The argmax now runs inside the graph, built from reduce_max and elementwise ops so it stays on the ANE. 1.05x to 1.63x faster and 7 to 48 percent lower peak memory on an iPhone 16. XNNPACK variants are unchanged.
Core ML: emit a (1,520,520) int32 index map instead of (1,21,520,520) logits
The argmax now runs inside the graph, built from reduce_max and elementwise ops so it stays on the ANE. 1.05x to 1.63x faster and 7 to 48 percent lower peak memory on an iPhone 16. XNNPACK variants are unchanged.
Core ML: emit a (1,520,520) int32 index map instead of (1,21,520,520) logits
The argmax now runs inside the graph, built from reduce_max and elementwise ops so it stays on the ANE. 1.05x to 1.63x faster and 7 to 48 percent lower peak memory on an iPhone 16. XNNPACK variants are unchanged.
Stop restating quantized, and drop the unused default
quantized duplicated precisions.json, which is authoritative for the
quantized/non-quantized partition. default expressed a per-repository
recommendation nothing consumed: the library resolves its own per-platform
default. Both are gone from the schema.
Core ML: drop the fp32 variant, fp16 is equivalent in task terms
fp32 is both slower and larger than fp16 on device, and the ANE will not accept fp32 at all, so an fp32 Core ML build is structurally excluded from the accelerator that makes Core ML worth using.
Equivalence was gated with a task metric rather than raw tensor distance: pixel argmax agreement 99.96% across 8 real images. Raw relative L2 on logits reads ~1% for the segmentation models, but that is logit noise which does not change the predicted class per pixel.
Models where fp16 and fp32 genuinely diverge (ssdlite320 86% box agreement, fast-sam 93%) keep both variants.
v0.9.0 and below are untouched.
