All pull requests

Pull requests

• Fixed misaligned GGUF metadata access

  #4887·ipazarbasi opened Nov 14, 2026Contributor·

  ·

• Fixed misaligned stores in RandomBits on the CPU

  #4886·ipazarbasi opened Nov 14, 2026Contributor·

  ·

• Fixed the race on the Metal allocator's memory stats

  #4885·ipazarbasi opened Nov 14, 2026Contributor·

  ·

• Fix GPU compiled crash on an empty matmul output

  #4884·ZYZ666-RGB opened Nov 14, 2026·

  ·

• Do not wire user provided buffers

  #4883·aleroot opened Nov 14, 2026Contributor·

  ·

• Make the scheduler's active task count atomic

  #4882·ipazarbasi opened Nov 13, 2026Contributor·

  ·

• MetalAllocator::malloc(0) leads to undefined behavior

  #4881·ipazarbasi opened Nov 13, 2026Contributor·

  ·

• Fix IndexError caused by a compiled function returning stop_gradient

  #4880·katamyra opened Nov 12, 2026·

  ·
  1

• CPU fp qmm: accumulate sums in float32, similar to #4803 for affine; …

  #4879·kadyrbekovhamit-cyber opened Nov 12, 2026·

  ·

• guard against empty output before kernel launches

  #4874·nwh opened Nov 12, 2026Collaborator·

  ·

• Fix conv weight gradient with empty batch

  #4873·OgShadoww opened Nov 12, 2026·

  ·
  1

• [CUDA] Fix slow host-built arrays on platforms without managed memory

  #4868·arcusbuilds opened Nov 11, 2026Contributor·

  ·
  1

  4

• [CUDA] Avoid releasing DXGI adapters during DLL teardown

  #4867·strayberry opened Nov 11, 2026Contributor·1 task done·

  ·

• [CUDA] Synchronize LayerNorm reduction scratch reuse

  #4866·strayberry opened Nov 11, 2026Contributor·1 task done·

  ·

• [draft] fuse reductions in mx.compile on metal

  #4864·nwh opened Nov 10, 2026Collaborator·

  ·

• Note when fused SDPA kernels are available

  #4863·electron-rare opened Nov 10, 2026Contributor·

  ·

• [Metal] Skip gemv/gemv_wide for large half-precision projections on M5 Ultra

  #4862·AIvashov opened Nov 10, 2026·1 task done·

  ·

• Add bool to Metal sort kernel instantiations

  #4861·utsav1033 opened Nov 10, 2026·

  ·

• [Metal] Add 2, 3, 5 and 6 bits to affine qmv_mma

  #4860·XXXXRT666 opened Nov 10, 2026Contributor·

  ·

• [CPU] Support float64 in to_fp8 and from_fp8

  #4858·siddhantchaudhary036 opened Nov 9, 2026·

  ·
  1

  1

• Fix identical default random keys across threads

  #4857·Chr0mium24 opened Nov 9, 2026·

  ·
  1

• Add missing step default to arange signature

  #4855·lipefree opened Nov 9, 2026·

  ·

• Add staged gemv_wide for small-M fp32 matmuls

  #4854·aditrip0 opened Nov 9, 2026·

  ·

• [CUDA] Add 64-column fast path for gather_qmm

  #4853·dhiltgen opened Nov 8, 2026Contributor·

  ·

• Reuse in-flight Metal buffers on the same stream

  #4852·XXXXRT666 opened Nov 8, 2026Contributor·

  ·