All pull requests

Pull requests

• test(microkernel): the tile layer picks the instruction the plane fragment microkernel picks

  #1139·louisfd opened Nov 10, 2026Member·

  ·

• Promote dev to main (batch 73)

  #1138·louisfd opened Nov 10, 2026Member·

  ·

• perf(attention): deep prompt chunks run the split routine, and its unmasked steps skip the bias

  #1136·louisfd opened Nov 10, 2026Member·

  ·

• Serving: attention reads listed blocks of keys

  #1135·louisfd opened Nov 9, 2026Member·

  ·

• Walk an additive update rule beside the delta rule

  #1130·SamuelBelanger opened Nov 9, 2026Contributor·

  ·

• WIP: compute 2D convolutions as one tile contraction

  #1092·SamuelBelanger opened Nov 5, 2026Contributor·

  ·

• test: validate cubecl#2027 (LLVM shr signedness)

  #1053·lega0208 opened Nov 1, 2026·

  ·

• feat(convolution): let a caller choose the plane tiling

  #1023·vboussot opened Oct 29, 2026·2 of 4 tasks·

  ·

• fix(matmul): refuse plane-op batch routines on a device without plane operations

  #1022·vboussot opened Oct 29, 2026·2 of 4 tasks·

  ·
  1

• fix(convolution): accumulate half-precision direct convolutions in f32

  #1021·vboussot opened Oct 29, 2026·2 of 4 tasks·

  ·

• fix(matmul): refuse a TMA stage whose box exceeds the 256-element tensor map limit

  #961·bayanasar opened Oct 27, 2026·

  ·

• feat: prefix scan

  #952·Lichtso opened Oct 25, 2026·22 of 29 tasks·

  ·
  2

• docs rs all features

  #702·Redhawk18 opened Sep 25, 2026·4 tasks·

  ·

• fix: conv2d bench allocates NHWC tensors for the NHWC kernel

  #683·onenewcode opened Sep 22, 2026·

  ·
  1