QLoRA Memory

Problem

Estimate QLoRA weight memory in bytes: base parameters stored at base_bits plus LoRA adapters stored at lora_bytes, converted with / 8: (base_params * base_bits + lora_params * 8 * lora_bytes) / 8.

The signature is qlora_memory(base_params, lora_params, base_bits=4, lora_bytes=4). Return a float number of bytes.

Hint

4-bit base weights plus fp32 adapters gives large savings with almost no quality loss.

Test cases

case 1

in: [8000000,160000,4,4]

expected: 4640000

case 2

in: [1000000,0,4,4]

expected: 500000

case 3

in: [1000000,1000000,4,4]

expected: 4500000

case 4

in: [1000,1000,2,2]

expected: 2250

Starter codePython

def qlora_memory(base_params, lora_params, base_bits=4, lora_bytes=4):
    # Your code here
    pass

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