MFU From Step Time

Problem

Compute model FLOPs utilization from a measured training step: achieved_flops_per_s = tokens_per_step * flops_per_token / step_time_s, and the hardware capacity is num_gpus * peak_flops_per_gpu.

The signature is mfu_from_step_time(tokens_per_step, step_time_s, num_gpus, peak_flops_per_gpu, flops_per_token). Return the fraction of peak FLOPs achieved.

Hint

Divide the measured token throughput times FLOPs per token by the total peak throughput of all GPUs.

Test cases

case 1

in: [200000,2,8,312000000000000,6000000000]

expected: 0.2403846154

case 2

in: [0,2,8,312000000000000,6000000000]

expected: 0

case 3

in: [1000000,1.5,4,1000000000000000,2000000000]

expected: 0.3333333333

case 4

in: [50000,10,1,50000000000000,2000000000]

expected: 0.2

Starter codePython

def mfu_from_step_time(tokens_per_step, step_time_s, num_gpus, peak_flops_per_gpu, flops_per_token):
    # Your code here
    pass

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