diff --git a/paddle/fluid/operators/jit/CMakeLists.txt b/paddle/fluid/operators/jit/CMakeLists.txt
index 0f213c5898fa6558af4bb80b38a0420f3acd2638..ced29741253e72a17413de51fb2c24a7fb1257d3 100644
--- a/paddle/fluid/operators/jit/CMakeLists.txt
+++ b/paddle/fluid/operators/jit/CMakeLists.txt
@@ -20,4 +20,6 @@ endif()
 
 cc_library(jit_kernel_helper SRCS ${jit_kernel_cc_srcs} DEPS ${JIT_KERNEL_DEPS})
 cc_test(jit_kernel_test SRCS test.cc DEPS jit_kernel_helper)
-cc_binary(jit_kernel_benchmark SRCS benchmark.cc DEPS jit_kernel_helper)
+if(NOT WIN32)
+    cc_binary(jit_kernel_benchmark SRCS benchmark.cc DEPS jit_kernel_helper)
+endif()
diff --git a/paddle/fluid/operators/jit/more/intrinsic/crf_decoding.cc b/paddle/fluid/operators/jit/more/intrinsic/crf_decoding.cc
index f06892a40775d37ae88bed9f396427d8d7b8e599..f446226c9354899d4fda885c42082d434acdaae6 100644
--- a/paddle/fluid/operators/jit/more/intrinsic/crf_decoding.cc
+++ b/paddle/fluid/operators/jit/more/intrinsic/crf_decoding.cc
@@ -22,11 +22,16 @@ namespace operators {
 namespace jit {
 namespace more {
 namespace intrinsic {
+// Note: intrinsic code is not runtime build.
+// For example, if you build code on AVX, and run on AVX512 it can only use AVX
 
 void CRFDecoding(const int seq_len, const float* x, const float* w,
                  float* alpha, int* track, int tag_num) {
-  const int step_size =
-      platform::MayIUse(platform::avx512f) ? ZMM_FLOAT_BLOCK : YMM_FLOAT_BLOCK;
+#ifdef __AVX512F__
+  const int step_size = ZMM_FLOAT_BLOCK;
+#else
+  const int step_size = YMM_FLOAT_BLOCK;
+#endif
   const int end = tag_num / step_size;
   const int rest = tag_num % step_size;
   /* Setup the alpha initial value.*/
@@ -157,7 +162,12 @@ void CRFDecoding(const int seq_len, const float* x, const float* w,
 }
 
 bool CRFDecodingKernel::UseMe(const int& d) const {
-  return platform::MayIUse(platform::avx);
+#ifdef __AVX512F__
+  constexpr int block = ZMM_FLOAT_BLOCK;
+#else
+  constexpr int block = YMM_FLOAT_BLOCK;
+#endif
+  return platform::MayIUse(platform::avx) && d >= block;
 }
 
 }  // namespace intrinsic
diff --git a/paddle/fluid/operators/jit/more/intrinsic/layer_norm.cc b/paddle/fluid/operators/jit/more/intrinsic/layer_norm.cc
index bac709bc9eaeb8ae4e0da0f0f991ea7f6a3a8004..4b502fbf6bca85e4eb5faea2523f0323c1f20a8f 100644
--- a/paddle/fluid/operators/jit/more/intrinsic/layer_norm.cc
+++ b/paddle/fluid/operators/jit/more/intrinsic/layer_norm.cc
@@ -154,7 +154,7 @@ void LayerNorm(float* x, float* out, float* mean, float* var,
 }
 
 bool LayerNormKernel::UseMe(const int& d) const {
-  return platform::MayIUse(platform::avx);
+  return platform::MayIUse(platform::avx) && d >= YMM_FLOAT_BLOCK;
 }
 
 }  // namespace intrinsic
diff --git a/paddle/fluid/operators/jit/test.cc b/paddle/fluid/operators/jit/test.cc
index 5be7cc5d1c873eab2723600bda84d31c4507ac3f..c9194ec6b1fafea5dfd8a42e0cf3545a8bc8a7b9 100644
--- a/paddle/fluid/operators/jit/test.cc
+++ b/paddle/fluid/operators/jit/test.cc
@@ -37,7 +37,7 @@ template <typename T>
 void ExpectEQ(const T* target, const T* refer, int n) {
   if (std::is_floating_point<T>::value) {
     for (int i = 0; i < n; ++i) {
-      EXPECT_NEAR(target[i], refer[i], 1e-3);
+      EXPECT_NEAR(target[i], refer[i], 1e-5);
     }
   } else {
     for (int i = 0; i < n; ++i) {