diff --git a/example/ck_tile/18_hstu_attention/hstu_attention_fwd_pipeline.hpp b/example/ck_tile/18_hstu_attention/hstu_attention_fwd_pipeline.hpp index 998fdb595b..86566a2ed0 100644 --- a/example/ck_tile/18_hstu_attention/hstu_attention_fwd_pipeline.hpp +++ b/example/ck_tile/18_hstu_attention/hstu_attention_fwd_pipeline.hpp @@ -303,11 +303,12 @@ struct HstuAttentionFwdPipelineQRKSVS do { static_for<0, k1_loops, 1>{}([&](auto i_k1) { + // load v_tile for current unroll + auto v_tile = load_tile(v_dram_window); + store_tile(k_lds_windows[number{}], tile_elementwise_in(k_element_func, k_tile)); - // load v_tile for current unroll - auto v_tile = load_tile(v_dram_window); move_tile_window(v_dram_window, {0, kK1}); __builtin_amdgcn_sched_barrier(0);