All pull requests

Pull requests

• Split concat-backed multi-slice elementwise ops and register TransposeLikeBroadcastElementwise

  #3957·wsmoses opened Nov 13, 2026Member·

  ·
  2

  1

• Export dynamic symbols in enzymexlamlir-opt for crash backtraces

  #3956·wsmoses opened Nov 13, 2026Member·

  ·
  1

  1

• Simplify Reduce(Reduce/DotGeneral) across transposes and Slice(While(DUS)) along loop axis

  #3955·wsmoses opened Nov 13, 2026Member·

  ·
  2

  1

• Eliminate chlo.conj across closed commutative elementwise and structural cones

  #3954·wsmoses opened Nov 13, 2026Member·

  ·
  1

  1

• Eliminate reverse ops across closed reversible cones

  #3953·wsmoses opened Nov 13, 2026Member·

  ·
  1

  1

• Fold gather transposes and hoist reduce before broadcast_in_dim

  #3952·wsmoses opened Nov 13, 2026Member·

  ·
  2

  1

• Sink slice through single-use rotate and reuse common powers

  #3951·wsmoses opened Nov 13, 2026Member·

  ·
  2

  1

• Hoist broadcasts and splat constants out of enzyme.batch

  #3950·wsmoses opened Nov 13, 2026Member·

  ·
  1

  1

• Factor scaled shared-operand dot_generals and fuse adjacent reduce_windows

  #3949·wsmoses opened Nov 13, 2026Member·

  ·
  2

  1

• Fix complex ADJOINT derivatives in stablehlo.triangular_solve

  #3948·wsmoses opened Nov 13, 2026Member·

  ·
  1

  1

• Remove unused DenseMap locals in enzyme_ad flagged by future LLVM.

  #3946·copybara-service[bot] opened Nov 13, 2026Bot·

  ·

• Support gather/offset diagonals in DotGeneralOnlyDiagonalAccess and iota-select symmetry

  #3945·wsmoses opened Nov 13, 2026Member·

  ·
  2

  1

• Pad non-divisible sharded dimensions in MultiRotate communication optimization

  #3944·wsmoses opened Nov 13, 2026Member·

  ·
  1

  1

• Speculate loop-invariant reduction load/store when provably in bounds

  #3943·wsmoses opened Nov 13, 2026Member·

  ·
  1

  1

• Enable ConcatSliceStack in EnzymeHLOOpt and add CSE for stablehlo.not

  #3942·wsmoses opened Nov 13, 2026Member·

  ·
  2

  1

• Run gpu-eliminate-barriers after gpu1, infer llvm.nocapture, and preserve kernel arg attrs

  #3941·wsmoses opened Nov 13, 2026Member·

  ·
  3

  1

• feat: support batched CPU potrf and lower stablehlo.cholesky to LAPACK on CPU

  #3940·wsmoses opened Nov 13, 2026Member·

  ·
  1

• fix: avoid 0/0 NaN in reverse-mode autodiff of multiply reductions

  #3939·wsmoses opened Nov 13, 2026Member·

  ·
  1

  1

• fix: raise scalar/struct constant globals and lower CPU kernel barriers & shmem

  #3938·wsmoses opened Nov 13, 2026Member·

  ·
  1

  1

• fix: support shifted wrap convlike/reduce_window and dynamic_reshape in raising (#1206, #1678)

  #3937·wsmoses opened Nov 13, 2026Member·

  ·
  2

  1

• Lower chlo.top_k(k=1) to reduce, factor scalars out of FFT, and fold finite Inf compares

  #3936·wsmoses opened Nov 12, 2026Member·

  ·
  3

  1

• Update jax-ml/jax to commit eeb73fe61e293db9ba74596b9c3a86859ae791b2

  dependencies
  #3935·enzymead-bot[bot] opened Nov 12, 2026Bot·

  ·

• Avoid widening ConvertConcat and reorder BroadcastInDim of widening Convert

  #3934·wsmoses opened Nov 12, 2026Member·

  ·
  2

• EnzymeHLOOpt: simplify ReshapeBroadcast, DUSToDynamicPad, and SyrkSimplifyOutputUplo

  #3933·wsmoses opened Nov 12, 2026Member·

  ·
  3

  1

• Optimize communication for nested Extend(Extend(x)) operations

  #3932·wsmoses opened Nov 12, 2026Member·

  ·
  1

  1