All pull requests

Pull requests

• perf(flex): vectorize f32 sinh and cosh

  flex
  #6354·antimora opened Oct 30, 2026Collaborator·2 of 4 tasks·

  ·
  1

  

• perf(flex): add a two-input SIMD driver for atan2, hypot and backward ops

  flex
  #6353·antimora opened Oct 30, 2026Collaborator·2 of 4 tasks·

  ·
  1

  

• perf(flex): vectorize f32 atan, asin and acos

  flex
  #6352·antimora opened Oct 30, 2026Collaborator·2 of 4 tasks·

  ·
  1

  

• perf(flex): vectorize clamp, sign and integer powi

  flex
  #6351·antimora opened Oct 30, 2026Collaborator·3 of 4 tasks·

  ·
  3

  

• refactor(fusion): resolve a collective's members through SpawnedCall

  fusionrefactor
  #6349·ThierryCantin-Demers opened Oct 30, 2026Member·1 of 4 tasks·

  ·
  2

• fix(cubecl): treat an empty tensor as non-overlapping

  bug
  #6348·ThierryCantin-Demers opened Oct 30, 2026Member·2 of 4 tasks·

  ·
  2

• perf(flex): vectorize f32 sin, cos and tan

  flex
  #6347·antimora opened Oct 30, 2026Collaborator·3 of 4 tasks·

  ·
  2

  

• perf(flex): vectorize f32 log1p, atanh, log_sigmoid and its backward

  flex
  #6345·antimora opened Oct 30, 2026Collaborator·3 of 4 tasks·

  ·
  2

  

• perf(cubecl): gather im2col columns in a single kernel

  #6343·EthanAlmloff opened Oct 30, 2026Contributor·4 tasks done·

  ·
  2

• perf(flex): vectorize f32 ln, tanh, erf, gelu and gelu_backward

  flex
  #6341·antimora opened Oct 30, 2026Collaborator·3 of 4 tasks·

  ·
  2

  

• feat(optim): train a split model as a pipeline with GPipe or 1F1B

  enhancement
  #6340·ThierryCantin-Demers opened Oct 30, 2026Member·2 of 4 tasks·

  ·
  1

• perf(fused-cubecl): add fused concat equal size fast case

  #6337·EthanAlmloff opened Oct 30, 2026Contributor·4 tasks done·

  ·
  3

• fix(tensor): prevent softmax collapse at high confidence

  #6336·istaari opened Oct 30, 2026·4 tasks done·

  ·
  1

  3

• ComputeEnvironment: a device's compiled kernels and autotune answers, wherever it runs

  #6335·nathanielsimard opened Oct 30, 2026Member·

  ·
  2

• chore: bump cubecl and cubek; follow cubek's autotune names

  #6333·louisfd opened Oct 30, 2026Member·4 tasks·

  ·
  1

• perf(cubecl): respect memory order in comparison and mask fill

  #6331·EthanAlmloff opened Oct 30, 2026Contributor·4 tasks done·

  ·
  2

• Metal: Device::metal() is cubecl's own Metal runtime, wgpu-msl the wgpu path

  #6330·louisfd opened Oct 30, 2026Member·

  ·

• perf(flex): native 1D convolution avoiding pseudo-3D expansion and im2col

  #6322·onenewcode opened Oct 30, 2026Contributor·2 tasks done·

  ·
  1

  4

• Test: a shape that cannot store quants is refused when read

  #6320·louisfd opened Oct 30, 2026Member·

  ·

• perf(cubecl) add wgrad_split_implicit candidate

  #6319·EthanAlmloff opened Oct 30, 2026Contributor·4 tasks done·

  ·
  3

• fix(tensor): return NaN from median when the slice contains NaN

  #6307·b84330808 opened Oct 29, 2026·4 tasks done·

  ·
  1

• fix(cubecl): report exclusive memory pools

  #6303·vboussot opened Oct 29, 2026Contributor·4 tasks done·

  ·
  1

  2

• perf(cubecl): split a long-k matmul and a long-axis sum into chunks

  #6302·vboussot opened Oct 29, 2026Contributor·4 tasks done·

  ·
  1

• perf(cubecl): take an extremum over a long axis on the cube routine

  #6301·vboussot opened Oct 29, 2026Contributor·4 tasks done·

  ·
  2

• perf(cubecl): run a small matmul over a long k one thread per output

  #6299·vboussot opened Oct 29, 2026Contributor·4 tasks done·

  ·
  2