pytorch / torchtitan

A native PyTorch Library for large model training

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Modify FLOPs in MFU calculation for casual mask when using FlashAttention.

Yuxin-CV opened this issue · comments

Hi, I suggest we modify the FLOPs calculation in the MFU according to the FlashAttention benchmark script.

Specifically, the current calculation for the casual mask can exceed 100% MFU for seq_len = 16k (189 * 2 / 312 = 1.21), which is inaccurate. The FLOPs for the casual mask setting should be divided by 2 when using FlashAttention.

flash2_a100_fwd_bwd_benchmark

There was some past discussion on this (#280).