Hardswish

Quantized lucid.nn.Hardswish — hardswish(x) then output fake-quant.

The inference-time replacement for a calibrated float nn.Hardswish, installed by lucid.quantization.convert / from_float. Hardswish is the cheap piecewise-linear approximation of x·sigmoid(x) used throughout MobileNetV3 / EfficientNet-lite, chosen precisely because its clamp-and-multiply form quantizes cleanly. Under the sidecar representation (design B) the quantized layer is the float activation followed by a fake-quantize to the calibrated output grid.

Unlike Sigmoid / Tanh, hardswish is unbounded above (it grows like $x$ for large $x$), so its output range is not known a priori — it must be calibrated with a real observer rather than a lucid.quantization.FixedQParamsObserver.

with $\operatorname{fake\_quant}(t) = (\operatorname{clamp}( \operatorname{round}(t/S) + Z, q_{\min}, q_{\max}) - Z)\, S$ and the calibrated output (scale, zero_point) $S, Z$.