All pull requests

Pull requests

• feat(tensor): add NaN-ignoring nan_max/nan_min reductions

  #6278·HarpeLm opened Oct 28, 2026Contributor·3 of 4 tasks·

  ·
  2

• feat(core): stream batches through a split model with a thread per stage

  enhancement
  #6277·ThierryCantin-Demers opened Oct 28, 2026Member·

  ·
  2

• Select cubecl's native Metal runtime with a metal-native feature

  #6276·louisfd opened Oct 28, 2026Member·

  ·

• Tile attention keys its autotune at its own extents; pin cubek 625fabd

  #6275·louisfd opened Oct 28, 2026Member·

  ·

• perf(autodiff): run float collectives on the inner backend's collective

  performance
  #6274·ThierryCantin-Demers opened Oct 28, 2026Member·

  ·
  1

• feat(cubecl-fusion): fused matmul on cubek's tile routines through a trace sink (tile-kernels-only)

  #6273·louisfd opened Oct 28, 2026Member·2 of 4 tasks·

  ·

• feat(group): draw a random tensor split by its first use on each member

  enhancement
  #6272·ThierryCantin-Demers opened Oct 28, 2026Member·2 of 4 tasks·

  ·

• feat(cubecl): attention backward kernels by default, run cubek matmul plans

  #6268·louisfd opened Oct 28, 2026Member·

  ·
  1

• Sliding window masking for attentions using the cubek implementation

  #6259·gzsombor opened Oct 27, 2026Contributor·4 tasks done·

  ·
  1

• feat(train)!: add optimizer-step metrics and grouped learning-rate reporting

  #6256·laggui opened Oct 27, 2026Member·4 tasks·

  ·
  1

• ci: save Rust caches only from main and let main runs finish

  #6233·agarg21 opened Oct 26, 2026·2 of 4 tasks·

  ·
  1

  4

• Tensor-carried dtype propagation, with quantized signals

  #6220·nathanielsimard opened Oct 25, 2026Member·1 of 4 tasks·

  ·
  10

• fix(cubecl): preserve maxpool3d index strides and validate vector reductions

  #6057·HongchengZhao opened Oct 13, 2026·2 of 3 tasks·

  ·

• perf(cubecl): F32 unit convolution, im2col gather, and split depthwise wgrad in conv autotune

  #6048·znicelya opened Oct 12, 2026·3 tasks·

  ·
  2

• Add regression metrics (MSE, MAE, R²)

  #6042·Infinty-ux opened Oct 11, 2026Contributor·1 of 3 tasks·

  ·
  1

  5

• feat(data): add drop_last to the data loader

  datasetfeature
  #5911·ProgrammerProfi opened Sep 30, 2026·2 tasks done·

  ·
  1

  2

• feat(optim)!: accept a device learning rate for graph-captured steps

  bug
  #5872·nathanielsimard opened Sep 28, 2026Member·

  ·
  3

• feat(tensor): add differentiable trilinear grid_sample_3d

  feature
  #5803·thewtex opened Sep 23, 2026·3 tasks done·

  ·
  24

• chore(deps): validate CubeCL complex cast lowering

  dependencies
  #5541·shinaoka opened Aug 31, 2026·1 of 2 tasks·

  ·
  2

• feat(tensor): add GroupNorm backend primitive and CubeCL kernel

  feature
  #5504·bungogood opened Aug 27, 2026·

  ·
  6

• chore(bench): improve benchmark defaults and the comparison script

  benchmark
  #5367·laundmo opened Aug 14, 2026Contributor·2 tasks·

  ·
  5

• feat(cubecl): native argwhere/nonzero/mask_select via stream compaction

  feature
  #5309·Yashiru opened Aug 6, 2026Contributor·2 tasks done·

  ·
  5

• feat(backend): support bitcast ops for non-int kinds

  feature
  #5105·skewballfox opened Jun 24, 2026Contributor·2 tasks·

  ·
  1

• refactor: use CBOR as the serialization format across crates

  refactor
  #5100·nathanielsimard opened Jun 22, 2026Member·

  ·
  3

• feat(autodiff): retain the graph for partial derivatives

  feature
  #5043·AHA-HH opened Jun 3, 2026·2 tasks done·

  ·
  1