Draft Cost Ratio From Time

Problem

Compute the cost ratio of running the draft model versus the target model per token: draft_ms_per_token / target_ms_per_token. This is the per-token time ratio that speculative decoding multiplies by the number of drafted tokens.

The signature is draft_cost_ratio(draft_ms_per_token, target_ms_per_token).

Hint

A draft ten times faster per token has a cost ratio of 0.1.

Test cases

case 1

in: [1,10]

expected: 0.1

case 2

in: [0,5]

expected: 0

case 3

in: [5,5]

expected: 1

case 4

in: [2,8]

expected: 0.25

Starter codePython

def draft_cost_ratio(draft_ms_per_token, target_ms_per_token):
    # Your code here
    pass

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