diff --git a/tests/unit/test_umma_qk_hd64.cu b/tests/unit/test_umma_qk_hd64.cu
index e8d09579..144810d8 100644
--- a/tests/unit/test_umma_qk_hd64.cu
+++ b/tests/unit/test_umma_qk_hd64.cu
@@ -54,10 +54,18 @@ test_umma_qk_hd64(const bf16_t* q, const bf16_t* k,
     }
     __syncthreads();
 
-    // Simple sanity check: thread 0 writes to sQ[0]
-    if (tid == 0) sQ[0] = f32_to_bf16(1.0f);
+    // Sanity check: write to sQ[0]
+    if (tid == 0) {
+        uint16_t one_bf16 = f32_to_bf16(1.0f);
+        sQ[0] = one_bf16;
+    }
     __syncthreads();
-    if (tid == 0) s_out[250] = bf16_to_f32(sQ[0]);  // Should be 1.0
+    if (tid == 0) {
+        uint16_t val = sQ[0];
+        float fval = bf16_to_f32(val);
+        s_out[250] = fval;  // Should be 1.0
+        s_out[251] = (float)one_bf16;  // Raw BF16 bits
+    }
     __syncthreads();
 
     // Write Q (1, hd) to row 0 of sQ in canonical layout