v0.2.31

Reactant v0.2.31

Diff since v0.2.30

Merged pull requests:

• Bump actions/checkout from 3 to 4 (#762) (@dependabot[bot])

v0.2.30

Reactant v0.2.30

Diff since v0.2.29

Merged pull requests:

• feat: use parameter shardings from XLA (#743) (@avik-pal)
• feat: JLL changes to expose HloModule (#749) (@avik-pal)
• [IFRT] add ifrt-proxy server and client bindings (#750) (@mofeing)
• fix: ordering of arguments need to be according to device (#753) (@avik-pal)
• Support tracing of rem with only one operand being a ConcreteRNumber (#754) (@giordano)
• Fix for ocean (#756) (@wsmoses)
• Bump to 0.2.30 (#757) (@glwagner)
• Fix implementation of mod (#758) (@giordano)
• [ReactantCUDAExt] Remove extra method (#760) (@giordano)

Closed issues:

• mod is JIT-ed to HLO operator with the semantic of Julia's rem (#755)
• Method Overwritting in CUDAExt (#759)

v0.2.29

Reactant v0.2.29

Diff since v0.2.28

Merged pull requests:

• Format code of branch "main" (#729) (@github-actions[bot])
• fix: prevent method ambiguity for CartesianIndex{1} (#730) (@avik-pal)
• [GHA] Some improvement to CI setup (#731) (@giordano)
• fix: improve generated mlir for wrapped arrays (#732) (@avik-pal)
• fix Type(value) instead of type(value) (#733) (@jumerckx)
• fix: don't expand all ranges by default (#737) (@avik-pal)
• ci: add cpp format check (#739) (@avik-pal)
• feat: sharding via IFRT (#740) (@avik-pal)
• fix: unqualified Sharding access (#741) (@avik-pal)
• Force tracing of type to act as noop (#747) (@wsmoses)
• Support for dicts (#748) (@wsmoses)

Closed issues:

• Types as struct fields (#745)
• Dictionaries with to_rarray (#746)

v0.2.28

Reactant v0.2.28

Diff since v0.2.27

Merged pull requests:

• feat: add sign dispatches (#727) (@avik-pal)
• fix: correct dims handling in mapreducedim! (#728) (@avik-pal)

v0.2.27

Reactant v0.2.27

Diff since v0.2.26

Merged pull requests:

• Format code of branch "main" (#711) (@github-actions[bot])
• feat: overload ifelse for more types (#712) (@avik-pal)
• fix: multi-device execution and sharding [take III] (#713) (@avik-pal)
• Replace capture maps with Holded wrapper (#715) (@mofeing)
• refactor: split XLA.jl into multiple files (#716) (@avik-pal)
• feat: enable async on CPU (#717) (@avik-pal)
• [ReactantExtra] IFRT bindings (round 4) (#718) (@mofeing)
• [ReactantExtra] feat: OpSharding bindings for Julia (#721) (@avik-pal)
• [ReactantExtra] fix: build on mac (#722) (@avik-pal)
• Update WORKSPACE (#723) (@avik-pal)
• Fix jll (#724) (@wsmoses)

Closed issues:

• shardy functions not visible on macos (#714)

v0.2.26

Reactant v0.2.26

Diff since v0.2.25

Merged pull requests:

• @trace function calls (#366) (@jumerckx)
• chore: missing upstream optimization passes (#624) (@avik-pal)
• feat: shardy and multi device execution (#637) (@avik-pal)
• Regenerate MLIR Bindings (#686) (@github-actions[bot])
• Misc fixes (#687) (@wsmoses)
• dict value fix (#688) (@wsmoses)
• [deps] Some improvements to the build_local.jl script (#689) (@giordano)
• Multiple device error (#690) (@wsmoses)
• feat: API changes for multi-device execution [ReactantExtra JLL changes] (#692) (@avik-pal)
• Wrapping RCReferences (#697) (@hhkit)
• Ref ptr fix (#698) (@wsmoses)
• Add GPUCompiler and LLVM as deps to CUDA extension and run CUDA tests on macOS (#700) (@giordano)
• vendor optimize (#703) (@wsmoses)
• [ReactantExtra] Stop removing references to hardware_interference_size (#704) (@giordano)
• Update Project.toml (#705) (@wsmoses)
• JLL related fixups (#706) (@wsmoses)
• Regenerate MLIR Bindings (#708) (@github-actions[bot])
• Format code of branch "main" (#709) (@github-actions[bot])
• fix: don't trace val (#710) (@avik-pal)

Closed issues:

• @trace function_call() to introduce function barrier (#346)
• Is there any practical benefit of tracing Val? (#602)
• Missing adjoint of stablehlo.gather (#676)
• Segfault on convert_simplify optimization with complex numbers (#695)
• Integration with SpeedyWeather.jl (#696)
• Integration with NFFT.jl (#699)

v0.2.25

Reactant v0.2.25

Diff since v0.2.24

Merged pull requests:

• make similar return empty tensors. (#632) (@jumerckx)
• Use LLVMOpenMP_jll to call OpenMP functions (#673) (@giordano)
• [ReactantCUDAExt] Skip precompile load on Julia v1.11.3 (#675) (@giordano)
• Regenerate MLIR Bindings (#680) (@github-actions[bot])
• [ReactantExtra] Add argument to ClientCompile to pass CUDA data dir (#683) (@giordano)
• CUDA: fix gc issues (#685) (@wsmoses)

Closed issues:

• We aren't actually using the ptxas and libdevice.bc shipped in the CUDA packages (#663)
• Segmentation faults on aarch64-linux starting from introduction of extension of KernelAbstractions (#677)

v0.2.24

Reactant v0.2.24

Diff since v0.2.23

Merged pull requests:

• [IFRT] Compile error hotfix (#641) (@hhkit)
• KA without cuda backend (#670) (@wsmoses)

v0.2.23

Reactant v0.2.23

Diff since v0.2.22

Merged pull requests:

• Regenerate MLIR Bindings (#627) (@github-actions[bot])
• [CI] Add workflow to clean up docs previews (#628) (@giordano)
• fix: build error with shardy (#629) (@avik-pal)
• [ReactantExtra] Improvements to BUILD file to compile CUDA for aarch64 (#631) (@giordano)
• fix cuda abi setting (#633) (@wsmoses)
• Format code of branch "main" (#634) (@github-actions[bot])
• [tests] Replace random custom type numbers with fixed set of numbers (#636) (@giordano)
• Add IR dumping (#638) (@wsmoses)
• [ReactantExtra] Bump XLA version (#640) (@giordano)
• TPU profiler (#642) (@Pangoraw)
• Applehw (#643) (@wsmoses)
• Regenerate MLIR Bindings (#644) (@github-actions[bot])
• feat: add dispatch for KA get_backend (#645) (@avik-pal)
• Use xla/stream_executor/cuda:cuda_compute_capability_proto_cc_impl only on non CUDA (#646) (@giordano)
• CPU backend (#647) (@wsmoses)
• docs: add shardy to docs (#648) (@avik-pal)
• chore: generate shardy c wrappers (#650) (@avik-pal)
• Regenerate MLIR Bindings (#651) (@github-actions[bot])
• chore: missing header files in API (#652) (@avik-pal)
• feat: the big jll PR (#653) (@avik-pal)
• [CI] Fix path of previews directory in PreviewCleanup workflow (#656) (@giordano)
• Detect TPU using PCI devices (#659) (@Pangoraw)
• Replace trim -> strip (#661) (@giordano)
• Silence various warnings in tests (#662) (@giordano)
• Feature: allow colon indexing of traced vectors (#664) (@floffy-f)
• Format code of branch "main" (#665) (@github-actions[bot])
• Regenerate MLIR Bindings (#666) (@github-actions[bot])
• KA ext (#667) (@wsmoses)
• [docs] Add information about configuration on GPU and TPU systems (#668) (@giordano)
• Fix ntuple traced type issue on unionall (#669) (@wsmoses)

Closed issues:

• Document solution of #526 (#584)
• ceil method not defined when casting to an integer (#618)
• TPU Profiler (#630)
• Failing tests on custom number types (#635)
• Enzyme.autodiff(::ReverseMode) returns nothing derivs while Enzyme.gradient works (#657)

v0.2.22

Reactant v0.2.22

Diff since v0.2.21

Merged pull requests:

• [CI] Move tests on aarch64 linux to GitHub Actions (#543) (@giordano)
• feat: multi GPU support (#587) (@avik-pal)
• feat: expose gpu memory allocation options (#589) (@avik-pal)
• Fix condition to skip CUDA tests on aarch64 (#592) (@giordano)
• feat: add the new optimization passes (#595) (@avik-pal)
• feat: support lowering custom fp types (#596) (@avik-pal)
• Update ReactantCUDAExt.jl (#597) (@wsmoses)
• Add convert (#598) (@wsmoses)
• feat: support dynamic indexing for reshaped arrays (#601) (@avik-pal)
• Fix dense elements attribute in Enzyme.autodiff #593 (#604) (@mofeing)
• feat: overload LinearAlgebra.kron (#607) (@avik-pal)
• feat: more indexing support (#608) (@avik-pal)
• feat: forward more base ops to chlo (#611) (@avik-pal)
• Add hermetic cuda getter (#612) (@wsmoses)
• [tests] Always skip CUDA tests on non-CUDA machines (#615) (@giordano)
• Typed rounding (#619) (@wsmoses)
• Regenerate MLIR Bindings (#621) (@github-actions[bot])
• feat: build the shardy dialect (#622) (@avik-pal)
• feat: support more set indexing (#625) (@avik-pal)
• Add bound optimizations (#626) (@wsmoses)

Closed issues:

• Tenet + Reactant + Enzyme.gradient broken on last releases (#593)
• CartesianIndex when broadcasting (#599)
• Inefficient host-device communication on GH200 (#609)
• Computing With Array of Structs on GPU (#610)
• setindex! error when copyto! with TracedRArray (#617)