Function quant_weight_fp16

gpu/convert_checkpoint.py:29–32

(weight)

        return new_weight, new_scale.reshape(1)

    def quant_weight_fp16(weight):
        s = 1.0 / weight.abs().mean().clamp_(min=1e-5)
        new_weight = (weight * s).round().clamp(-1, 1) / s
        return new_weight

    def convert_int8_to_int2(weight):
        return convert_weight_int8_to_int2(weight)