The records behind the numbers of the card, copied from the conversion’s work directory with its paths replaced by
placeholders (<work> = the work directory, <eg2-transcribe> = the reader / app repository, <venv>, <home>).
SOURCES.tsv lists each copy, the record it was copied from, the speed-ladder row it backs, the sha256 of the record
as written and the sha256 of the copy.
| Folder | Records |
|---|---|
encoder/ |
the fp16 encoder: Mac M4 Max GPU gate and timing; iPhone 18 Pro GPU gate and timing, AOT h19p and JIT |
reader-v2/ |
reader v2 + word table, fp16 (the reader of the HF repo): the graph’s PyTorch parity and float16 headroom, its export record, the PyTorch labels against the release’s decoding; Mac GPU and Neural Engine gates (AOT h16c and JIT) and timing in one measurement window; the 3,257.4 s recording end to end on the Mac (pipeline/); its strip records, the AOT compiles from the stripped bundle and the Mac compare (strip/); the iPhone 18 Pro GPU timing of its ios-h19p/ compile, interleaved with reader v1’s in one launch (iphone-gpu-aot-timing.json; reader v1’s: iphone-gpu-aot-timing-reader-v1-same-launch.json); the eg2-transcribe app on the Mac M4 Max and on the iPhone 18 Pro (app/); its own SOURCES.tsv |
reader/ |
reader v1 (HF revision a4b7895d) + word table, fp16: Mac GPU gate, timing and the zero-half red arm; Mac Neural Engine gate; the iPhone 18 Pro GPU timing of the same compiled graph with an interim checkpoint’s weights |
pipeline/ |
reader v1: the 3,257.4 s recording end to end on the Mac (encoder GPU ∥ reader Neural Engine with two windows in flight; encoder GPU → reader GPU one window at a time) and the float32 PyTorch path on the same windows |
fixture-20.md |
the transcripts of the 20 fixture utterances (reader v1’s PyTorch path) and the one the Core AI fp16 reader v1 writes differently |
strip/ |
strip_<name>.json: the bundles of the HF repo’s macos/ and ios/ are the gated bundles with their MLIR debug locations removed (coreai-torch strip_debug_info); aot_h19p_<name>.json: the bundles of ios-h19p/ are compiled from those stripped .aimodel bundles and hold the same resources.bin as the compiles of the gated graphs; compare_mac.json: on the Mac the stripped bundles return outputs bit-identical to the gated ones (the encoder, and reader v1; reader v2’s are in reader-v2/strip/) |
encoder-split7/ |
the second encoder form (the HF repo’s macos-encoder-split7-30s/, Mac only): its gate, timing and measurement window, the one-bundle encoder’s timing in that window, the strip records of its two bundles and the Mac compare; its own README.md gives the contract |
app/ |
the eg2-transcribe app with reader v1 on HF revision a4b7895d’s files: on the Mac M4 Max its default form and both bundles on the GPU, on the iPhone 18 Pro its default form and the timing record of its passes (the 3,257.4 s recording, the 20 fixture utterances, launch to transcript after a Core AI cache clear and from the cache); its own SOURCES.tsv |
The ios-h19p/ bundles were compiled from the stripped .aimodel bundles; the app ran the HF repo’s on the iPhone 18
Pro with reader v2 (reader-v2/app/: fixture transcripts 20/20 = PyTorch, long-recording WER 3.81 %) and HF revision
a4b7895d’s with reader v1 (app/: 20/20, WER 5.77 %). The card’s iPhone reader row ran the HF repo’s ios-h19p/
reader; its iPhone encoder rows were measured on the compile of the unstripped graph.