardDiff/vXysl/ext/ForwardDiffStaticArraysExt.jl:44
 [10] (::ReverseDiff.var"#df#110"{Base.Broadcast.var"#12#14"{Base.Broadcast.var"#18#20"{Base.Broadcast.var"#11#13", Base.Broadcast.var"#18#20"{Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#19"}, Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#19"}, Base.Broadcast.var"#25#26"{Base.Broadcast.var"#27#28"}, Base.Broadcast.var"#21#22"{Base.Broadcast.var"#23#24"}, DataType}, Base.Broadcast.var"#25#26"{Base.Broadcast.var"#25#26"{Base.Broadcast.var"#27#28"}}, Base.Broadcast.var"#21#22"{Base.Broadcast.var"#21#22"{Base.Broadcast.var"#23#24"}}, typeof(/)}, typeof(abs2)}, DiffResults.ImmutableDiffResult{1, Float64, Tuple{StaticArraysCore.SVector{2, Float64}}}, Tuple{}, Val{(1, 2)}})(::Float64, ::Vararg{Float64})
    @ ReverseDiff ~/.julia/packages/ReverseDiff/wIfrd/src/derivatives/broadcast.jl:148
 [11] _broadcast_getindex_evalf
    @ ./broadcast.jl:683 [inlined]
 [12] _broadcast_getindex
    @ ./broadcast.jl:656 [inlined]
 [13] getindex
    @ ./broadcast.jl:610 [inlined]
 [14] copy
    @ ./broadcast.jl:912 [inlined]
 [15] materialize
    @ ./broadcast.jl:873 [inlined]
 [16] broadcast(::ReverseDiff.var"#df#110"{Base.Broadcast.var"#12#14"{Base.Broadcast.var"#18#20"{Base.Broadcast.var"#11#13", Base.Broadcast.var"#18#20"{Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#19"}, Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#19"}, Base.Broadcast.var"#25#26"{Base.Broadcast.var"#27#28"}, Base.Broadcast.var"#21#22"{Base.Broadcast.var"#23#24"}, DataType}, Base.Broadcast.var"#25#26"{Base.Broadcast.var"#25#26"{Base.Broadcast.var"#27#28"}}, Base.Broadcast.var"#21#22"{Base.Broadcast.var"#21#22"{Base.Broadcast.var"#23#24"}}, typeof(/)}, typeof(abs2)}, DiffResults.ImmutableDiffResult{1, Float64, Tuple{StaticArraysCore.SVector{2, Float64}}}, Tuple{}, Val{(1, 2)}}, ::Vector{Float64}, ::Vector{Float64})
    @ Base.Broadcast ./broadcast.jl:811
 [17] ∇broadcast(::Base.Broadcast.var"#12#14"{Base.Broadcast.var"#18#20"{Base.Broadcast.var"#11#13", Base.Broadcast.var"#18#20"{Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#19"}, Base.Broadcast.var"#15#16"{Base.Broadcast.var"#17#19"}, Base.Broadcast.var"#25#26"{Base.Broadcast.var"#27#28"}, Base.Broadcast.var"#21#22"{Base.Broadcast.var"#23#24"}, DataType}, Base.Broadcast.var"#25#26"{Base.Broadcast.var"#25#26"{Base.Broadcast.var"#27#28"}}, Base.Broadcast.var"#21#22"{Base.Broadcast.var"#21#22"{Base.Broadcast.var"#23#24"}}, typeof(/)}, typeof(abs2)}, ::Vector{Float64}, ::Vararg{Any})
    @ ReverseDiff ~/.julia/packages/ReverseDiff/wIfrd/src/derivatives/broadcast.jl:154
 [18] copy(_bc::Base.Broadcast.Broadcasted{ReverseDiff.TrackedStyle, Tuple{Base.OneTo{Int64}}, typeof(abs2), Tuple{Base.Broadcast.Broadcasted{ReverseDiff.TrackedStyle, Nothing, typeof(/), Tuple{Base.Broadcast.Broadcasted{Base.Broadcast.DefaultArrayStyle{1}, Nothing, Type{Float64}, Tuple{Vector{Float64}}}, ReverseDiff.TrackedArray{Float64, Float64, 1, Vector{Float64}, Vector{Float64}}}}}})
    @ ReverseDiff ~/.julia/packages/ReverseDiff/wIfrd/src/derivatives/broadcast.jl:94
 [19] materialize
    @ ./broadcast.jl:873 [inlined]
 [20] (::var"#7#8")(x::ReverseDiff.TrackedArray{Float64, Float64, 1, Vector{Float64}, Vector{Float64}})
    @ Main ./REPL[4]:1
 [21] ReverseDiff.GradientTape(f::var"#7#8", input::Vector{Float64}, cfg::ReverseDiff.GradientConfig{ReverseDiff.TrackedArray{Float64, Float64, 1, Vector{Float64}, Vector{Float64}}})
    @ ReverseDiff ~/.julia/packages/ReverseDiff/wIfrd/src/api/tape.jl:199
 [22] gradient(f::Function, input::Vector{Float64}, cfg::ReverseDiff.GradientConfig{ReverseDiff.TrackedArray{Float64, Float64, 1, Vector{Float64}, Vector{Float64}}})
    @ ReverseDiff ~/.julia/packages/ReverseDiff/wIfrd/src/api/gradients.jl:22
 [23] gradient(f::Function, input::Vector{Float64})
    @ ReverseDiff ~/.julia/packages/ReverseDiff/wIfrd/src/api/gradients.jl:22
 [24] top-level scope
    @ REPL[4]:1

So it seems this bug was introduced in FillArrays 1.1.1 (presumably #264) and therefore tests still failed when I checked with FillArrays 1.1.1 even though the initial issues were (seemingly) fixed (#252 (comment)).

devmotion · 2023-07-07T23:48:56Z

Hmm, seems these broadcasting specializations are a bit tricky for AD (and Tracker + ReverseDiff in particular). Changing

FillArrays.jl/src/fillbroadcast.jl

Line 208 in 9be3ba2

broadcasted(TT, a)

to _copy_oftype(a, TT) would fix the MWE but I'm not sure if it could be problematic in other cases.

devmotion · 2023-07-07T23:52:40Z

Oh, actually the problem might be promote_type. Using

TT = typeof($op(zero(T), zero(V)))

seems to fix it (and seems more correct if e.g. T = V = Bool).

devmotion mentioned this issue May 18, 2023

Add the Kumaraswamy distribution JuliaStats/Distributions.jl#1699

Merged

devmotion mentioned this issue Jul 4, 2023

CompatHelper: bump compat for FillArrays to 1, (keep existing compat) TuringLang/Turing.jl#2017

Closed

devmotion mentioned this issue Jul 7, 2023

Use return types of calculations instead of promote_type #273

Merged

dlfivefifty closed this as completed in #273 Jul 11, 2023

devmotion mentioned this issue Jul 11, 2023

Fix ReverseDiff bug #278

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Latest release broke DistributionsAD #252

Latest release broke DistributionsAD #252

devmotion commented May 18, 2023

jishnub commented May 19, 2023 •

edited

Loading

jishnub commented Jun 5, 2023

devmotion commented Jun 5, 2023

torfjelde commented Jun 16, 2023

jishnub commented Jun 17, 2023

devmotion commented Jul 7, 2023

devmotion commented Jul 7, 2023

devmotion commented Jul 7, 2023

Latest release broke DistributionsAD #252

Latest release broke DistributionsAD #252

Comments

devmotion commented May 18, 2023

jishnub commented May 19, 2023 • edited Loading

jishnub commented Jun 5, 2023

devmotion commented Jun 5, 2023

torfjelde commented Jun 16, 2023

jishnub commented Jun 17, 2023

devmotion commented Jul 7, 2023

devmotion commented Jul 7, 2023

devmotion commented Jul 7, 2023

jishnub commented May 19, 2023 •

edited

Loading