Attach activation observers to a float model ahead of calibration.

prepare is step one of the eager-mode static post-training quantization (PTQ) workflow — the classic prepare → calibrate → convert pipeline. It does not touch weights or change any numerics; it walks the model and, for every quantizable module the QConfigMapping selects, attaches an activation observer as a forward hook. As calibration data flows through the returned model, each observer records the running range (min / max, or a histogram) of that module's output, and convert later reads those ranges to freeze the per-layer activation scale / zero_point.

Only module types in the observed set — Linear, the Conv / ConvTranspose family, the intrinsic fused Conv/Linear+ReLU modules, a handful of activations, and the QuantStub / FloatFunctional boundaries — receive an observer, and only where the mapping returns a non-None QConfig for them. The model is switched to eval mode so BatchNorm / Dropout stay deterministic during calibration. Use this when you want the best accuracy and can afford a calibration pass over representative data; for a calibration-free path see quantize_dynamic, and to make quantization error visible to the optimizer see prepare_qat.