Implement a blocked/tiled attention forward pass (FlashAttention-style) with running max/s

Hard

Attention Efficiency and Long Context4 / 5

Chapter 034 / 5

Attention Efficiency and Long Context

Implement a blocked/tiled attention forward pass (FlashAttention-style) with running max/s

Hard

Implement a blocked/tiled attention forward pass (FlashAttention-style) with running max/sum and verify it matches naive attention.

Target interface

flash_attention(Q, K, V, block=16, causal=True)

Complete the starter code in the editor. Any correct approach is accepted.

Hints0 / 2

Reference solution

Your own code stays in the editor.

solution.pyPython

Local draft

import torch

def flash_attention(Q, K, V, block=16, causal=True):
    raise NotImplementedError

def naive(Q, K, V, causal=True):
    raise NotImplementedError

Local draft

⌘/Ctrl↵to submit

AI review

Ready when you are

Submit your solution for a score and feedback on correctness, edge cases, and clarity. Any correct approach is accepted.

Chapter 034 / 5

Attention Efficiency and Long Context

Implement a blocked/tiled attention forward pass (FlashAttention-style) with running max/s

Hard

Implement a blocked/tiled attention forward pass (FlashAttention-style) with running max/sum and verify it matches naive attention.

Target interface

flash_attention(Q, K, V, block=16, causal=True)

Complete the starter code in the editor. Any correct approach is accepted.

Hints0 / 2

Reference solution

Your own code stays in the editor.