Quantize

Runtime entry into the quantized region — fake-quantizes the input.

The converted, inference-time form of a calibrated QuantStub, installed by lucid.quantization.convert (or built directly via from_float). Where the stub was a passive marker, Quantize is active: every forward maps a real-valued activation onto the calibrated (scale, zero_point) grid the observer recorded, which is the numerical "entry" into the quantized region.

Under the sidecar representation (design B) the quantized activation is carried as float — the result has int8 numerics (values snapped to the grid) but remains an ordinary float32 tensor, so it flows through the rest of the quantized graph without any packed-integer container. Integer inputs such as token indices are recognised and passed through unchanged, since only real-valued activations have a meaningful quantization grid.

where $S, Z$ are the calibrated activation (scale, zero_point) and $q_{\min}, q_{\max}$ the grid bounds (0, 255 for the default quint8). The inner round-and-clamp is what quantizes; the outer $(\cdot - Z)\, S$ re-expresses the integer code as the nearest representable float.