Repository navigation
v0.7.16
Zygote v0.7.16
Merged pull requests:
- Bump JuliaRegistries/TagBot from 1.25.13 to 1.25.14 (#1683) (@dependabot[bot])
- Make the
repeat(x; inner, outer)adjoint a reshape and sum (#1703) (@CarloLucibello) - Avoid large temporaries in the pinv and Diagonal solve adjoints (#1704) (@CarloLucibello)
- Use literal_pow in the broadcast x.^p adjoint so the backward vectorises (#1705) (@CarloLucibello)
- Remove host round trips in the tr-adjoint accumulation and ProjectTo(::CuSparseMatrixCSC) (#1707) (@CarloLucibello)
- Unroll tuple and n-ary accum (quadratic backward over long tuples) (#1708) (@CarloLucibello)
- Add mapreduce/reduce rules for + and * over arrays (#1709) (@CarloLucibello)
- Broaden the precompile workload (#1710) (@CarloLucibello)
- Route vcat/hcat splats of a vector of arrays to the reduce rule (#1711) (@CarloLucibello)
- Project sparse cotangents onto CuSparseMatrixCSC without densifying (#1712) (@CarloLucibello)
- Accumulate tuple cotangents per position and skip nothing (#1713) (@CarloLucibello)
- Reuse the factorization in the dense left-division adjoint (#1714) (@CarloLucibello)
- Defer cotangents of scalar broadcast operands (#1719) (@CarloLucibello)
- Fix splat cotangent accumulation and first-call time (#1716, #1717) (#1723) (@CarloLucibello)
- Differentiate GPU map and captured variables of GPU broadcasts (#1724) (@CarloLucibello)
- Unthunk implicit
Gradsentries when the pullback returns (#1726) (@CarloLucibello) - Cache pullback decompositions and remove quadratic block scans in the transform (#1728) (@CarloLucibello)
- Note in AGENTS.md that runtests.jl runs only
cudaon GPU machines (#1729) (@CarloLucibello) - Send unary broadcasts over a
Fillto the genericFillrule (#1730) (@CarloLucibello) - Recompute the
Duals in the pullback of GPU broadcasts (#1731) (@CarloLucibello) - Add a broadcast rule for
x .^ pwith a runtime exponent (#1732) (@CarloLucibello) - Add broadcast rules for
exp,log,sqrtandinv(#1733) (@CarloLucibello) - Fix the Julia 1.10 test failures on master (#1734) (@CarloLucibello)
- Type the pullbacks of code with branches unless the type gets large (#1735) (@CarloLucibello)
- Type the stacks of code with loops after the values they record (#1736) (@CarloLucibello)
- Differentiate
foldl/accumulatewithinit,dimsand on the GPU (#1737) (@CarloLucibello) - Accumulate cotangents in place in the reverse pass (#365) (#1738) (@CarloLucibello)
Closed issues:
- Wrong gradient results for multiple
getindexchained byfoldl(#1558) - Triage 15 June 2026 (#1631)
- accumulate(...; init=x) silently drops the gradient of x (#1684)
- First gradient of
vcat(xs...)/hcat(xs...)with a vector splat takes seconds and grows with the number of arguments (#1687) - Ambiguity
MethodErrorwhen broadcasting-,identity,tanh,conj,real,imag,abs2,xlogx,log1pexpover aFill(#1688) repeat(x; inner, outer)adjoint is a scalar loop: ~140x slower on CPU, errors on CuArray (#1689)- Broaden the precompile workload to cut time-to-first-gradient (#1690)
mapreduce/reducewith+/*over arrays has a quadratic gradient; add rules (#1691)- Broadcast
x .^ pbackward uses a runtime-exponent^and is not vectorised (#1692) - Broadcast adjoint eagerly computes cotangents for constant and data arguments (extra kernel, reduction and D2H sync) (#1694)
- Broadcast Dual path: retained Dual array and extra kernels; cheaper alternatives and dedicated
sqrt/inv/exp/logrules (#1696) - Pullbacks of functions with branches or loops are
Pullback{F,Any}(type-unstable, allocating) (#1697) - Pullback generation: redundant IR transforms and compile time superlinear in basic blocks (#1698)
- Linear-algebra adjoints:
pinvbuilds m×m temporaries,Diagonal \and/densify,\refactorizes (#1699) - CUDA: host round trips in
tr/Diagonalaccumulation and inProjectTo(::CuSparseMatrixCSC)(#1700) - First gradient of
vcat(xs...)with a vector of scalars or a mixed splat takes seconds and grows with N (follow-up to #1687) (#1716) MethodErroraccumulating the cotangent of a splatted vector invcat(a, vs..., b)whenvsis also used elsewhere (#1717)- Implicit
Gradsstore raw thunks:pairs(gs)returns a thunk andgs[w]recomputes it (#1721)