All pull requests

Pull requests

• Plan: order runtime-symbol consumers after their defining node

  #2932·willwade opened Sep 25, 2026·

  ·

• (perf) Skinny-K GEMM groundwork: cost model + kernel selection

  #2930·czoli1976 opened Sep 25, 2026Contributor·

  ·
  3

• Fix typed translation of dynamic-length Range used for shape construction

  #2929·willwade opened Sep 25, 2026·

  ·
  1

• metal: a binary merged two axes broadcasting opposite ways

  #2927·kali opened Sep 25, 2026Collaborator·

  ·
  2

• (perf) depthwise: ZoneScanner reuse in parallel path

  #2925·czoli1976 opened Sep 25, 2026Contributor·

  ·
  1

• (perf) multithread: adaptive par_chunks_mut chunking

  #2924·czoli1976 opened Sep 24, 2026Contributor·

  ·
  1

• (perf) depthwise: plane-size guard + smarter par_chunks_mut chunking

  #2923·czoli1976 opened Sep 24, 2026Contributor·

  ·
  11

• transformers: add routed mixture-of-experts FFN operator

  #2911·JulienBalianSonos opened Sep 22, 2026Collaborator·

  ·
  27

• 5090: TMA for flash-attention K/V tiles

  #2893·czoli1976 opened Sep 17, 2026Contributor·

  ·
  25

• metal: drain autoreleased command buffers and encoders

  #2891·JulienBalianSonos opened Sep 17, 2026Collaborator·

  ·
  2

• H100: parked — not TK L07

  #2890·czoli1976 opened Sep 17, 2026Contributor·

  ·
  1

• H100: parked — TMA WGMMA GEMM, wait for #2880

  #2889·czoli1976 opened Sep 17, 2026Contributor·

  ·
  1

• H100: Hopper flash TMA — parked (5090 TMA is a separate PR)

  #2886·czoli1976 opened Sep 17, 2026Contributor·

  ·
  1

• H100: parked — TK L06 is TMA, not cp.async on WGMMA

  #2884·czoli1976 opened Sep 16, 2026Contributor·

  ·
  1

• H100: Hopper WGMMA GEMM — parked until hardware is available

  #2880·czoli1976 opened Sep 16, 2026Contributor·

  ·
  3

• Jetson/4060/5090: allocate dynamic shared memory the way ThunderKittens does

  #2877·czoli1976 opened Sep 16, 2026Contributor·

  ·
  23

• B6: device storage slicing

  #2856·JulienBalianSonos opened Sep 14, 2026Collaborator·

  ·
  3

• B5: device-resident outputs

  #2855·JulienBalianSonos opened Sep 14, 2026Collaborator·

  ·
  5

• GPU arena reuse and Metal buffer pool

  #2823·JulienBalianSonos opened Sep 10, 2026Collaborator·

  ·
  11

• cli: glibc's allocator carries a lot of overhead on the graphs the be…

  #2789·kali opened Sep 7, 2026Collaborator·

  ·
  6

• (perf) core/cnn: NCHW spatial kernels (direct conv, maxpool, nearest, deconv)

  #2771·czoli1976 opened Sep 3, 2026Contributor·

  ·
  41

• (perf) core/einsum: do not pick AMX/SME when C's n axis is strided

  #2733·czoli1976 opened Aug 27, 2026Contributor·

  ·
  1

• (perf) core: fuse a squeezed unit matmul axis and a unicast Mul+Add pair

  #2728·czoli1976 opened Aug 26, 2026Contributor·

  ·
  9

• (fix)core/cnn: fuse ONNX QDQ Conv into int8 GEMM instead of running it in float

  #2725·czoli1976 opened Aug 25, 2026Contributor·

  ·
  1

• Feat/gelu f16 lut

  #2687·kali opened Aug 20, 2026Collaborator·

  ·
  1