diff --git a/include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qr_ks_vs_async.hpp b/include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qr_ks_vs_async.hpp index e31d534cd0..87c0997ffc 100644 --- a/include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qr_ks_vs_async.hpp +++ b/include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qr_ks_vs_async.hpp @@ -519,6 +519,7 @@ struct BlockFmhaPipelineQRKSVSAsync constexpr auto p_spans = decltype(p_compute)::get_distributed_spans(); sweep_tile_span(p_spans[number<0>{}], [&](auto idx0) { constexpr auto i_idx = make_tuple(idx0); + // will only record the max value from first tile block, neve update m(i_idx) = m[i_idx] == -numeric::infinity() ? m_local[i_idx] : m[i_idx]; }); (void) m_old;