apple/coreai-torch

View on GitHub

Bridges PyTorch and Core AI. Convert existing models to Core AI IR, or author new ones from PyTorch via composite ops, custom op lowerings, and inline Metal GPU kernels.

Stars: 152
Forks: 45
Open beginner issues: 0
Indexed issues: 22
Avg merge: 1d 7m
Merged PRs (30d): 12

Dominant language: Python
License: BSD-3-Clause
Last GitHub push: Sep 10, 2026
Latest indexed: Sep 14, 2026
Contributing guide: Contributing guide
Code of conduct: Code of conduct
Beginner labels: No beginner labels indexed

Open All

22 open issues indexed

externalize: SDPA submodule re-export drops the upper bound on the key-length dim with a static query + dynamic KV context Open

apple/coreai-torch#1 · 0 comments · 1 reaction · 1 assignee · Indexed Sep 14, 2026
Runtime MPSGraph null-deref (mlir::FloatType::getWidth) executing 2+ GatedDeltaUpdate layers + an attention layer with dynamic KV context Open

apple/coreai-torch#2 · 2 comments · 0 reactions · 1 assignee · Indexed Sep 14, 2026
Support for transposed conv3d Open

apple/coreai-torch#3 · 0 comments · 3 reactions · 1 assignee · Indexed Sep 14, 2026
Contribution Proposal: Add stable softplus, mish, and logsumexp conversion support Open

apple/coreai-torch#5 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 14, 2026
ANECompiler crash (ANECPlistInterface::addOpToNetwork EXC_BAD_ACCESS) at AIModel.load when slice_update begin/end indices are runtime values Open

apple/coreai-torch#6 · 0 comments · 0 reactions · 1 assignee · Indexed Sep 14, 2026
Converter folds float→int→float cast round-trips, dropping truncation semantics (CPU too) Open

apple/coreai-torch#9 · 0 comments · 1 reaction · 1 assignee · Indexed Sep 14, 2026
GPU delegate executes aten.floor/trunc/ceil as identity; round uses away-from-zero ties; div(x,1,floor) folds to identity Open

apple/coreai-torch#10 · 0 comments · 2 reactions · 2 assignees · Indexed Sep 14, 2026
Runtime clobbers an unrelated live tensor when an int64-comparison bool mask chain executes (deformable-attention sampler pattern) Open

apple/coreai-torch#11 · 0 comments · 1 reaction · 1 assignee · Indexed Sep 14, 2026
softplus, mish, logsumexp, logcumsumexp overflow in fp16 on ANE — missing stable decompositions Open

apple/coreai-torch#21 · 0 comments · 0 reactions · 1 assignee · Indexed Sep 14, 2026
AIProgram.optimize() removes broadcasting-significant axis moves and silently miscompiles N×N distance expressions Open

apple/coreai-torch#49 · 0 comments · 0 reactions · 1 assignee · Indexed Sep 14, 2026
[ANE] FP16 numerical discrepancy in MobileNetV3 (2D MatMul + Hardswish) Open

apple/coreai-torch#51 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
Depthwise conv_transpose aborts in MPSGraph (explicit_padding = -1) when output_padding > padding Open

apple/coreai-torch#58 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
aten.topk costs a fixed 1.75 ms on device when the rest of the graph is on the ANE Open

apple/coreai-torch#66 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 14, 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 Open

apple/coreai-torch#67 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
Zero-sized tensors abort the process on GPU/ANE: a 0-length split section fails MPSCommonRuntimeCanonicalization, a width-0 output fails in MPSNDArray Open

apple/coreai-torch#68 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
Support Python 3.14 on macOS Open

enhancement

apple/coreai-torch#70 · 0 comments · 0 reactions · 1 assignee · Indexed Sep 14, 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 Open

apple/coreai-torch#73 · 0 comments · 0 reactions · 1 assignee · Indexed Sep 14, 2026
CPU delegate silently miscompiles a bool bitwise_and chain cast to float — GPU lane is correct, no diagnostic emitted Open

apple/coreai-torch#74 · 0 comments · 0 reactions · 1 assignee · Indexed Sep 14, 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) Open

apple/coreai-torch#75 · 2 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
debugging comparator cannot execute the source ExportedProgram — TypeError: 'int' object is not callable in _guards_fn (0.4.1 and 0.4.2) Open

apple/coreai-torch#76 · 0 comments · 0 reactions · 1 assignee · Indexed Sep 14, 2026
Grouped LUT convolution returns wrong per-group scaled values on ANE Open

apple/coreai-torch#93 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026
QDQ multiply returns scale-dependent wrong values on ANE Open

apple/coreai-torch#112 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 14, 2026

apple/coreai-torch

Get new issues in your inbox