How did flash-attn compute attention for cu_seqlens

[zigzagcai]

Hi,

We know that cu_seqlens is for the compute efficiency when we do training over multiple variable-length samples. And the attention mask can be calculated through cu_seqlens. We can cut the original cumulative sequence into a batch sequence and pad the empty positions with zeros. But this approach will waste training efficiency, since computing resources are consumed for meaningless padding tokens.

I am not quite familiar with the implementation details of flash-attn. So, I am just curious about where can I find the implementation or mechanism that how did flash-attn compute attention directly over cumulative sequence and get the separate results?

Thanks!

[tridao]

We just launch parallel work (i.e. thread blocks) to process each attn head of each sequence, and each thread block will figure out the start and end idx of each sequence from cu_seqlens.

[zigzagcai]

We just launch parallel work (i.e. thread blocks) to process each attn head of each sequence, and each thread block will figure out the start and end idx of each sequence from cu_seqlens.

Got it. Thanks for the explanation!

[zigzagcai]

We just launch parallel work (i.e. thread blocks) to process each attn head of each sequence, and each thread block will figure out the start and end idx of each sequence from cu_seqlens.

We observe that the flash API provides fwd/bwd and varlen_fwd/varlen_bwd API to handle inputs wo/w cu_seqlen. Different patterns of inputs are all passed into run_mha_fwd and run_mha_bwd, and ultimately are processed by the flash_fwd_kernel and flash_bwd_kernel within the templates.

In the kernel of flash attention, a structure named BlockInfo is defined to store the offsets of qkv. These offsets are calculated based on cu_seqlen_q and cu_seqlen_k, allowing for the computation of attention on a row-by-row basis (compute_attn_1rowblock). Through BlockInfo, we can specify which thread block will compute attention on which row of qkv.

Hence, each row of qkv would supports variable lengths since the gemm computation is break down into row-by-row computation, thereby preventing the waste of computational resources on meaningless padding tokens.

[2hiTee]

Hi,

We know that cu_seqlens is for the compute efficiency when we do training over multiple variable-length samples. And the attention mask can be calculated through cu_seqlens. We can cut the original cumulative sequence into a batch sequence and pad the empty positions with zeros. But this approach will waste training efficiency, since computing resources are consumed for meaningless padding tokens.

I am not quite familiar with the implementation details of flash-attn. So, I am just curious about where can I find the implementation or mechanism that how did flash-attn compute attention directly over cumulative sequence and get the separate results?

Thanks!

Hi, you mentioned that original code for calculating attention output needs attention_mask, which can be calculated using cu_seqlens. So I am wondering how to convert cu_seqlens to attention_mask, I would be glad if you could give some advice. Thanks!