All pull requests

Pull requests

• Fix fp16 input rounding in gated delta NAX matmuls

  #4788·mycomputercrashsalot opened Oct 29, 2026·

  ·
  1

• [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·

  ·

• Fix CUDA matmul and addmm reduction precision

  #4784·tetsuo opened Oct 28, 2026·

  ·

• [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

• Fix complex power VJP conjugation

  #4778·devYRPauli opened Oct 27, 2026Contributor·

  ·

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

  #4776·ArupKantiDas opened Oct 26, 2026Contributor·

  ·

• Keep the existing file when save, save_safetensors or save_gguf fails

  #4773·aleroot opened Oct 26, 2026Contributor·

  ·

• [Metal] Fix fill loop in layer_norm

  #4772·luciusacarr opened Oct 26, 2026Contributor·

  ·
  1

  3

• Fix SliceUpdate vmap with unbatched update

  #4770·mahimn01 opened Oct 26, 2026Contributor·

  ·

• Fix premature rounding in addmm

  #4769·tetsuo opened Oct 25, 2026·

  ·
  1

• Check scatter and gather indices are in bounds

  #4768·turingH opened Oct 25, 2026Contributor·

  ·

• [jaccl] Detect dead RDMA links instead of crashing

  #4767·ipazarbasi opened Oct 25, 2026·

  ·
  4

• Route of per-channel gate in gated_delta_update fixed

  #4766·Exorust opened Oct 25, 2026·

  ·
  1

• Add optional source IP binding for ring connections

  low priority
  #4757·ronaldmannak opened Oct 24, 2026Contributor·

  ·

• 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

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

  ·

• Make GPU scheduler know the spin is waiting for something

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

  ·
  4

• Fix wrong results in left_shift/right_shift on overflow

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

  ·
  5