All pull requests

Pull requests

• Adding sliding window attention

  #4734·jagrit06 opened Oct 20, 2026Member·

  ·
  2

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

  #4732·shadyuwugurl opened Oct 20, 2026·

  ·

• Fix zero-copy numpy arrays dtype mismatch

  #4731·michalk8 opened Oct 20, 2026Collaborator·

  ·

• Fix normal loc scale shape

  #4730·michalk8 opened Oct 20, 2026Collaborator·

  ·

• Fix mx.take's negative indices

  #4729·michalk8 opened Oct 20, 2026Collaborator·

  ·

• Widen affine qmv_wide row tiles to 8

  await verification
  #4728·Ambisphaeric opened Oct 20, 2026·1 task done·

  ·
  1

• Use data_size for Metal gather/scatter large-offset checks

  #4724·begcdn opened Oct 19, 2026·

  ·

• [Metal] Add SDPA fast paths for speculative decoding

  await verification
  #4723·dhiltgen opened Oct 18, 2026Contributor·

  ·

• Add optional int8 path for quantized matmul with affine weights

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

  ·
  1

• Fix silent Metal grid truncation in softmax, logsumexp, rms_norm and layer_norm past 2^27 rows

  low priority
  #4717·ArupKantiDas opened Oct 17, 2026·

  ·
  1

  1

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

  await verificationlow priority
  #4715·ronaldmannak opened Oct 17, 2026Contributor·

  ·

• [Metal] Tile fp quantized_matmul for 6-16 rows

  await verification
  #4712·dhiltgen opened Oct 16, 2026Contributor·

  ·
  1

• Debug time graph verification

  #4711·zcbenz opened Oct 16, 2026Member·

  ·

• Fix Windows CUDA 13 DLL loading

  await verification
  #4706·luciusacarr opened Oct 15, 2026·

  ·

• Fix fftn size one axis aliasing a temporary

  #4705·StockerMC opened Oct 15, 2026·

  ·
  1

  2

• Make GPU scheduler know the spin is waiting for something

  low priority
  #4703·aleroot opened Oct 15, 2026Contributor·

  ·
  3

• Fix fast.cross_entropy VJP on non-contiguous Metal inputs

  #4702·StockerMC opened Oct 15, 2026·

  ·
  1

  1

• Fixed lshift and rshift to return 0 for unsigned and -1 for signed value when shift is greater than dtype size

  low priority
  #4694·aaishwarymishra opened Oct 14, 2026Contributor·

  ·

• Simplify docs generation by ditching doxygen

  #4690·zcbenz opened Oct 14, 2026Member·

  ·

• Fix negative padding with input_dilation > 1 in conv_general

  await verificationlow priority
  #4688·MarcosAsh opened Oct 13, 2026Contributor·

  ·
  2

• Add "config" module

  #4647·zcbenz opened Oct 8, 2026Member·

  ·
  2

  2

• 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

  await response
  #4568·STRML opened Sep 26, 2026·

  ·
  3

• Support uneven tensor parallel sharding in distributed linear layers

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

  ·

• Fix strided_scan reading/writing out-of-bounds

  #4430·zcbenz opened Aug 30, 2026Member·

  ·
  3