diff --git a/tests/unit/test_umma_qk.cu b/tests/unit/test_umma_qk.cu
index e014e53c..f8337ec8 100644
--- a/tests/unit/test_umma_qk.cu
+++ b/tests/unit/test_umma_qk.cu
@@ -71,14 +71,25 @@ test_umma_qk_hd16(
     __syncthreads();
     uint32_t tmem_base = *sTmemBase;
 
-    // Zero TMEM — skip for now
-    // if (wid == 0) {
-    //     for (int col = 0; col < 32; col++) {
-    //         tmem_store(tmem_base + col, 0, 0, 0, 0);
-    //     }
-    //     tmem_fence_store();
-    // }
-    // __syncthreads();
+    // Debug: write tmem_base to output
+    if (tid == 0) {
+        s_out[138] = 999.0f;  // sentinel before tmem ops
+    }
+    __syncthreads();
+
+    // Zero TMEM — re-enable with debug
+    if (wid == 0) {
+        for (int col = 0; col < 128; col++) {
+            tmem_store(tmem_base + col, 0, 0, 0, 0);
+        }
+        tmem_fence_store();
+    }
+    __syncthreads();
+
+    if (tid == 0) {
+        s_out[138] = (float)tmem_base;  // write tmem_base after tmem ops
+    }
+    __syncthreads();
 
     // ================================================================
     // Load Q and K into SMEM in canonical layout