All pull requests

Pull requests

• [Metal] Route qqmm without global scales like quantized_matmul

  #4796·XXXXRT666 opened Oct 29, 2026Contributor·

  ·

• use self-hosted macos runners

  #4795·madrob opened Oct 29, 2026Contributor·

  ·

• Check output size for 64-bit indexing in GPU SliceUpdate

  #4794·arian262785-cell opened Oct 29, 2026Contributor·

  ·
  1

• [Metal] Prefer qmm_nax over legacy split-k on NAX devices

  #4793·arian262785-cell opened Oct 29, 2026Contributor·

  ·
  1

• Respect MLX_ENABLE_TF32 in gated_delta_update

  #4792·zcbenz opened Oct 29, 2026Member·

  ·
  1

• Validate trace input ids match trace inputs in import_function

  #4790·turingH opened Oct 29, 2026Contributor·

  ·

• Keep double precision in CPU exp, sin, cos, erf

  #4789·mycomputercrashsalot opened Oct 29, 2026·

  ·

• [Metal] Add affine qmv_mma for 3-16 rows

  #4787·ahuzmeza opened Oct 28, 2026·

  ·

• Build and test for macOS 27

  #4786·zcbenz opened Oct 28, 2026Member·

  ·

• metal: read FP16 and BF16 affine constants directly for FP32 inputs

  #4785·bri-prism opened Oct 28, 2026Contributor·

  ·
  1

• [CUDA] Sequential kernel for gated_delta_update

  #4783·zcbenz opened Oct 28, 2026Member·

  ·

• Speed up QMM & QMV

  #4781·XXXXRT666 opened Oct 28, 2026Contributor·

  ·
  1

• Fix Metal expm1 argument reduction accuracy

  #4780·arcusbuilds opened Oct 28, 2026·

  ·
  1

• Launch fast.cross_entropy in slices past the Metal grid limit

  #4776·ArupKantiDas opened Oct 26, 2026Contributor·

  ·

• [Metal] Fix fill loop in layer_norm

  #4772·luciusacarr opened Oct 26, 2026Contributor·

  ·
  1

  3

• [jaccl] Detect dead RDMA links instead of crashing

  #4767·ipazarbasi opened Oct 25, 2026·

  ·
  4

• Bigger MoE tiles on M1 GPUs

  await response
  #4754·a3rosol opened Oct 23, 2026·

  ·
  1

• Use gemv_wide for fp32 matmuls with 2-3 rows

  await verification
  #4742·utsav1033 opened Oct 22, 2026·

  ·

• feat(nn): add freqs and partial_rotary_factor support to RoPE

  low priority
  #4732·shadyuwugurl opened Oct 20, 2026·

  ·

• Add optional int8 path for quantized matmul with affine weights

  await verification
  #4722·magnuslundstedt opened Oct 18, 2026Contributor·

  ·
  1

• Fix sharded-to-all bias and integer split positions on quantized layers

  low priority
  #4715·ronaldmannak opened Oct 17, 2026Contributor·

  ·
  4

• Fix negative padding with input_dilation > 1 in conv_general

  low priority
  #4688·MarcosAsh opened Oct 13, 2026Contributor·

  ·
  4

• Fix custom_function vmap and jvp traversal with opaque bodies

  await responselow priority
  #4612·axiom-of-choice opened Oct 2, 2026Contributor·1 task done·

  ·
  1

  2

• Perf: Metal kernel source hash cache

  #4568·STRML opened Sep 26, 2026·

  ·
  9

• Support uneven tensor parallel sharding in distributed linear layers

  await verificationlow priority
  #4465·twallgren opened Sep 4, 2026·

  ·