Releases: EnzymeAD/Reactant.jl
Releases · EnzymeAD/Reactant.jl
v0.2.171
Reactant v0.2.171
Merged pull requests:
- Make
RArraya subtype ofDenseArray(#1696) (@albertomercurio) - feat: reverse mode batched AD (#1740) (@avik-pal)
- Fix jl-generators for AttrSizedOperandSegments string (#1741) (@sbrantq)
- Format Julia code of branch "main" (#1742) (@github-actions[bot])
- Regenerate MLIR Bindings (#1743) (@github-actions[bot])
- feat: print bytes using format_bytes (#1744) (@avik-pal)
- feat: implement zeros/ones/fill (#1746) (@avik-pal)
- test: fix failing TPU test for oneelement (#1747) (@avik-pal)
- feat: new optimization passes + new jll (#1748) (@avik-pal)
- Regenerate MLIR Bindings (#1750) (@github-actions[bot])
- docs: fix has_tpu (#1752) (@avik-pal)
- fix: loading reactant on 1.12 (#1755) (@avik-pal)
- fix: disable 1.12 precompile (#1756) (@avik-pal)
Closed issues:
- Int32 Array compiling got AssertionError: ElType != Any && ElType != Union{} (#470)
- Bad XLA codegen for exact TopK (#886)
- Cannot compile function with more than 32 args (#922)
- Argmax fails with reactant (#1079)
- Add
Aquatests (#1201) - Alternate implementation of
BatchDuplicated(#1295) - GB-25 code OOM'ing on GitHub-hosted runners while compiling code (#1551)
- Ambigous method for fill! (#1601)
- Print the memory info in Stats.jl using
Base.format_bytes(#1646) - Make
Reactant.RArraya subtype ofDenseVector(#1695) - Incorrect results for nested AD (#1733)
- Forward mode AD fails on vector-output functions (#1749)
- Reactant.has_tpu is not defined (#1751)
v0.2.170
Reactant v0.2.170
Merged pull requests:
- ProbProg: JLL changes for trace/symbol/constraint types (#1719) (@sbrantq)
- feat: updates for new JLL (#1721) (@avik-pal)
- add Colon dispatch and tests (#1725) (@yolhan83)
- feat(JLL): expose cuda cc functions (#1726) (@avik-pal)
- Format Julia code of branch "main" (#1728) (@github-actions[bot])
- Regenerate MLIR Bindings (#1730) (@github-actions[bot])
- feat: c-api auto-wrap for EnzymeXLA (#1732) (@avik-pal)
- feat: support for fold/unfold (#1735) (@avik-pal)
- feat: new compiler optimizations (#1737) (@avik-pal)
- ci: use 1.11 and prevent python 3.14 for now (#1738) (@avik-pal)
Closed issues:
- Support for unfold (and fold) with Lux.jl & Reactant.jl (#1731)
v0.2.169
Reactant v0.2.169
Merged pull requests:
- Add support for log2 and log10 (#1714) (@bartvanerp)
- Format Julia code of branch "main" (#1715) (@github-actions[bot])
- fix: add additional check in ignore_derivatives (#1717) (@avik-pal)
Closed issues:
- No method matching
log10(::Reactant.TracedRNumber{Float32})(#1713)
v0.2.168
v0.2.167
v0.2.166
v0.2.165
Reactant v0.2.165
Merged pull requests:
- Bump actions/checkout from 4 to 5 (#1564) (@dependabot[bot])
- feat: overlay Zygote.gradient and use Enzyme instead (#1658) (@avik-pal)
- Regenerate MLIR Bindings (#1688) (@github-actions[bot])
- adpt to activity changes (#1691) (@wsmoses)
- Update sed command to modify multiple files (#1692) (@wsmoses)
Closed issues:
@tracefunction definition (#1683)
v0.2.164
Reactant v0.2.164
Merged pull requests:
- Format Julia code of branch "main" (#1679) (@github-actions[bot])
- Add section about loops in AD tutorial (#1680) (@Pangoraw)
- docs: cross reference lux docs on training neural networks (#1682) (@avik-pal)
- feat: new upstream opt passes (#1684) (@avik-pal)
- feat:
@traceon function definition (#1685) (@avik-pal)
Closed issues:
v0.2.163
v0.2.162
Reactant v0.2.162
Merged pull requests:
- fix: overlay searchsorted functions (#1596) (@avik-pal)
- feat: scatter add gradient (#1609) (@avik-pal)
- Add partial evaluation tutorial (#1624) (@oschulz)
- feat: ensure i32 indexing based on input sizes (#1630) (@avik-pal)
- ci: update the continuous benchmarking ci to use the new machines (#1632) (@avik-pal)
- fix: specialize on abstractmatrix and abstractvector for repeat outer (#1633) (@avik-pal)
- fix: force call with reactant if using inferencebarrier (#1635) (@avik-pal)
- refactor: standardize use of
promote_toandbroadcast_to_size(#1637) (@avik-pal) - feat: bump jll (#1638) (@avik-pal)
- fix: set-indexing into subarray (#1639) (@avik-pal)
- fix: special case cartesian indexing of reshaped arrays (#1640) (@avik-pal)
- chore: exclude auto-generated files from appearing in coverage reports (#1641) (@avik-pal)
- feat: use
traced_callwhen unrolling iterators and generators (#1642) (@avik-pal) - [ReactantExtra] Adapt WORKSPACE to jax-ml/jax@006b290 (#1644) (@giordano)
- Don't initialize gpu backend if no relevant jll (#1647) (@wsmoses)
- Regenerate MLIR Bindings (#1648) (@github-actions[bot])
- doc: fix docs build (#1650) (@avik-pal)
- docs: autodiff tutorial (#1651) (@avik-pal)
- docs: should have been marked
@example(#1655) (@avik-pal) - fix: clamp! dispatch (#1661) (@avik-pal)
- feat: support generators + force calling into interp (#1665) (@avik-pal)
- test: code quality testing (#1668) (@avik-pal)
- Regenerate MLIR Bindings (#1671) (@github-actions[bot])
Closed issues:
- HLO spmd-partitioning failure (#1010)
@compilegradient of NNlib.scatter! with Enzyme fails (#1423)- Overlayed functions not detected inside another overlayed function (#1589)
- Weird indexing error with similar + fill! (#1602)
- Error in forward run with Oceananigans (#1629)
- Wrong number of indices when broadcasting reshaped array (#1634)
- Unable to copy to array slice (#1636)
- Reactant's traced scalar types causing issues with Optimisers.jl's ClipNorm() (#1645)
- Trouble building Reactant with GPU (CUDA) support on remote cluster (#1653)