diff --git a/dsv4/kernels/attention/fmha_6warp_multirow.cuh b/dsv4/kernels/attention/fmha_6warp_multirow.cuh
index 4c95504c..b434987e 100644
--- a/dsv4/kernels/attention/fmha_6warp_multirow.cuh
+++ b/dsv4/kernels/attention/fmha_6warp_multirow.cuh
@@ -164,6 +164,10 @@ fmha_6warp_multirow_kernel(FmhaMultiRowParams params) {
         __syncthreads();
     }
 
+    // Ensure TMEM visibility across all warps after QK GEMM
+    asm volatile("fence.sc.gpu;" ::: "memory");
+    __syncthreads();
+
     // ================================================================
     // SOFTMAX PASS 1: compute row_max for each row
     // ================================================================