All problems
Grouped-Query Head Ratio
Problem
Compute how many query heads share each key/value head: num_query_heads / num_kv_heads.
The signature is gqa_ratio(num_query_heads, num_kv_heads).
Hint
A ratio of 1 is standard multi-head attention and a ratio equal to the head count is multi-query.
Test cases
case 1
in: [32,8]
expected: 4
case 2
in: [12,12]
expected: 1
case 3
in: [8,1]
expected: 8
case 4
in: [64,16]
expected: 4
Starter codePython
def gqa_ratio(num_query_heads, num_kv_heads):
# Your code here
pass4 test cases · run your code against them in the editor