LoRA Parameter Count

Problem

Count the trainable parameters of a LoRA adapter on a d_in x d_out weight matrix: rank * (d_in + d_out) for the low-rank A and B matrices (biases are frozen).

The signature is lora_param_count(d_in, d_out, rank). Return an integer.

Hint

A tiny rank makes adapters hundreds of times smaller than the base matrix.

Test cases

case 1

in: [768,768,8]

expected: 12288

case 2

in: [512,512,4]

expected: 4096

case 3

in: [1024,1024,16]

expected: 32768

case 4

in: [64,32,2]

expected: 192

Starter codePython

def lora_param_count(d_in, d_out, rank):
    # Your code here
    pass

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