xnor
Extreme low-bit inference lab: pack weights to 1–1.58 bits and run real XNOR/popcount kernels on CPU/edge — honest STE training, not fake GPU 32× from sign().
Logit-space logical activation functions for pytorch