Per-Channel Scale Count

Problem

Count per-channel, per-group quantization scales for a weight matrix: out_features * ceil(in_features / group_size).

The signature is per_channel_scale_count(out_features, in_features, group_size). Return an integer.

Hint

Group size 128 is the common sweet spot between accuracy and scale overhead.

case 1

in: [4096,4096,128]

expected: 131072

case 2

in: [4096,4096,4096]

expected: 4096

case 3

in: [8,16,4]

expected: 32

case 4

in: [1,1,1]

expected: 1

Starter codePython

def per_channel_scale_count(out_features, in_features, group_size):
    # Your code here
    pass

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