All pull requests

Pull requests

• Fix CPU float16 SIMD comparison masks

  #4602·crabel99 opened Sep 30, 2026·

  ·
  1

• Enhancement: Replaces Black formatter with ruff

  #4600·aaishwarymishra opened Sep 30, 2026Contributor·

  ·

• Fix slicing's non contiguous start

  bug
  #4599·michalk8 opened Sep 30, 2026Collaborator·

  ·

• metal/quantized: add gather_qmv_fast_medium for K%256 (bits>=4) shapes

  #4598·Stonesjtu opened Sep 30, 2026·

  ·
  1

• metal/sdpa: unroll TK=4 keys per iter in sdpa_vector_2pass_1

  #4596·Stonesjtu opened Sep 30, 2026·

  ·
  2

• Fix complex to bool

  bug
  #4593·michalk8 opened Sep 29, 2026Collaborator·

  ·
  

• Reject non-aligned pointers when converting to mx.array

  bug
  #4592·michalk8 opened Sep 29, 2026Collaborator·

  ·

• Fix fft size one

  bug
  #4591·michalk8 opened Sep 29, 2026Collaborator·

  ·

• fix(metal): sanitize '-' in make_template_hash for negative template ints

  #4584·gokay-ai opened Sep 29, 2026·

  ·
  1

• docs: Correct the export callback type in the guide

  low priority
  #4578·ssmall256 opened Sep 28, 2026·

  ·

• Fix as_strided contiguity

  low priority
  #4575·sofinvalery opened Sep 27, 2026Contributor·

  ·
  1

• Fix MultiOptimizer crash when an optimizer's group is empty

  low priority
  #4570·ishtihoss opened Sep 26, 2026Contributor·

  ·

• Perf: Metal kernel source hash cache

  #4568·STRML opened Sep 26, 2026·

  ·
  3

• metal: runtime setter for command-buffer limits

  await discussion
  #4562·aidiffuser opened Sep 25, 2026·

  ·

• Fix default.metallib lookup under swift test with Swift Build

  low priority
  #4527·heckj opened Sep 17, 2026Contributor·

  ·
  1

  1

• Use the fast qmv kernel for affine outputs not divisible by 8

  low priority
  #4516·wyanzhao opened Sep 15, 2026Contributor·

  ·
  1

  2

• Use NAX attention for short causal D256 prefill

  await response
  #4476·wyanzhao opened Sep 8, 2026Contributor·

  ·
  3

• Support uneven tensor parallel sharding in distributed linear layers

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

  ·

• Fix 32 bit dispatch grid overflow in Metal reductions

  low priority
  #4437·Lyxot opened Aug 31, 2026Contributor·

  ·
  1

  1

• Fix strided_scan reading/writing out-of-bounds

  #4430·zcbenz opened Aug 30, 2026Member·

  ·
  3

• Small input linear and cubic upsampling should use dense matmuls

  await verification
  #4418·slink opened Aug 28, 2026·

  ·

• [CUDA] Grouped mm

  await response
  #4390·nastya236 opened Aug 24, 2026Collaborator·

  ·
  1

• Add fused reduction support

  await discussionawait verificationenhancement
  #4369·sakupan102 opened Aug 21, 2026·

  ·
  1

  2

• Collapse the non sorted axes to pick the contiguous sort kernel

  await verificationlow priority
  #4366·kapellirohith opened Aug 21, 2026Contributor·

  ·

• Add nanobind 3.0 split-mode support

  await discussion
  #4359·XXXXRT666 opened Aug 19, 2026Contributor·

  ·
  2