Function get_scale_zero_point

tests/unit/ops/quantizer/test_quantize.py:50–65

(q_bits, is_symmetric_quant, max, min, absmax, scales=None, zero_points=None)

def get_scale_zero_point(q_bits, is_symmetric_quant, max, min, absmax, scales=None, zero_points=None):

    q_range, q_max, q_min = get_q_props(q_bits)

    if is_symmetric_quant:
        scale = torch.empty_like(absmax)
        for i, x in enumerate(absmax):
            scale[i] = torch.ones_like(x) if x == 0 else q_range / (2 * x)
        zero_point = torch.zeros(scale.shape, dtype=torch.float32, device=get_accelerator().device_name())
    else:
        scale = torch.empty_like(max)
        for i, x in enumerate(max):
            scale[i] = torch.ones_like(x) if max[i] == min[i] else q_range / (max[i] - min[i])
        zero_point = q_min - (min * scale)

    return scale, zero_point


def int4x2to2xint4(int4X2tensor):