The public API reserves NemotronComputePreference.neuralEngine, but the 0.1.0 model package intentionally contains no ANE artifact.
Current Xcode 27 beta findings:
- forced source inference terminates with
ANE inference operation failed due to unknown error;
coreai-build --preferred-compute neural-engine --architecture h15d fails verification around a rank-6 cache transpose (memref.transpose result rank 6), exceeding the ANE-oriented rank-5 authoring constraint;- automatic specialization resolves to the same MPSGraph/GPU implementation as explicit GPU for this graph.
The SDK therefore throws a recoverable missing-artifact error for explicit ANE requests instead of allowing a process-level abort.
Acceptance criteria:
- re-author encoder/cache layout so ANE-bound tensors are rank 5 or lower;
- validate forced-ANE source inference without aborts;
- compile device-matched ANE AOT artifacts (at minimum h15d and h18p);
- demonstrate streaming/cache and transcript parity against fp16 GPU;
- record latency, power, and memory benchmarks;
- add
aot_320ms.<platform>.neural-engine.<architecture> manifest entries and selection tests.
The public API reserves
NemotronComputePreference.neuralEngine, but the 0.1.0 model package intentionally contains no ANE artifact.Current Xcode 27 beta findings:
ANE inference operation failed due to unknown error;coreai-build --preferred-compute neural-engine --architecture h15dfails verification around a rank-6 cache transpose (memref.transposeresult rank 6), exceeding the ANE-oriented rank-5 authoring constraint;The SDK therefore throws a recoverable missing-artifact error for explicit ANE requests instead of allowing a process-level abort.
Acceptance criteria:
aot_320ms.<platform>.neural-engine.<architecture>manifest entries and selection tests.