Dequant Fusion Cost

Problem

Compute memory traffic with and without dequantization fusion. Unfused traffic is quant_bytes + 2*dequant_bytes (write then re-read); fused traffic is quant_bytes. Return [unfused_bytes, fused_bytes, saved_bytes].

The signature is dequant_fusion_cost(quant_bytes, dequant_bytes).

Hint

Fusing dequant into the consumer avoids materializing the fp16 tensor in memory.

Test cases

case 1

in: [100,400]

expected: [900,100,800]

case 2

in: [0,100]

expected: [200,0,200]

case 3

in: [50,0]

expected: [50,50,0]

case 4

in: [100,100]

expected: [300,100,200]

Starter codePython

def dequant_fusion_cost(quant_bytes, dequant_bytes):
    # Your code here
    pass

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