Method forward

gpu/model.py:73–75

(self, input)

        return (input * s).round().clamp(-128, 127).to(torch.int8), s

    def forward(self, input):
        input, s = self.quant_input(input)
        return bitnet_int8xint2_linear(input, self.weight, s, self.weight_scale)

class BitLinear(nn.Linear):
    @torch.compile