Bake a calibrated model into quantized inference modules.

convert is the final step of the static-PTQ pipeline (prepare → calibrate → convert). It walks the calibrated model and swaps every observed float submodule for its quantized counterpart: the learned float weight is quantized once into int8 codes plus per-channel scale / zero_point (dropping the float weight), and each activation observer's recorded range is frozen into that layer's output qparams. What comes back is a self-contained, eval-mode model with no observers left — ready to run or serialize, with a checkpoint roughly 4x smaller than the float original.

A submodule is convertible when calibration attached an activation_post_process to it; a small always-convert set (DeQuantStub, Embedding, EmbeddingBag) converts even without an activation observer. Replacements are spliced in through _modules rather than setattr so a Sequential's child order is preserved.

When lucid.backends.quantized.use_mlx() is true (the MLX quantized engine is present), the Linear family is routed to the real weight-only quantized_matmul GEMM — the genuine compute + memory win, at W8A16 numerics (weights int8/int4, activations kept float) — instead of the dequantize-to-float reference path. On a fully-Metal model this is bandwidth- bound and measures ~3.15x faster at the memory-bound decode shape M = 1 with a weight payload ~3.55x smaller; layers whose in_features is not divisible by a supported MLX group size fall back to the dequant path automatically. Set backends.quantized.engine = "reference" to force the exact W8A8 dequant numerics on every layer.