Fix Tracker gradient on RAT v4 by preserving the ODESolution wrapper#3665
Merged
ChrisRackauckas merged 2 commits intoMay 21, 2026
Merged
Conversation
In RecursiveArrayTools v4 `AbstractVectorOfArray <: AbstractArray`, so the `sol isa AbstractArray` branch in `Tracker.@Grad function DiffEqBase.solve_up` now matches ODESolution and returns the nested `sol.u :: Vector{Vector{Float64}}` directly. Tracker tracks the vector-of-vectors, and downstream `sum(solve(...))` reduces the outer vector element-wise into a `Vector{Float64}`, breaking `Tracker.gradient(loss, p)` callers with "Function output is not scalar". Return the ODESolution wrapper itself for `AbstractVectorOfArray` inputs so callers reduce through the RAT v4 AbstractArray interface and get a scalar as before. Earlier attempt (SciML#3663) stacked into a fresh matrix via `Array(sol)`; that was reverted in SciML#3664 because it changed the return type and broke downstream consumers. Preserving the wrapper keeps the contract. Refs SciML/SciMLSensitivity.jl#1331. Co-Authored-By: Chris Rackauckas <accounts@chrisrackauckas.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Please ignore until reviewed by @ChrisRackauckas.
Summary
Re-take on the Tracker + RAT v4 fix after the original #3663 was reverted in #3664. Same root cause, different correct shape.
In RecursiveArrayTools v4,
AbstractVectorOfArray <: AbstractArray, so thesol isa AbstractArraybranch inTracker.@grad function DiffEqBase.solve_upnow matches ODESolution and returns the nestedsol.u :: Vector{Vector{Float64}}directly. Tracker tracks the vector-of-vectors, and downstreamsum(solve(...))reduces the outer vector element-wise into aVector{Float64}, breakingTracker.gradient(loss, p)callers with"Function output is not scalar".The previous fix (#3663) stacked the data into a fresh
Matrix{Float64}viaArray(sol). That changed the return type ofsolve(...)from a wrapper to a matrix and broke downstream consumers — hence the revert in #3664. This PR returns the ODESolution wrapper itself forAbstractVectorOfArrayinputs, so callers reduce through the RAT v4 AbstractArray interface and produce a scalar without losing the solution type.Verification
Local on Julia 1.12.6 + RAT 4.3.0 + Tracker 0.2.38 + SciMLSensitivity master, with DiffEqBase 7.5.1 from this branch:
sum(solve(remake(prob; p), Tsit5()))sum(solve(...; save_idxs = [1]))sum(solve(...; save_everystep = false))sum(solve(...; saveat = 2.3))The five
@test_broken falseguards intest/concrete_solve_derivatives.jl(lines 518, 548, 579, 611, 735) on SciMLSensitivity are slated for removal in SciML/SciMLSensitivity.jl#1452, gated on this PR landing.Refs
@test_brokenguardsTest plan
solve_up @gradupdated to explain RAT v4 semantics