v0.2.161

Reactant v0.2.161

Diff since v0.2.160

Merged pull requests:

• fix: accidental promotion to i64 (#1605) (@avik-pal)
• fix: add overloads for Base.complex (#1607) (@avik-pal)
• chore: update jll (#1610) (@avik-pal)
• fix: remove unnecessary permutedims in batched_mul (#1612) (@avik-pal)
• [CI] Run job updating MLIR bindings on Julia v1.11 (#1613) (@giordano)
• fix: add within_autodiff_rewrite for new Enzyme versions (#1614) (@avik-pal)
• feat: support multiple argument map and map! (#1620) (@avik-pal)
• Specialize transpose! and adjoint! for AnyTracedRMatrix (#1621) (@oschulz)
• Update Clang compatibility version to 0.19 (#1622) (@giordano)
• Regenerate MLIR Bindings (#1623) (@github-actions[bot])
• fix: missing licm passes + accidental promotions (#1628) (@avik-pal)
• Update Reactant_jll version to 0.0.238 (#1631) (@avik-pal)

Closed issues:

• Compile errors when using AcceleratedKernels.mapreduce on a StructVector of Tuples (#1555)
• Abort trap: mismatched ranks of types1 vs 0 with Oceananigans script with ImmersedBoundaryCondition (#1600)
• Reactant_jll cannot be built with CUDA13.0 (#1603)
• copy of transpose fails due to scalar indexing (#1611)

v0.2.160

Reactant v0.2.160

Diff since v0.2.159

Merged pull requests:

• windows ci (#1562) (@wsmoses)
• fix: logsumexp (#1594) (@avik-pal)
• Reduce cpu size (#1599) (@wsmoses)

Closed issues:

• Error when precompiling in Windows (#198)
• Cannot register CUDA backend: "Could not find registered platform with name: 'cuda'" (#1591)
• Going from Reactant v0.2.152 to v0.2.153 breaks compilation of NNlib.logsumexp (#1593)
• Segfault with cumulative sum in stack/eachrow (#1597)

v0.2.159

Reactant v0.2.159

Diff since v0.2.158

Merged pull requests:

• Format Bazel code of branch "main" (#1588) (@github-actions[bot])
• fix: unwanted promotion in copysign (#1590) (@avik-pal)

v0.2.158

Reactant v0.2.158

Diff since v0.2.157

Merged pull requests:

• [build_local] Symlink libdevice and fatbinary for GPU builds (#1579) (@giordano)
• feat: new batching opt passes (#1582) (@avik-pal)
• [CI] Run local build on GCP runner (#1584) (@giordano)
• [CI] Use Codecov token (#1585) (@giordano)
• A comic number of backslashes to support rewriting patch files define… (#1587) (@wsmoses)

v0.2.157

Reactant v0.2.157

Diff since v0.2.156

Merged pull requests:

• test: mark tests as broken instead of skipping them (#1536) (@avik-pal)
• Format Julia code of branch "main" (#1580) (@github-actions[bot])

v0.2.156

Reactant v0.2.156

Diff since v0.2.155

Merged pull requests:

• fix: setup for new jll (#1538) (@avik-pal)
• feat: define a @opcall macro for better debug info (#1574) (@avik-pal)
• fix: stack for numbers (#1576) (@avik-pal)
• override ifelse (#1577) (@wsmoses)

Closed issues:

• Replace Ops.<...> with a custom @op macro (#671)

v0.2.155

Reactant v0.2.155

Diff since v0.2.154

Merged pull requests:

• Fix typedef and undef ctors for ConcreteRArray and ConcreteRNumber (#1571) (@oschulz)
• fix: mapreduce with unitrange dims (#1572) (@avik-pal)
• CompatHelper: add new compat entry for FillArrays at version 1 for package test, (keep existing compat) (#1573) (@github-actions[bot])
• Update Project.toml (#1575) (@wsmoses)

v0.2.154

Reactant v0.2.154

Diff since v0.2.153

Merged pull requests:

• test: make TPU ci green (#1534) (@wsmoses)
• fix: iteration ordering for findfirst (#1537) (@avik-pal)
• Updating branch with main (#1548) (@jlk9)
• No static libstdc++ (#1553) (@wsmoses)
• Format Bazel code of branch "main" (#1556) (@github-actions[bot])
• feat: persistent compilation cache (#1557) (@avik-pal)
• Add ConcreteRArray undef ctors and fix KA.allocate (#1558) (@oschulz)
• feat: persistent compile cache jll changes (#1560) (@avik-pal)
• feat: FillArrays support (#1568) (@avik-pal)

Closed issues:

• Enzyme.make_zero! not applying when compiled inside a gradient call (#1504)
• Reactant with Nvidia GPU on x86_64-linux-gnu taking forever to precompile on 0.2.153 vs 0.2.152 (#1550)
• Support for FillArrays? (#1563)

v0.2.153

Reactant v0.2.153

Diff since v0.2.152

Merged pull requests:

• Don't own a null array (#1524) (@wsmoses)
• More precompile memory fixups (#1525) (@wsmoses)
• Actually use c api (#1526) (@wsmoses)
• Prepare for new jll passes (#1527) (@wsmoses)
• Bump julia-actions/julia-downgrade-compat from 1 to 2 (#1528) (@dependabot[bot])
• Regenerate MLIR Bindings (#1529) (@github-actions[bot])
• Reduce TPU errors (#1531) (@wsmoses)
• fix: init for julia and shlo are semantically different (#1532) (@avik-pal)
• Format Bazel code of branch "main" (#1540) (@github-actions[bot])
• Regenerate MLIR Bindings (#1541) (@github-actions[bot])
• Format Bazel code of branch "main" (#1542) (@github-actions[bot])
• Regenerate MLIR Bindings (#1543) (@github-actions[bot])
• Fix enzyme.make_zero! (#1544) (@wsmoses)
• Update Project.toml (#1546) (@wsmoses)
• JLL: fix no nan (#1547) (@wsmoses)

Closed issues:

• TPU ifrt nondeterministically segfaults on initialization (#1497)

v0.2.152

Reactant v0.2.152

Diff since v0.2.151

Merged pull requests:

• docs: new sharding docs (#1370) (@avik-pal)
• Add widen_wrap opt (#1511) (@wsmoses)
• fix: default to shardy again (#1512) (@avik-pal)
• fix: fixes for random sampling (#1513) (@avik-pal)
• Sdy comm (#1516) (@wsmoses)
• Format Julia code of branch "main" (#1517) (@github-actions[bot])
• Format Bazel code of branch "main" (#1518) (@github-actions[bot])
• Regenerate MLIR Bindings (#1519) (@github-actions[bot])
• refactor: move patterns to diff list (#1520) (@avik-pal)
• fix: more dispatches for mapreduce (#1521) (@avik-pal)
• fix: force finalize the buffers after precompile (#1522) (@avik-pal)

Closed issues:

• rand! without rng falls-back to sampler interface (#1446)
• Host-device copying (#1458)
• GSPMD sharding propgagtion is deprecated (#1472)
• Diagonal Matrix Mult is Slower than Dense (#1483)
• Error with rand and eltype (#1508)