SmoothQuant Migration Factor

Problem

Compute the SmoothQuant migration scale for one channel: s = act_max_abs^alpha / weight_max_abs^(1 - alpha). Activations are divided by s and weights multiplied by s, moving quantization difficulty from activations to weights.

The signature is smoothquant_scale(act_max_abs, weight_max_abs, alpha).

Hint

Alpha = 0.5 balances the two dynamic ranges in the geometric mean.

Test cases

case 1

in: [10,1,0.5]

expected: 3.1622776602

case 2

in: [4,1,0.5]

expected: 2

case 3

in: [10,1,0]

expected: 1

case 4

in: [10,1,1]

expected: 10

case 5

in: [8,2,0.5]

expected: 2

Starter codePython

def smoothquant_scale(act_max_abs, weight_max_abs, alpha):
    # Your code here
    pass

5 test cases · run your code against them in the editor