Support arrays with non-concrete eltype (fixes #264) - #300
Merged
Merged
Conversation
`istracked(::AbstractArray)` only promises that an array *may* hold tracked
elements: for a non-concrete eltype it cannot know. Two consumers mishandled
that, in opposite directions.
`increment_deriv!`/`decrement_deriv!` treated it as a promise that every
element *is* tracked and indexed straight into `increment_deriv!(t[i], ...)`,
so a `Vector{Real}` of plain `Float64`s threw `MethodError`. They now skip
untracked elements. A direct `increment_deriv!(::Float64, ::Float64)` still
throws, so a mis-wired call elsewhere stays loud.
`_add_to_deriv!` went the other way and matched only
`AbstractArray{<:TrackedReal}`, which a `Vector{Real}` does not satisfy, so
every `@grad` rule fell through to the `nothing` fallback and *silently*
discarded the derivative:
f(x) = sum([Real[x[1], x[2]]; x[3]])
ForwardDiff.gradient(f, [1.0, 2.0, 3.0]) # [1.0, 1.0, 1.0]
ReverseDiff.gradient(f, [1.0, 2.0, 3.0]) # [0.0, 0.0, 1.0]
It now asks `istracked` directly. This is the same bound widened once more:
50b47bf took it from `TrackedArray` to `AbstractArray{<:TrackedReal}`, which
incidentally fixed the MWE of #145 without closing it; the non-concrete
eltype case is what remained.
Finally, `istracked(::AbstractArray{T})` used `!isconcretetype(T)` as a proxy
for "could an element be tracked". That is wrong in both directions: it
admits `AbstractString`, `Integer` and `Function` eltypes, and rejects
`Vector{TrackedArray}`, for which `tape` returned `NULL_TAPE` even though the
elements carry a tape. Asking whether a tracked type can inhabit `T` is exact
and folds to a compile-time constant.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
devmotion
force-pushed
the
dmw/nonconcrete-eltype-arrays
branch
from
September 18, 2026 13:35
57a7505 to
3a75bac
Compare
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## master #300 +/- ##
==========================================
+ Coverage 86.69% 86.92% +0.23%
==========================================
Files 19 19
Lines 1916 1950 +34
==========================================
+ Hits 1661 1695 +34
Misses 255 255 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
Pullbacks may return their tangents lazily, and ChainRules' own rules
routinely do, but nothing in the test suite exercised that path: no rrule
returned a thunk. Cover both call sites of `_add_to_deriv!`, both concrete
`AbstractThunk` subtypes, and both legs of its `Union{TrackedReal,
AbstractArray}` signature.
The `g_thunk` cases also pin the `istracked` guard added here: thunks aimed
at an untracked argument are dropped rather than forced, and a repeated
tracked element in a `Vector{Real}` accumulates.
Without the `AbstractThunk` method every new assertion fails with a
`MethodError` from `increment_deriv!`, which only accepts arrays and reals.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
devmotion
added a commit
that referenced
this pull request
Sep 22, 2026
…, and pin #126 (#302) * Test indexing a zero-dimensional `TrackedArray` (#126) `gradient(x -> x[], fill(1.0))` threw a `MethodError` for an ambiguity between the `AbstractRange...` and `Colon...` methods of `getindex`: a bare vararg matches zero arguments, so `t[]` matched both equally. 4a50321 fixed it incidentally while fixing #202, by rewriting every `getindex(t::TrackedArray, i::T...)` as `i1::T, is::T...`. Requiring a leading index sends `t[]` down `Base`'s generic zero-index `AbstractArray` path, which lands on the `Integer` method via linear indexing -- so, like any scalar `getindex`, it aliases its parent through the origin and records nothing. Nothing pinned that, and the ambiguity would come back unnoticed with any future widening of those signatures. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * Give tracked reals with a free origin a `zero` and a `one` (fixes #172) `map`ping over a tracked array mixes elements that carry an origin with ones that do not, and the resulting eltype leaves the origin free. `reducedim_init` allocates a reduction's accumulator by asking that eltype for its `zero`, or `one` for `prod`, which fell through to `convert(T, 0)` and then to `valtype`, defined only for a pinned origin: MethodError: no method matching valtype(::Type{TrackedReal{Float64, Float64}}) Neither identity depends on the origin, so both are deducible from the type alone, as `zero` requires. `Nothing` is the origin `zero` has always used, and `TrackedReal{V,D,Nothing}` is an instance of `TrackedReal{V,D}`, so the documented `zero(T) isa T` holds. The signature matches that type exactly rather than its subtypes: a `Union` of two origins is also a subtype, and no single member of it could be returned. The `zero` method is the one #173 added and #177 reverted wholesale over #175, whose cause -- `increment_deriv!` indexing into untracked elements -- was fixed by #300. `one` is its counterpart, which #173 never added. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This was referenced Sep 22, 2026
devmotion
added a commit
that referenced
this pull request
Sep 24, 2026
`Diagonal` and `diagm` of a `Vector{Real}` give tracked diagonal entries
and untracked zeros, whose reverse pass threw `increment_deriv!(::Float64,
::Float64)` until #300 made it skip untracked elements. Check against the
analytical Jacobians, directly and on a compiled tape.
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.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
1 participant
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.
Fixes #264.
istracked(::AbstractArray)only promises that an array may hold tracked elements. Two consumers mishandled that, in opposite directions.increment_deriv!/decrement_deriv!assumed every element is tracked, so aVector{Real}of plainFloat64s threwMethodError: increment_deriv!(::Float64, ::Float64). They now skip untracked elements; a direct call with an untracked scalar still throws._add_to_deriv!matched onlyAbstractArray{<:TrackedReal}, so every@gradrule fell through to thenothingfallback and silently discarded the derivative:It now asks
istrackeddirectly. Same bound as 50b47bf widened once more — that commit incidentally fixed the MWE of #145 (bisected: broken in v1.5.0, fixed in v1.6.0) without closing it, and this covers what remained.Finally,
istracked(::AbstractArray{T})used!isconcretetype(T)as a proxy for "could an element be tracked". Asking whether a tracked type can inhabitTis exact and folds to a compile-time constant. One behaviour change:istracked(::Vector{TrackedArray})goesfalse→true, sovaluenow unwraps it andtapefinds its tape.🤖 Generated with Claude Code