Skip to content

fix(fmha_v2): enable flash_attention for Q_PAGED_KV regardless of s_kv #3106

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
blake-snc wants to merge 1 commit into
base: main
Choose a base branch
from
+7 −1
Open

fix(fmha_v2): enable flash_attention for Q_PAGED_KV regardless of s_kv #3106

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 7 additions & 1 deletion csrc/fmha_v2_run.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -250,9 +250,15 @@ static inline void determine_launch_params(
launch_params.device_l2_cache_size = props.l2CacheSize;

// threshold for adopting flash attention or warp_specialized kernels.
// For Q_PAGED_KV layouts, flash attention is required regardless of
// seqlen — the paged KV dispatch path only supports flash attention
// kernels, and `s` is the padded max_kv_len (not per-request seq_lens,
// which the kernel predicates on separately). Falling back to the
// non-flash path produces silently wrong output when max_kv_len < 16.
bool const is_paged_kv = input_layout == Attention_input_layout::Q_PAGED_KV;
launch_params.flash_attention =
(data_type == DATA_TYPE_FP16 || data_type == DATA_TYPE_BF16 || data_type == DATA_TYPE_E4M3) &&
(s >= 16 && d >= 16) && !force_non_flash_attention;
(is_paged_kv || (s >= 16 && d >= 16)) && !force_non_flash_attention;

// enable warp_speialized kernels when s >= 512 on hopper
// note that warp_speialized kernels need flash attention + tma
Expand Down
Loading