diff --git a/dsv4/kernels/attention/fmha_qk_verify.cuh b/dsv4/kernels/attention/fmha_qk_verify.cuh
index e420814b..8b58c0f1 100644
--- a/dsv4/kernels/attention/fmha_qk_verify.cuh
+++ b/dsv4/kernels/attention/fmha_qk_verify.cuh
@@ -74,13 +74,15 @@ fmha_qk_verify(
         for (int d = 0; d < HD; d++) {
             dot += bf16_to_f32(sQ[d]) * bf16_to_f32(sK[d]);
         }
-        s_out[0] = dot * scale;  // sanity: scalar QK result
-        s_out[1] = (float)(sQ_smem);  // debug: SMEM address
-        s_out[2] = (float)(sK_smem);  // debug: SMEM address
-        s_out[3] = (float)(tmem_base);  // debug: TMEM base
+        s_out[0] = dot * scale;
+        s_out[1] = (float)(sQ_smem);
+        s_out[2] = (float)(sK_smem);
+        s_out[3] = (float)(tmem_base);
     }
     __syncthreads();
-    return;  // TEMPORARILY SKIP MMA — just check SMEM loads
+    // TEMPORARILY SKIP MMA — just verify SMEM loads
+    if (wid == 0) tmem_dealloc(tmem_base, 128);
+    return;  // ALL threads return — we're just testing SMEM loads
     // MMA: called by ONE lane (elect_one_sync pattern)
     if (tid == 0) printf("[qk] tmem_base=%u sQ_smem=0x%x sK_smem=0x%x\n", tmem_base, sQ_smem, sK_smem);
     __syncthreads();