Eval before anything. Ground truth + one scalar metric (F-measure, accuracy — whatever the task has) + a script that runs any model variant against it. Every later decision — quantize? fp16? distill? — is this script's output, never a guess. Keep 5–10 held-out samples as a smoke set.
Export ONNX at a fixed shape (torch.onnx.export, one export per batch size — see Gotchas). Verify against the original: max abs diff and the eval score. onnxruntime CPU is the portable baseline: web (onnxruntime-web WASM), Android (ORT mobile), desktop.
Quantize weights to int8 (onnxruntime.quantization / ct.optimize.coreml.linear_quantize_weights). Weights-only int8 is usually free — measured twice on a 20M-param transformer: F 0.872→0.871 (ONNX) and 0.864→0.863 (Core ML), 4x smaller. Re-run the eval anyway; "usually" is not "always".
Apple: convert to native Core ML (coremltools), never the ORT Core ML execution provider (see Don't). Conversion of transformer-ish models fails on Python-int shape math; the fixes that work:
- Replace einops layers/rearrange with explicit permute/reshape.
- Bake shapes as constants: capture tuple(int(s) for s in x.shape) on a module attribute during a warm eager pass, use those ints in forward — the trace then contains zero aten::size/aten::Int ops. Fixed export shape makes this sound.
- Derive dims from weights (linear.out_features), not from tensors.
- After EVERY replacement: run the model, assert max abs diff ≈ 0 vs the original. A stack of "obviously equivalent" rewrites drifts; a stack of asserted ones doesn't.
Identify failing ops by graph census, not by reading code. When the converter names a node (blocks/0/partial/76), dump ts.inlined_graph, filter nodes by scope and kind (aten::Int, aten::size, prim::NumToTensor), and match. The class you think is in the call path may be a similarly-named neighbour that is never called — a patch that changes nothing (diff 0.000e+00 because the code never ran) looks exactly like a patch that is perfectly equivalent.
Benchmark on the target device, with the model's native window size and deterministic non-zero input. Compare compute units (cpuOnly, GPU, ANE) per model — a rotary-attention transformer measured 2x FASTER on iPhone CPU than on the ANE, and the ANE's first load cost 43s of one-time compilation. Dev-machine numbers do not transfer; a busy dev machine's numbers don't even reproduce.
Deliver as a download, not in the bundle. Tens of MB belong in a FaceModelStore-style on-demand store: download → MLModel.compileModel → Application Support (excluded from backups) → validate by loading before declaring installed. An .mlpackage is a directory: host it as its files under one base URL (Manifest.json, Data/com.apple.CoreML/model.mlmodel, Data/com.apple.CoreML/weights/weight.bin) and pin that list with a test.
Leave a bench door open: a debug HTTP endpoint in the app that loads the installed model and times N predicts turns every future variant into a one-curl measurement on a real device over Wi-Fi.

Gotchas

ONNX tracing bakes batch size and lies silently. A model traced at batch 1 accepts (4, N, D) and returns garbage shaped like an answer (four sequences fused into one attention pass) — wrong result, not an error. One export per batch size, or stay at batch 1.
Version pairs are load-bearing: coremltools supports a narrow torch range (e.g. torch 2.7.x + coremltools 9.0), and torchaudio must match torch to the minor or imports die with ABI errors. Pin the pair in a dedicated venv and write it in the doc.
Unsubstituted build variables pass URL parsing. $(MODEL_URL) left in Info.plist is non-empty and URL(string:) accepts it — the download then fails at runtime with a nonsense error. Reject any value containing $(.
Batching probably doesn't help on-device inference (measured 0.91–1.01x) and an unshuffled benchmark invents speedups from machine-load drift — interleave variants within each round.
fp16 conversion may fail where int8 succeeds (Cast nodes in attention); int8 is smaller anyway. Don't fight fp16 for its own sake.

Don't

Don't use ONNX Runtime's Core ML execution provider. Measured: 11x slower than ORT CPU on the same model, plus a memory leak that grew to 41 GB over repeated runs. Native coremltools conversion of the same model ran 7x faster than ORT CPU. The EP's failure mode — graph split across unsupported ops, tensors ping-ponging between CPU and ANE — is structural.
Don't start with distillation. A converted+quantized original at 20 MB delivered on demand ships today; a distilled student is weeks of teacher infrastructure. Distill only after conversion is proven impossible or the size budget is single-digit MB.
Don't tune performance from a laptop. Simulators lack hardware codecs, ANEs differ per chip, and a loaded machine turns 14s into 29s with no code change. One number from the target device beats twenty from the Mac.

solo-model-shrink

model-shrink — ship a trained model to phones and browsers

Workflow