diff --git a/tests/unit/test_tma_kqk.cu b/tests/unit/test_tma_kqk.cu
index d0d95610..e8acc343 100644
--- a/tests/unit/test_tma_kqk.cu
+++ b/tests/unit/test_tma_kqk.cu
@@ -164,13 +164,13 @@ int main() {
     float* h_out = (float*)malloc(SK * sizeof(float));
     cudaMemcpy(h_out, d_out, SK * sizeof(float), cudaMemcpyDeviceToHost);
 
-    float scale = 1.0f / sqrtf((float)HD);
+    // Reference: raw dot product (MMA is unscaled, scale applied during softmax)
     int fail = 0; float max_rel = 0;
     for (int j = 0; j < SK; j++) {
         float dot = 0;
         for (int d = 0; d < HD; d++)
             dot += bf16_to_f32_host(h_q[d]) * bf16_to_f32_host(h_k[j * HD + d]);
-        float ref = dot * scale;
+        float ref = dot;  // NO scale — MMA output is unscaled
         float got = h_out[j];
         float rel = fabsf(ref) > 1e-4f ? fabsf(got - ref) / fabsf(ref) : fabsf(got - ref);
         if (rel > max_rel) max_rel = rel;