diff --git a/paddle/fluid/operators/fake_quantize_op.cu.h b/paddle/fluid/operators/fake_quantize_op.cu.h
index d85d47f54613132a8d16198e5c6a10e27fb1cc9b..ae448b7ff2c8b96418b2c920cfed1befca05434c 100644
--- a/paddle/fluid/operators/fake_quantize_op.cu.h
+++ b/paddle/fluid/operators/fake_quantize_op.cu.h
@@ -305,7 +305,7 @@ __global__ void ChannelClipAndQuantKernelQuantAxisN(
   int64_t idx = blockDim.x * blockIdx.x + threadIdx.x;
   for (int64_t i = idx; i < n; i += blockDim.x * gridDim.x) {
     T s = scale[(i / quant_stride) % nScale];
-    T inv_s = 1.0 / s;
+    T inv_s = inverse(s);
     T x = in[i];
     T v = x > s ? s : x;
     v = v < -s ? -s : v;