All pull requests

Pull requests

• Fuse adjacent jit calls

  #3770·romanlee opened Oct 29, 2026Collaborator·

  ·

• CI: run fewer tests at once on the Linux runners, with a single-directory Python runtime

  #3769·wsmoses opened Oct 29, 2026Member·

  ·
  1

• Grid rewrites: a scalar the index carries moves the slice's start

  #3767·wsmoses opened Oct 29, 2026Member·

  ·
  1

• Batch fission: lift a compare or broadcast of the loop index only when it takes real work along

  #3765·wsmoses opened Oct 29, 2026Member·

  ·
  1

• [DO NOT MERGE] lit timing diagnostics 2: symlink creation, trimmed Python toolchain, tmpfs output base

  #3764·wsmoses opened Oct 29, 2026Member·

  ·

• gather_op_canon: a grid reaching past a 1-D operand reads it padded with its last element

  #3762·wsmoses opened Oct 29, 2026Member·

  ·
  1

• dynamic_slice_iota: a dynamic slice of an iota is the start plus the iota of the slice

  #3761·wsmoses opened Oct 29, 2026Member·

  ·
  1

• div_rem_of_sum_with_multiple: a multiple of the divisor in a sum contributes whole blocks

  #3760·wsmoses opened Oct 29, 2026Member·

  ·
  1

• Index reading: evaluate a value the structure does not follow at every level

  #3759·wsmoses opened Oct 29, 2026Member·

  ·
  1

• index_divrem_split: a quotient or remainder of an iota in a scatter/gather index splits the dimension

  #3758·wsmoses opened Oct 29, 2026Member·

  ·
  1

• Differentiate jit_region in reverse mode

  #3756·Pangoraw opened Oct 29, 2026Collaborator·2/4·

  ·
  7

• raise math.acos, math.acosh, math.asin, math.asinh, math.atanh, math.cosh, math.sinh to CHLO equivalents

  #3754·mofeing opened Oct 29, 2026Collaborator·

  ·
  10

• Fuse the jit_calls of kernels exposed to Enzyme

  #3752·Pangoraw opened Oct 29, 2026Collaborator·4/4·

  ·
  1

• ScatterIndicesAreUnique: through a guard's out-of-bounds select and a batch-only broadcast

  #3751·wsmoses opened Oct 29, 2026Member·

  ·
  1

• and_or_pad_intersect: an and of pads with false only keeps the windows' intersection

  #3750·wsmoses opened Oct 29, 2026Member·

  ·
  1

• Raising: a launch dimension is specialized like a loop bound

  #3749·wsmoses opened Oct 29, 2026Member·

  ·
  1

• reduce_or_and_pad_window: an or-reduce only sees the window its padded factor keeps

  #3748·wsmoses opened Oct 29, 2026Member·

  ·
  1

• [DO NOT MERGE] lit timing diagnostics on linux-x86-n4-32

  #3745·wsmoses opened Oct 29, 2026Member·

  ·

• WhileLoopInfo: memoize invariance; fission: stop copying per attempt

  #3733·wsmoses opened Oct 28, 2026Member·

  ·
  2

• XLA patch: a transpose the tiled emitter would not handle, within a 1 KB block, is coalesced (fusable)

  #3727·wsmoses opened Oct 28, 2026Member·

  ·
  1

• Add enzymexla.ml.attention and enzymexla.ml.attention_grad ops

  #3726·wsmoses opened Oct 28, 2026Member·

  ·
  1

• Raise complex accesses at a non-affine index

  #3725·dkytezab opened Oct 28, 2026Collaborator·

  ·
  1

• WhileSplitConcatCarry: carry the pieces of a concatenate the loop only reads piecewise

  #3719·wsmoses opened Oct 28, 2026Member·

  ·
  1

• Select-to-slices rewrites: only cut floating-point branches that are cheap to slice

  #3713·wsmoses opened Oct 27, 2026Member·

  ·
  1

• End-of-pipeline repair of split stencil kernels: concat_pieces_to_select, pad_computed_to_select, sink_elementwise_into_concat

  #3712·wsmoses opened Oct 27, 2026Member·

  ·
  1