coreml

Lucid computes on Accelerate (CPU) and MLX (Metal). Neither targets the Neural Engine, so a whole processor on every machine Lucid supports was unreachable; the same gap meant a model trained here could not ship inside an iOS or macOS app. Core ML is the only public route to either.

No third-party dependency. The package format is written by lucid/_C/coreml — the MIL protobuf, the weight blob, the bundle — and executed through Apple's own CoreML.framework, which stands beside Accelerate and Metal rather than beside a pip package. Nothing under lucid/ imports anything external (H4).

Two things worth knowing before trusting an export.

The Neural Engine only runs float16. A float32 program asked for with ComputeUnits.CPU_AND_NE does not warn, does not error, and runs at CPU speed — measured: zero of its operations are scheduled on the ANE. Pass Precision.FLOAT16 to actually reach it, and expect the ~1e-4 that half precision costs. CoreMLModel.compute_plan reports where the operations really landed, so this is checkable rather than inferred from a stopwatch.

Coverage is narrow and loud. The mapped operations are the ones real models were measured to emit; anything else raises UnsupportedOp naming the gap, because a package quietly missing a layer still loads and still returns plausible numbers.

Smaller packages. weights= stores the weights compressed — WeightPrecision for eight bits with a scale per output channel, Palettize for one to eight bits through a lookup table, Sparsify for the survivors plus a mask. Measured on a trained ResNet-50, against its top-1 over five inputs: int8 is 3.9x smaller and keeps 4/5, six-bit palettization 4.3x and 4/5. Below six bits post-training palettization changes every prediction, and no export setting recovers it — which is what CompressionAware is for.

Models from lucid.quantization export. A model prepared for quantization-aware training carries observers that record what passes through them, and tracing is not calibration, so they are paused around it and put back; its weights were trained onto their grid, so storing them there costs nothing — measured at 0.00e+00 against the eager model where the same network quantized afterwards lands at 7e-04. A converted model — quantize_dynamic, or convert — holds a packed MLX linear whose Metal-only kernel the tracer cannot follow, and is exported through that form's own dequantize-to-float reference path, which is the same arithmetic to 5e-07 rather than an approximation of it. Activations decides whether a prepared model's activation quantization is carried into the package as arithmetic or left out; Core ML quantizes weights only, so carrying it is faithful and is work the accelerator did not ask for.