MinMaxObserver

Per-tensor observer tracking the global running min / max of the input.

The simplest and most common calibration statistic: on every batch it folds the tensor's overall min / max into a cumulative running range, and calculate_qparams distills that single [min_val, max_val] into one per-tensor (scale, zero_point) pair. It is the natural default for activations and a serviceable choice for weights, but because the running range only ever grows, a single spike in any calibration batch permanently widens it — wasting codes on a range that most values never occupy.

When to pick it. Reach for MinMaxObserver when the tensor's distribution is well-behaved (roughly bounded, light-tailed) and you want the cheapest possible calibration. If a few calibration batches carry outliers, prefer MovingAverageMinMaxObserver, whose EMA lets the range relax back toward the bulk of the data; for heavy-tailed activations where clipping the tails buys real accuracy, prefer HistogramObserver. For weights, the per-channel PerChannelMinMaxObserver is almost always tighter.

Per batch the running range is the cumulative min/max:

seeded from $(+\infty, -\infty)$, then reduced to qparams by the affine map of lucid.quantization.calculate_qparams (range first widened to include 0):

(or the symmetric form of the base class if a symmetric qscheme is passed).