Scale Granularity Error Ratio

Problem

Compute the ratio of per-tensor to per-group quantization MSE for a weight matrix. Groups are contiguous chunks of group_size columns using a per-group symmetric scale; return mse_tensor / mse_group, or 1.0 when the group error is zero.

The signature is granularity_error_ratio(W, group_size, bits=8).

Hint

The ratio grows with the spread of per-group dynamic ranges.

Test cases

case 1

in: [[[1,0.7,50,0.3],[100,0,1,0.6]],2,8]

expected: 43.05562614711239

case 2

in: [[[3,2,1,0]],2,8]

expected: 2

case 3

in: [[[0,0]],2,8]

expected: 1

case 4

in: [[[1,0],[100,0]],2,8]

expected: 1

Starter codePython

def granularity_error_ratio(W, group_size, bits=8):
    # Your code here
    pass

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