[EvoArt]

Add Piste.jl as a backend, in forward and reverse mode, following @gdalle's "PR welcome" on Slack.

Piste is small and written so that its derivatives compile under juliac --trim=safe. That still holds through DI. I compiled small apps that call DI with AutoPiste using JuliaC (Julia 1.12.7, --trim=safe). Every one built with zero verifier errors, ran with codegen disabled (JULIA_LOAD_CODEGEN_LIB=0), and printed the same output as the JIT:

Through DI, prepared	Forward	Reverse
gradient	✅ (explicit chunksize)	✅
derivative, pushforward	✅
pullback, jacobian		✅
gradient with Constant / Cache contexts	✅	✅
jacobian	❌ see #1088

The binary for a prepared DI gradient is 2.5 MB, against 2.4 MB for the same program calling Piste directly. One app exercising all the ✅ operators above comes to 2.8 MB.

The one ❌ is not specific to Piste. BatchSizeSettings(B::Int, N::Int) computes its singlebatch/aligned type parameters from the runtime N, so any forward backend's jacobian prep has an unknown type under --trim. I've opened #1088 with a proposed fix and kept it out of this PR.

Usage

using DifferentiationInterface, Piste
import ADTypes

f(x) = sum(abs2, x)
prep = prepare_gradient(f, AutoPiste(; chunksize = 4), x)  # forward, dual numbers
prep = prepare_gradient(f, AutoPiste(; mode = ADTypes.ReverseMode()), x)  # reverse, tape

AutoPiste itself is added in SciML/ADTypes.jl#191.

What the extension implements

• Forward (Piste.Dual{B,T}, untagged): one- and two-argument pushforward with batches of tangents. gradient for Vector{<:AbstractFloat} inputs goes straight to Piste's chunked driver with a reusable GradientWorkspace. Other inputs fall back to DI's batched pushforwards. pick_batchsize and threshold_batchsize mirror the ForwardDiff extension. With chunksize = nothing the width is picked from length(x) at prep time, which works fine in the JIT, but trimmed code should pass a fixed chunksize.
• Reverse (Piste.TapedReal on a tape): one- and two-argument pullback, recording once per cotangent. gradient is a single pullback reusing a ReverseWorkspace. Two-argument functions are recorded through a tracked copy of y, initialized to constant zeros. This relies on a new public rev_value_and_pullback(!) in Piste 0.4.
• Contexts: Constant and Cache in both modes. Caches are rebuilt with the active number type through recursive_similar.
• inplace_support is InPlaceSupported().

Not supported: second order. Piste's Duals are untagged, and nesting Dual over TapedReal isn't implemented, so SECOND_ORDER scenarios are excluded from the tests. The docs table shows ❌ for those operators. DI's own fallbacks still work for first-order operators the extension doesn't define, such as forward jacobian and reverse derivative.

Tests

test/Back/Piste runs default_scenarios (including the constantified, cachified, constant-or-cachified, tuple and smaller variants) and sparse_scenarios on AutoPiste(), AutoPiste(; chunksize = 2) and AutoPiste(; mode = ReverseMode()), plus test_counterparts. Locally (Julia 1.12.7), 76,419 tests pass and none fail or error. I've added the Piste group to the CI matrix.

Running the suite turned up one Piste bug: eltype(::Type{Dual}) returned the float type, which broke broadcasting with a scalar dual. It's fixed in Piste 0.4.

Order of merging

1. Piste 0.4.0 (adds the pullback API and the eltype fix). Registration is pending.
2. Add AutoPiste SciML/ADTypes.jl#191, then an ADTypes release.
3. This PR. Its compat bounds (Piste = "0.4", ADTypes = "1.25.0") assume both releases, so CI for the Piste group will fail until then. Happy to leave this as a draft until they land, and to adjust the ADTypes bound if the release gets a different number.

The docs (backends table, a short "Piste" section), the README list and the CHANGELOG are updated.

🤖 Generated with Claude Code