Implement FlashAttention-2 in Triton in PyTorch

Expertgpu-systemsauto-graded

Write the tiled, fused attention kernel that computes exact attention in O(n) memory using online softmax in Triton.

Solve it

Check your answer

The grader verifies properties of your implementation, so a correct solution written differently from ours still passes.

pip install torchleet

from torchleet import check
check("flash-attention-triton", flash_attention_pytorch)

Company tags

How these tags are sourced