diff --git a/dsv4/kernels/attention/fmha_6warp_multihead.cuh b/dsv4/kernels/attention/fmha_6warp_multihead.cuh
index 8d5094c7..9d19e4b8 100644
--- a/dsv4/kernels/attention/fmha_6warp_multihead.cuh
+++ b/dsv4/kernels/attention/fmha_6warp_multihead.cuh
@@ -128,11 +128,11 @@ fmha_6warp_multihead_kernel(FmhaParams params) {
     float* s_p_vals = (float*)(sV + V_SUB_SZ);
     float* sOacc = (float*)(((uintptr_t)(s_p_vals + SK_TILE) + 15) & ~(uintptr_t)15);
 
-    // Initialize accumulator
-    if (is_softmax_warp && lane == 0) {
+    // Initialize accumulator (single thread to avoid race)
+    if (tid == 0) {
         *sRowMax = -INFINITY;
         *sRowSum = 0.0f;
-        for (int d = lane; d < HD; d += 32) sOacc[d] = 0.0f;
+        for (int d = 0; d < HD; d++) sOacc[d] = 0.0f;
     }
 
     // TMEM allocation