Tanh

Quantized lucid.nn.Tanh — tanh(x) then output fake-quant.

The inference-time replacement for a calibrated float nn.Tanh, installed by lucid.quantization.convert / from_float. Under the sidecar representation (design B) the quantized layer is the float tanh followed by a fake-quantize to the calibrated output grid, so a standalone tanh inside a quantized region is requantized rather than left at full precision.

Because $\tanh(x) \in [-1, 1]$ the range is known a priori, making this another fixed-qparams activation. Since the range is signed, the fixed grid centres the zero-point in the middle of the quint8 codes so that both the -1 and +1 extremes are representable.

so codes 0 .. 255 tile $[-1, 1]$ with the real value 0 landing on code 128. $\operatorname{fake\_quant}(t) = (\operatorname{clamp}( \operatorname{round}(t/S) + Z, q_{\min}, q_{\max}) - Z)\, S$.