diff --git a/example/ck_tile/01_fmha/utils.hpp b/example/ck_tile/01_fmha/utils.hpp index c26098e39b..70a5844cde 100644 --- a/example/ck_tile/01_fmha/utils.hpp +++ b/example/ck_tile/01_fmha/utils.hpp @@ -150,7 +150,7 @@ decode_seqlen(mode_enum mode, if(1 < batch && use_kvcache) { - // we always use seqlen_k_max in first batch + // to keep the original s_k value, we always use seqlen_k_max in first batch randints(std::next(seqlen_ks.begin()), seqlen_ks.end(), seqlen_k_min,