Issues

• Neural Engine FP16 GELU: GELU(0) = -0.000754 and 5.04% error on recorded activations (macOS 27.2, M4)

  apple/coreai-torch#149 · Filter by author Tai-Wei opened on Nov 10, 2026
  1

• NDArray.numpy() skips the BF16/FP8 fallback on NumPy 2.5 (BufferError)

  apple/coreai-torch#148 · Filter by author cdeil opened on Nov 7, 2026
  1

• to_coreai() segfaults on an assignment to an empty slice

  apple/coreai-torch#142 · Filter by author john-rocky opened on Nov 6, 2026

• Incorrect outputs during concurrent Conv3d/GroupNorm inference

  apple/coreai-torch#140 · Filter by author sandovaljoseph opened on Oct 28, 2026

• Neural Engine flushes FP16 subnormals inside RMSNorm: up to 84% error for small inputs; GPU path of the same .aimodel is correct

  apple/coreai-torch#138 · Filter by author cowboycoderhq opened on Oct 22, 2026
  1

• Additive attention mask built with arithmetic + permute gives wrong attention output on GPU and ANE (93% error); torch.where form is correct

  apple/coreai-torch#137 · Filter by author cowboycoderhq opened on Oct 22, 2026
  1

• avg_pool2d/avg_pool3d ignore count_include_pad=False and divide by kernel volume

  apple/coreai-torch#122 · Filter by author rohith500 opened on Sep 28, 2026
  1

• [Proposal] Support aten.grid_sampler_2d (bilinear, zeros/border padding)

  apple/coreai-torch#118 · Filter by author cdeil opened on Sep 23, 2026

• argmax returns a wrong result with default specialization on iPhone 17 Pro (uint8 and int32 outputs); CPU-only is correct

  apple/coreai-torch#116 · Filter by author glenn-jocher opened on Sep 18, 2026
  1

• FP16 model silently returns wrong outputs with default specialization on iPhone 17 Pro; CPU-only, GPU-preferred and FP32 are correct

  apple/coreai-torch#115 · Filter by author glenn-jocher opened on Sep 18, 2026
  1

• [Feature request] Python APIs for non-blocking Core AI submission and GPU-resident autoregressive token feedback

  apple/coreai-torch#113 · Filter by author adityavaid opened on Sep 17, 2026
  5

  

• QDQ multiply returns scale-dependent wrong values on ANE

  apple/coreai-torch#112 · Filter by author cadamcat opened on Sep 14, 2026
  1

• Grouped LUT convolution returns wrong per-group scaled values on ANE

  apple/coreai-torch#93 · Filter by author cadamcat opened on Sep 12, 2026

• debugging comparator cannot execute the source ExportedProgram — TypeError: 'int' object is not callable in _guards_fn (0.4.1 and 0.4.2)

  apple/coreai-torch#76 · Filter by author xocialize opened on Aug 30, 2026
  

• Process hard-crashes after exactly 2^14 = 16384 output NDArray allocations — floor(16384/num_outputs) inferences (all compute units; not fixed by releasing outputs or reloading the model)

  apple/coreai-torch#75 · Filter by author xocialize opened on Aug 30, 2026
  2

• CPU delegate silently miscompiles a bool bitwise_and chain cast to float — GPU lane is correct, no diagnostic emitted

  apple/coreai-torch#74 · Filter by author xocialize opened on Aug 30, 2026
  

• avg_pool2d: count_include_pad reads node.args[5] but guards node.args[4] — IndexError on 5-arg nodes (ceil_mode=True), still present in 0.4.2

  apple/coreai-torch#73 · Filter by author xocialize opened on Aug 30, 2026
  

• Zero-sized tensors abort the process on GPU/ANE: a 0-length split section fails MPSCommonRuntimeCanonicalization, a width-0 output fails in MPSNDArray

  apple/coreai-torch#68 · Filter by author john-rocky opened on Aug 25, 2026

• Error occurred when loading ANE module (MPSGraphExecutable.mm:3543; appleneuralengine Code=53 "Program load failed — no memory") — FP16 asset aborts the process on ANE load instead of returning an error

  apple/coreai-torch#67 · Filter by author john-rocky opened on Aug 25, 2026
  1

• aten.topk costs a fixed 1.75 ms on device when the rest of the graph is on the ANE

  apple/coreai-torch#66 · Filter by author john-rocky opened on Aug 25, 2026
  1

• Depthwise conv_transpose aborts in MPSGraph (explicit_padding = -1) when output_padding > padding

  apple/coreai-torch#58 · Filter by author eyupcanakman opened on Aug 17, 2026

• [ANE] FP16 numerical discrepancy in MobileNetV3 (2D MatMul + Hardswish)

  apple/coreai-torch#51 · Filter by author zli96 opened on Jul 23, 2026

• softplus, mish, logsumexp, logcumsumexp overflow in fp16 on ANE — missing stable decompositions

  apple/coreai-torch#21 · Filter by author Ashutosh0x opened on Jun 21, 2026
  

• Runtime clobbers an unrelated live tensor when an int64-comparison bool mask chain executes (deformable-attention sampler pattern)

  apple/coreai-torch#11 · Filter by author john-rocky opened on Jun 12, 2026
  1

  

• Converter folds float→int→float cast round-trips, dropping truncation semantics (CPU too)

  apple/coreai-torch#9 · Filter by author john-rocky opened on Jun 12, 2026
  1