Issues

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

FP16 asset aborts the process on ANE load instead of returning an error

#67

· john-rocky opened

on Aug 25, 2026

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

#66

· john-rocky opened

on Aug 25, 2026

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

#58

· eyupcanakman opened

on Aug 17, 2026

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

#51

· zli96 opened

on Jul 23, 2026

AIProgram.optimize() removes broadcasting-significant axis moves and silently miscompiles N×N distance expressions

#49

· dkomoroske opened

on Jul 23, 2026

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

#21

· Ashutosh0x opened

on Jun 21, 2026

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

#11

· john-rocky opened

on Jun 12, 2026

GPU delegate executes aten.floor/trunc/ceil as identity; round uses away-from-zero ties; div(x,1,floor) folds to identity

#10

· john-rocky opened

on Jun 12, 2026

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

#9

· john-rocky opened

on Jun 12, 2026

Converter aborts (bad_optional_access) on aten.arange with float start/end/step args

#8

· john-rocky opened

on Jun 12, 2026

Navigation Menu

Uh oh!

All issues

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

FP16 asset aborts the process on ANE load instead of returning an error

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

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

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

AIProgram.optimize() removes broadcasting-significant axis moves and silently miscompiles N×N distance expressions

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

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

GPU delegate executes aten.floor/trunc/ceil as identity; round uses away-from-zero ties; div(x,1,floor) folds to identity

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

Converter aborts (bad_optional_access) on aten.arange with float start/end/step args

Uh oh!

Search results