MaxBackward

Autograd node for the multi-axis reduction maximum $y = \max_{i \in \text{axes}} x_i$.

The gradient routes only to positions where $x_i$ equals the reduced maximum. The equality mask is built without a dedicated equality kernel by composing two ge masks ($a = b \iff a \ge b \wedge b \ge a$). Ties (multiple equal maxima) receive equal gradient — matching the reference framework's behaviour. The forward output is saved so the mask can be formed without rerunning the reduction.

Math