Issues

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

#116

· glenn-jocher opened

on Sep 18, 2026

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

#115

· glenn-jocher opened

on Sep 18, 2026

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

#113

· adityavaid opened

on Sep 17, 2026

QDQ multiply returns scale-dependent wrong values on ANE

#112

· cadamcat opened

on Sep 14, 2026

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

#93

· 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)

#76

· 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)

#75

· xocialize opened

on Aug 30, 2026

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

#74

· 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

#73

· xocialize opened

on Aug 30, 2026

Support Python 3.14 on macOS

enhancement

#70

· cdeil opened

on Aug 26, 2026

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

#68

· 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

#67

· john-rocky opened

on Aug 25, 2026

Navigation Menu

All issues

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

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

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

QDQ multiply returns scale-dependent wrong values on ANE

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

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

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)

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

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

Support Python 3.14 on macOS

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

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

Search results