diff --git a/paddle/fluid/memory/detail/memory_block.cc b/paddle/fluid/memory/detail/memory_block.cc
index f34b922b25a0110690671d487f190e1b977a67bb..15e2e856385a14acbbb4717681be5b5181e9e522 100644
--- a/paddle/fluid/memory/detail/memory_block.cc
+++ b/paddle/fluid/memory/detail/memory_block.cc
@@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/memory/detail/memory_block.h"
-#include "paddle/fluid/platform/assert.h"
+#include "paddle/fluid/platform/enforce.h"
 
 namespace paddle {
 namespace memory {
@@ -61,7 +61,7 @@ MemoryBlock* MemoryBlock::right_buddy(const MetadataCache& cache) const {
 
 void MemoryBlock::split(MetadataCache* cache, size_t size) {
   // make sure the split fits
-  PADDLE_ASSERT(total_size(*cache) >= size);
+  PADDLE_ENFORCE_GE(total_size(*cache), size);
 
   // bail out if there is no room for another partition
   if (total_size(*cache) - size <= sizeof(MemoryBlock::Desc)) {
@@ -102,8 +102,8 @@ void MemoryBlock::split(MetadataCache* cache, size_t size) {
 
 void MemoryBlock::merge(MetadataCache* cache, MemoryBlock* right_buddy) {
   // only free blocks can be merged
-  PADDLE_ASSERT(type(*cache) == FREE_CHUNK);
-  PADDLE_ASSERT(right_buddy->type(*cache) == FREE_CHUNK);
+  PADDLE_ENFORCE_EQ(type(*cache), FREE_CHUNK);
+  PADDLE_ENFORCE_EQ(right_buddy->type(*cache), FREE_CHUNK);
 
   auto metadata = cache->load(this);
 
@@ -129,8 +129,8 @@ void MemoryBlock::merge(MetadataCache* cache, MemoryBlock* right_buddy) {
 
 void MemoryBlock::mark_as_free(MetadataCache* cache) {
   // check for double free or corruption
-  PADDLE_ASSERT(type(*cache) != FREE_CHUNK);
-  PADDLE_ASSERT(type(*cache) != INVALID_CHUNK);
+  PADDLE_ENFORCE_NE(type(*cache), FREE_CHUNK);
+  PADDLE_ENFORCE_NE(type(*cache), INVALID_CHUNK);
   set_type(cache, FREE_CHUNK);
 }
 
diff --git a/paddle/fluid/memory/detail/meta_cache.cc b/paddle/fluid/memory/detail/meta_cache.cc
index b86e4f38c42a26e155f276f9b73cbed1d0d83f7d..f04b0c800e3d81419b408843e79ddfe74149a36d 100644
--- a/paddle/fluid/memory/detail/meta_cache.cc
+++ b/paddle/fluid/memory/detail/meta_cache.cc
@@ -14,7 +14,7 @@ limitations under the License. */
 
 #include "glog/logging.h"
 #include "paddle/fluid/memory/detail/memory_block.h"
-#include "paddle/fluid/platform/assert.h"
+#include "paddle/fluid/platform/enforce.h"
 
 namespace paddle {
 namespace memory {
@@ -25,12 +25,12 @@ MetadataCache::MetadataCache(bool uses_gpu) : uses_gpu_(uses_gpu) {}
 MemoryBlock::Desc MetadataCache::load(const MemoryBlock* block) const {
   if (uses_gpu_) {
     auto existing_desc = cache_.find(block);
-    PADDLE_ASSERT(existing_desc->second.check_guards());
+    PADDLE_ENFORCE_EQ(existing_desc->second.check_guards(), true);
     return existing_desc->second;
   } else {
     auto* desc = reinterpret_cast<const MemoryBlock::Desc*>(block);
     VLOG(10) << "Load MemoryBlock::Desc type=" << desc->type;
-    PADDLE_ASSERT(desc->check_guards());
+    PADDLE_ENFORCE_EQ(desc->check_guards(), true);
     return *reinterpret_cast<const MemoryBlock::Desc*>(block);
   }
 }
diff --git a/paddle/fluid/memory/detail/system_allocator.cc b/paddle/fluid/memory/detail/system_allocator.cc
index b0f48c455caf4606a4af63b54b6510f33f68894d..16a67160a3bae8d78bdc1bce819b7824b26dc0ad 100644
--- a/paddle/fluid/memory/detail/system_allocator.cc
+++ b/paddle/fluid/memory/detail/system_allocator.cc
@@ -56,7 +56,7 @@ void* AlignedMalloc(size_t size) {
   PADDLE_ENFORCE_EQ(posix_memalign(&p, alignment, size), 0, "Alloc %ld error!",
                     size);
 #endif
-  PADDLE_ENFORCE(p, "Fail to allocate CPU memory: size = %d .", size);
+  PADDLE_ENFORCE_NOT_NULL(p, "Fail to allocate CPU memory: size = %d .", size);
   return p;
 }
 
@@ -136,15 +136,10 @@ void* GPUAllocator::Alloc(size_t* index, size_t size) {
 
 void GPUAllocator::Free(void* p, size_t size, size_t index) {
   cudaError_t err;
-  if (index == 0) {
-    PADDLE_ASSERT(gpu_alloc_size_ >= size);
-    gpu_alloc_size_ -= size;
-    err = cudaFree(p);
-  } else {
-    PADDLE_ASSERT(fallback_alloc_size_ >= size);
-    fallback_alloc_size_ -= size;
-    err = cudaFreeHost(p);
-  }
+  PADDLE_ENFORCE_EQ(index, 0);
+  PADDLE_ENFORCE_GE(gpu_alloc_size_, size);
+  gpu_alloc_size_ -= size;
+  err = cudaFree(p);
 
   // Purposefully allow cudaErrorCudartUnloading, because
   // that is returned if you ever call cudaFree after the
@@ -194,9 +189,9 @@ void* CUDAPinnedAllocator::Alloc(size_t* index, size_t size) {
 
 void CUDAPinnedAllocator::Free(void* p, size_t size, size_t index) {
   cudaError_t err;
-  PADDLE_ASSERT(index == 1);
+  PADDLE_ENFORCE_EQ(index, 1);
 
-  PADDLE_ASSERT(cuda_pinnd_alloc_size_ >= size);
+  PADDLE_ENFORCE_GE(cuda_pinnd_alloc_size_, size);
   cuda_pinnd_alloc_size_ -= size;
   err = cudaFreeHost(p);
 
diff --git a/paddle/fluid/memory/detail/system_allocator.h b/paddle/fluid/memory/detail/system_allocator.h
index a0386a2dad1bb7faf54197a47ca7a5b6d9488817..42f0f23ec1d5d48276285dcef547a4d51054538b 100644
--- a/paddle/fluid/memory/detail/system_allocator.h
+++ b/paddle/fluid/memory/detail/system_allocator.h
@@ -52,7 +52,6 @@ class GPUAllocator : public SystemAllocator {
 
  private:
   size_t gpu_alloc_size_ = 0;
-  size_t fallback_alloc_size_ = 0;
   int gpu_id_;
 };
 
diff --git a/paddle/fluid/memory/detail/system_allocator_test.cc b/paddle/fluid/memory/detail/system_allocator_test.cc
index 268260142c579ea9301d89fcec1613ce5b0e15a5..65ac12323b099033cd4d2ecb4d161b294961594d 100644
--- a/paddle/fluid/memory/detail/system_allocator_test.cc
+++ b/paddle/fluid/memory/detail/system_allocator_test.cc
@@ -62,4 +62,10 @@ TEST(GPUAllocator, Alloc) {
   TestAllocator(&a, 2048);
   TestAllocator(&a, 0);
 }
+
+TEST(CUDAPinnedAllocator, Alloc) {
+  paddle::memory::detail::CUDAPinnedAllocator a;
+  TestAllocator(&a, 2048);
+  TestAllocator(&a, 0);
+}
 #endif
diff --git a/paddle/fluid/platform/float16_test.cc b/paddle/fluid/platform/float16_test.cc
index 3a937dfaec3acc7c116f0077694e9aee1b379061..f411c3863ffb39a943a863bc2b2ae0f327d51fb9 100644
--- a/paddle/fluid/platform/float16_test.cc
+++ b/paddle/fluid/platform/float16_test.cc
@@ -15,6 +15,7 @@ limitations under the License. */
 #define GLOG_NO_ABBREVIATED_SEVERITIES  // msvc conflict logging with windows.h
 #include "gtest/gtest.h"
 #include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/platform/enforce.h"
 #include "paddle/fluid/platform/init.h"
 
 namespace paddle {
@@ -144,7 +145,7 @@ TEST(float16, lod_tensor_cpu) {
 
 TEST(float16, floating) {
   // compile time assert.
-  PADDLE_ASSERT(std::is_floating_point<float16>::value);
+  PADDLE_ENFORCE_EQ(std::is_floating_point<float16>::value, true);
 }
 
 TEST(float16, print) {
diff --git a/paddle/fluid/platform/float16_test.cu b/paddle/fluid/platform/float16_test.cu
index 14cad927f06551ebbfbf1d166ae250c18591dd6b..bf2038419cb064c508af01a7e0cd085df9ed6d6d 100644
--- a/paddle/fluid/platform/float16_test.cu
+++ b/paddle/fluid/platform/float16_test.cu
@@ -19,6 +19,7 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/tensor_util.h"
+#include "paddle/fluid/platform/enforce.h"
 
 #define ARITHMETIC_KERNEL(op_type, sign)                                 \
   __global__ void op_type(const half* in1, const half* in2, half* out) { \
@@ -260,8 +261,8 @@ TEST(float16, typeid) {
   int b(0);
 
   // compile time assert
-  PADDLE_ASSERT(functor(a) == true);
-  PADDLE_ASSERT(functor2(b) == false);
+  PADDLE_ENFORCE_EQ(functor(a), true);
+  PADDLE_ENFORCE_EQ(functor2(b), false);
 }
 
 // GPU test