All pull requests

Pull requests

• Keep the iOS18 fused scaled_dot_product_attention op when scale is given

  #2864·Yigtwxx opened Sep 17, 2026·

  ·

• Fix torch gelu converter ignoring the approximate keyword argument

  #2863·Yigtwxx opened Sep 17, 2026·

  ·

• Fix FP16 weight detection in NeuralNetworkBuilder.make_updatable (only first layer was checked)

  #2858·Anai-Guo opened Sep 12, 2026·

  ·

• Clarify supported and preferred devices in quantization docs

  #2857·cadamcat opened Sep 12, 2026·

  ·

• Fix rank-3 softmax conversion to NeuralNetwork SoftmaxLayer

  #2856·cpruijsen opened Sep 12, 2026·

  ·
  1

• docs: state the supported PyTorch version range

  #2854·Ishkirat-Singh opened Sep 9, 2026·

  ·
  1

• Guard against zero scale for all-zero blocks in weight quantization

  #2851·TusharND12 opened Sep 8, 2026·

  ·
  1

  3

• Do not count ceil_mode overhang in the avg_pool divisor

  #2844·LeSingh1 opened Aug 30, 2026·

  ·
  1

• Add sym_float and trunc support for dynamic float-scale interpolate

  #2843·SakshamKapoor2911 opened Aug 27, 2026·

  ·
  5

• Only take the single-bool-mask index path when the mask is the only index

  #2842·LeSingh1 opened Aug 27, 2026·

  ·
  3

• Support torch.gather indices smaller than the input

  #2841·LeSingh1 opened Aug 27, 2026·

  ·
  1

• Keep the integer dtype in torch.clamp when only one bound is given

  #2839·LeSingh1 opened Aug 27, 2026·

  ·
  1

• Drop two no-op cpu() calls in the k-means cuda path

  #2835·darkdi opened Aug 25, 2026·

  ·

• Fix: shape ops are never folded into quantized weights, leaving them mirrored vs dense

  #2831·kasper0406 opened Aug 24, 2026·

  ·

• Use FindPython3 consistently for source builds

  #2830·cdeil opened Aug 23, 2026·

  ·

• Acquire the GIL when releasing NumPy-backed Core ML inputs

  #2829·cdeil opened Aug 23, 2026·

  ·
  1

• Fix weight_id collision when materializing multiple source functions

  #2828·kasper0406 opened Aug 23, 2026·

  ·

• Fix SIGSEGV: release prediction input arrays with the GIL held

  #2827·kasper0406 opened Aug 21, 2026·

  ·

• Fix rms_norm eps inflation from max-|x| rescale; guard all-zero input

  #2822·nschlaepfer opened Aug 19, 2026·

  ·
  1

• Add InstanceNorm3d alias-conversion regression coverage

  #2820·sfeng01-tech opened Aug 17, 2026·

  ·
  1

• Support PyTorch angle and trig ops (deg2rad, rad2deg, hypot)

  #2817·jaseem-kurikkal opened Aug 15, 2026·

  ·
  2

• Decompose Softplus in TF frontend to avoid fp16 overflow on ANE

  #2814·axiom-of-choice opened Aug 10, 2026·

  ·
  1

• Read torch rand's dtype from the argument that holds it

  #2813·LeSingh1 opened Aug 9, 2026·

  ·

• Fall back instead of baking a symbolic dim into softplus and nll_loss

  #2812·LeSingh1 opened Aug 9, 2026·

  ·

• State the limitation when chunk / unbind meet a dynamic dim

  #2811·LeSingh1 opened Aug 9, 2026·

  ·