Hacktoberfest 2026: the issues maintainers tagged for October, open and beginner-friendly. Browse Hacktoberfest issues

`@grad`/`@grad_from_chainrules` fail for rules returning structured arrays (`Diagonal`, `UpperTriangular`, ...)

Open
  • Is something wrong?
#321 0 comments 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 1 day

Nobody has claimed this yet.

Assessment

Difficulty
4/5
Estimated time
3-5 days
Newbie friendliness
56/100
Issue type
Bug
Clarity
Clearly specified
Activity status
Active
Tech stack
julia
Domain
devtools

Research direction

Start with the two tracking paths in src/macros.jl named in the issue: @grad and @grad_from_chainrules. Reproduce the Diagonal and UpperTriangular examples, then trace how output tracking and forward replay handle structured arrays. Done means both examples complete without the IndexStyle assertion and return the expected gradients.

Written by the indexing model from the issue text.

Description

If a rule's primal output is an array without linear indexing, such as Diagonal or UpperTriangular, ReverseDiff fails when it wraps the output in a TrackedArray. This affects both @grad_from_chainrules and @grad.

using ReverseDiff, ChainRulesCore, LinearAlgebra

f(x) = Diagonal(x)
ChainRulesCore.rrule(::typeof(f), x) = f(x), Δ -> (NoTangent(), diag(unthunk(Δ)))
ReverseDiff.@grad_from_chainrules f(x::ReverseDiff.TrackedArray)
ReverseDiff.gradient(x -> sum(f(x)), [1.0, 2.0])

ERROR: LoadError: AssertionError: IndexStyle(value) === IndexLinear()
Stacktrace:
  [1] ReverseDiff.TrackedArray{Float64, Float64, 2, Diagonal{Float64, Vector{Float64}}, Diagonal{Float64, Vector{Float64}}}(value::Diagonal{Float64, Vector{Float64}}, deriv::Diagonal{Float64, Vector{Float64}}, tape::ReverseDiff.InstructionTape)
  ...
  [5] track(::typeof(f), x::ReverseDiff.TrackedArray{Float64, Float64, 1, Vector{Float64}, Vector{Float64}})
    @ Main ~/.julia/dev/ReverseDiff/src/macros.jl:339

@grad with an UpperTriangular output fails the same way:

g(x) = UpperTriangular(x)
g(x::ReverseDiff.TrackedArray) = ReverseDiff.track(g, x)
ReverseDiff.@grad function g(x)
    return UpperTriangular(ReverseDiff.value(x)), Δ -> (triu(Δ),)
end
ReverseDiff.gradient(x -> sum(g(x)), [1.0 2.0; 3.0 4.0])
# ERROR: AssertionError: IndexStyle(value) === IndexLinear()

Expected: [1.0, 1.0] and [1.0 1.0; 0.0 1.0]. Many ChainRules rules return structured matrices (Diagonal, Symmetric, triangular factors, ...), so importing them fails as soon as the output is tracked.

Cause: the macros call track(output_value, tp) on the primal output as is (@grad, @grad_from_chainrules). TrackedArray requires IndexLinear() storage, and even with #216 a Diagonal/UpperTriangular deriv couldn't store the off-structure entries the reverse pass seeds. Materializing such outputs (e.g. collect) before tracking would avoid this. The forward replay value!(output, out_value) has to do the same.

ReverseDiff master (b796032, v1.18.4), Julia 1.13.1.

Dominant language
Julia
Stars
396
Forks
61
Avg merge
23h 18m
Merged PRs (30d)
16

Getting set up

This project ships no dev container, Dockerfile or contributing guide, so setting up is up to you: start from its README, and see our first-contribution guide for the general steps.

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from JuliaDiff/ReverseDiff.jl

All issues in JuliaDiff/ReverseDiff.jl

Similar issues

More Julia issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.