ELU

Quantized lucid.nn.ELU — elu(x, alpha) then output fake-quant.

The inference-time replacement for a calibrated float nn.ELU, installed by lucid.quantization.convert / from_float. Under the sidecar representation (design B) the quantized layer is the float ELU followed by a fake-quantize to the calibrated output grid. Unlike the bounded activations, ELU is unbounded above and saturates to $-\alpha$ below, so its range is data-dependent and must be calibrated with a real observer rather than a lucid.quantization.FixedQParamsObserver.

The negative-branch scale $\alpha$ is a parameter of the activation, not a learned weight; from_float copies it from the source nn.ELU (via the internal _configure hook) so the quantized layer reproduces the exact float curve before snapping the output to the grid.

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) $S, Z$.