mattmireles Profile Kokoro Core ML
Bind every measurement to the package hash, model/configuration, hardware, OS, Core ML compute-unit request, input shape/duration, warmup policy, and timing boundary. Inspect a device compute plan or physical trace when available; never infer Neural Engine placement from a requested compute unit or speed alone.
Start with scripts/dump_device_compute_plan.py,
scripts/inspect_coreml_compute_plan.m, and
README/Guides/apple-silicon/CoreML-Compute-Unit-Scheduling-guide.md.
Keep synthesis, phonemization, I/O, and individual model-stage timing separate.