diff --git a/tests/unit/test_umma_qk.cu b/tests/unit/test_umma_qk.cu index 82416d16..0097f1ed 100644 --- a/tests/unit/test_umma_qk.cu +++ b/tests/unit/test_umma_qk.cu @@ -78,9 +78,12 @@ test_umma_qk_hd16( } __syncthreads(); - // Zero TMEM — test just 1 store call, no fence + // Zero TMEM — test 2 stores with syncwarp between them if (wid == 0) { tmem_store(tmem_base, 0, 0, 0, 0); + __syncwarp(); + tmem_store(tmem_base + 1, 0, 0, 0, 0); + __syncwarp(); } __syncthreads();