diff --git a/paddle/fluid/framework/generator.cc b/paddle/fluid/framework/generator.cc
index d00e38784c2c0..9bde9e20b19a0 100644
--- a/paddle/fluid/framework/generator.cc
+++ b/paddle/fluid/framework/generator.cc
@@ -12,67 +12,122 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
+#include "paddle/fluid/framework/generator.h"
+
+#include <glog/logging.h>
+
 #include <deque>
 #include <memory>
 #include <unordered_map>
 #include <unordered_set>
 #include <utility>
 
-#include "paddle/fluid/framework/generator.h"
-
 namespace paddle {
 namespace framework {
 
-std::shared_ptr<Generator> Generator::gen_instance_ = NULL;
+const std::shared_ptr<Generator>& DefaultCPUGenerator() {
+  static auto default_cpu_generator =
+      std::make_shared<Generator>(GetRandomSeed());
+  VLOG(4) << "initial seed: " << default_cpu_generator->GetCurrentSeed()
+          << ", cpu engine: " << default_cpu_generator->GetCPUEngine().get();
+  return default_cpu_generator;
+}
+
+std::shared_ptr<std::mt19937_64> OpDefaultCPUEngine() {
+  static auto op_default_cpu_engine = std::make_shared<std::mt19937_64>();
+  return op_default_cpu_engine;
+}
+
+// NOTE(zhiqiu): there are 3 conditions:
+// (1) op seed is not set and DefaultCPUGenerator is inited, use
+// DefaultCPUGenerator
+// (2) op seed is not set and DefaultCPUGenerator is not inited, use se
+// OpDefaultCPUEngine() and set a radnom seed
+// (3) op seed is set, use OpDefaultCPUEngine() and set the seed
+std::shared_ptr<std::mt19937_64> GetCPURandomEngine(uint64_t seed) {
+  if (DefaultCPUGenerator()->GetIsInitPy() && seed == 0) {
+    VLOG(4) << "Use random engine from generator";
+    return DefaultCPUGenerator()->GetCPUEngine();
+  } else {
+    // NOTE(zhiqiu): creating an engine instance everytime instead of using
+    // OpDefaultCPUEngine(), this is the legacy behavior of random operators.
+    // The benefit is that when runing PE with fixed-seed in multiple thrads,
+    // each thread has their own engine, and doesn't affect each other.
+    //
+    // And we need to measure the determinacy of Generator in PE.
+    auto engine = std::make_shared<std::mt19937_64>();
+    if (seed == 0) {
+      seed = GetRandomSeed();
+      VLOG(4) << "Use default random engine with random seed = " << seed;
+    } else {
+      VLOG(4) << "Use default random engine with fixed random seed = " << seed;
+    }
+    static std::mutex mu_;
+    {
+      std::lock_guard<std::mutex> lock(mu_);
+      engine->seed(seed);
+    }
+    return engine;
+  }
+}
 
-GeneratorState* Generator::GetState() {
-  std::lock_guard<std::mutex> lock(this->mutex);
-  return this->state_.get();
+GeneratorState Generator::GetState() {
+  std::lock_guard<std::mutex> lock(this->mu_);
+  state_.cpu_engine = *engine_;
+  return this->state_;
 }
 
-void Generator::SetState(GeneratorState* state_in) {
-  std::lock_guard<std::mutex> lock(this->mutex);
-  *this->state_ = *state_in;
+void Generator::SetState(const GeneratorState& state) {
+  std::lock_guard<std::mutex> lock(this->mu_);
+  this->state_ = state;
+  this->engine_ = std::make_shared<std::mt19937_64>(state.cpu_engine);
 }
 
 uint64_t Generator::GetCurrentSeed() {
-  std::lock_guard<std::mutex> lock(this->mutex);
-  return this->state_->current_seed;
+  std::lock_guard<std::mutex> lock(this->mu_);
+  return this->state_.current_seed;
 }
 
 uint64_t Generator::Seed() {
-  std::lock_guard<std::mutex> lock(this->mutex);
+  std::lock_guard<std::mutex> lock(this->mu_);
   uint64_t seed;
   std::random_device de;
   seed = ((((uint64_t)de()) << 32) + de()) & 0x1FFFFFFFFFFFFF;
-  this->state_->current_seed = seed;
+  this->state_.current_seed = seed;
   std::seed_seq seq({seed});
-  this->state_->cpu_engine.seed(seq);
+  this->engine_->seed(seq);
 
-  return this->state_->current_seed;
+  return this->state_.current_seed;
 }
 
 void Generator::SetCurrentSeed(uint64_t seed) {
-  std::lock_guard<std::mutex> lock(this->mutex);
-  this->state_->current_seed = uint64_t(seed);
+  std::lock_guard<std::mutex> lock(this->mu_);
+  this->state_.current_seed = seed;
   std::seed_seq seq({seed});
-  this->state_->cpu_engine.seed(seq);
+  this->engine_->seed(seq);
 }
 
-std::mt19937_64& Generator::GetCPUEngine() {
-  std::lock_guard<std::mutex> lock(this->mutex);
-  return this->state_->cpu_engine;
+std::shared_ptr<std::mt19937_64> Generator::GetCPUEngine() {
+  std::lock_guard<std::mutex> lock(this->mu_);
+  return this->engine_;
 }
 
-void Generator::SetCPUEngine(std::mt19937_64 engine) {
-  std::lock_guard<std::mutex> lock(this->mutex);
-  this->state_->cpu_engine = std::mt19937_64(engine);
+void Generator::SetCPUEngine(std::shared_ptr<std::mt19937_64> engine) {
+  std::lock_guard<std::mutex> lock(this->mu_);
+  this->engine_ = engine;
 }
 
 uint64_t Generator::Random64() {
-  std::lock_guard<std::mutex> lock(this->mutex);
-  return this->state_->cpu_engine();
+  std::lock_guard<std::mutex> lock(this->mu_);
+  auto engine = this->engine_;
+  return (*engine)();
+}
+
+void Generator::SetIsInitPy(bool is_init_py) {
+  this->is_init_py_ = is_init_py;
+  VLOG(4) << "SetIsInitPy:" << this->is_init_py_;
 }
+bool Generator::GetIsInitPy() const { return this->is_init_py_; }
 
 }  // namespace framework
 }  // namespace paddle
diff --git a/paddle/fluid/framework/generator.h b/paddle/fluid/framework/generator.h
index 17870782ba72a..82b35f7ad550e 100644
--- a/paddle/fluid/framework/generator.h
+++ b/paddle/fluid/framework/generator.h
@@ -14,7 +14,9 @@ limitations under the License. */
 
 #pragma once
 
+#include <glog/logging.h>
 #include <stdint.h>
+
 #include <atomic>
 #include <deque>
 #include <iostream>  // temp for debug
@@ -27,6 +29,12 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
+static uint64_t GetRandomSeed() {
+  std::random_device rd;
+  // double has 53 bit significant, so limit uint64 to 53 bits
+  return ((((uint64_t)rd()) << 32) + rd()) & 0x1FFFFFFFFFFFFF;
+}
+
 struct GeneratorState {
   int64_t device = -1;
   uint64_t current_seed = 34342423252;
@@ -35,62 +43,67 @@ struct GeneratorState {
 
 struct Generator {
   Generator() {
-    GeneratorState default_gen_state_cpu;
-    default_gen_state_cpu.device = -1;
-    default_gen_state_cpu.current_seed = 34342423252;
-    std::seed_seq seq({34342423252});
-    default_gen_state_cpu.cpu_engine = std::mt19937_64(seq);
-    this->state_ = std::make_shared<GeneratorState>(default_gen_state_cpu);
+    auto seed = GetRandomSeed();
+    std::seed_seq seq({seed});
+    auto engine = std::make_shared<std::mt19937_64>(seq);
+    this->state_.cpu_engine = *engine;
+    this->state_.device = -1;
+    this->state_.current_seed = seed;
+    this->engine_ = engine;
+    VLOG(4) << "initial seed: " << this->state_.current_seed
+            << ", cpu engine: " << &this->state_.cpu_engine;
+  }
+  explicit Generator(uint64_t seed) {
+    std::seed_seq seq({seed});
+    auto engine = std::make_shared<std::mt19937_64>(seq);
+    this->state_.cpu_engine = *engine;
+    this->state_.device = -1;
+    this->state_.current_seed = seed;
+    this->engine_ = engine;
+    VLOG(4) << "initial seed: " << this->state_.current_seed
+            << ", cpu engine: " << &this->state_.cpu_engine;
+    this->is_init_py_ = true;  // TODO(zhiqiu): remove it in future
   }
-  explicit Generator(GeneratorState state_in)
-      : state_{std::make_shared<GeneratorState>(state_in)} {}
-  Generator(const Generator& other)
-      : Generator(other, std::lock_guard<std::mutex>(other.mutex)) {}
+  Generator(const Generator& other) = delete;
 
   // get random state
-  GeneratorState* GetState();
+  GeneratorState GetState();
   // set random state
-  void SetState(GeneratorState* state_in);
+  void SetState(const GeneratorState&);
   // get current seed
   uint64_t GetCurrentSeed();
   // random a seed and get
   uint64_t Seed();
-
   // set seed
   void SetCurrentSeed(uint64_t seed);
   // get cpu engine
-  std::mt19937_64& GetCPUEngine();
+  std::shared_ptr<std::mt19937_64> GetCPUEngine();
   // set cpu engine
-  void SetCPUEngine(std::mt19937_64 engine);
+  void SetCPUEngine(std::shared_ptr<std::mt19937_64>);
 
   uint64_t Random64();
 
-  bool is_init_py = false;
+  void SetIsInitPy(bool);
+  bool GetIsInitPy() const;
 
-  // CPU Generator singleton
-  static std::shared_ptr<Generator> GetInstance() {
-    if (NULL == gen_instance_) {
-      gen_instance_.reset(new paddle::framework::Generator());
-    }
-    return gen_instance_;
-  }
+ private:
+  GeneratorState state_;
+  std::shared_ptr<std::mt19937_64> engine_;
+  mutable std::mutex mu_;
+
+  // NOTE(zhiqiu): is_init_py_ is used to make generator be compatible with
+  // old seed, and it should be removed after all random-related operators
+  // and unittests upgrades to use generator.
+  bool is_init_py_ = false;
+};
 
-  static std::shared_ptr<Generator> GetInstanceX() {
-    if (NULL == gen_instance_) {
-      gen_instance_.reset(new paddle::framework::Generator());
-    }
-    gen_instance_->is_init_py = true;
-    return gen_instance_;
-  }
+// The DefaultCPUGenerator is used in manual_seed()
+const std::shared_ptr<Generator>& DefaultCPUGenerator();
 
- private:
-  static std::shared_ptr<Generator> gen_instance_;
-  std::shared_ptr<GeneratorState> state_;
-  mutable std::mutex mutex;
+// If op seed is set or global is not set, the OpDefaultCPUEngine is used.
+std::shared_ptr<std::mt19937_64> OpDefaultCPUEngine();
 
-  Generator(const Generator& other, const std::lock_guard<std::mutex>&)
-      : state_(std::make_shared<GeneratorState>(*(other.state_))) {}
-};
+std::shared_ptr<std::mt19937_64> GetCPURandomEngine(uint64_t);
 
 }  // namespace framework
 }  // namespace paddle
diff --git a/paddle/fluid/operators/bernoulli_op.cc b/paddle/fluid/operators/bernoulli_op.cc
index c525da5953d76..79c4e2c2bba31 100644
--- a/paddle/fluid/operators/bernoulli_op.cc
+++ b/paddle/fluid/operators/bernoulli_op.cc
@@ -64,11 +64,11 @@ class BernoulliOpKernel<platform::CPUDeviceContext, T>
 
     int64_t size = x->numel();
     std::uniform_real_distribution<T> dist(0.0, 1.0);
-    auto gen_ptr = framework::Generator::GetInstance();
-    std::mt19937_64 &gen_engine = gen_ptr->GetCPUEngine();
+    auto gen_ptr = framework::DefaultCPUGenerator();
+    auto engine = gen_ptr->GetCPUEngine();
 
     for (int64_t i = 0; i < size; ++i) {
-      out_data[i] = BernoulliFunctor(in_data[i], dist(gen_engine));
+      out_data[i] = BernoulliFunctor(in_data[i], dist(*engine));
     }
   }
 };  // namespace operators
diff --git a/paddle/fluid/operators/distributed/large_scale_kv.h b/paddle/fluid/operators/distributed/large_scale_kv.h
index 0d7032e286caa..9e39e68cba779 100644
--- a/paddle/fluid/operators/distributed/large_scale_kv.h
+++ b/paddle/fluid/operators/distributed/large_scale_kv.h
@@ -14,20 +14,19 @@
 
 #pragma once
 
+#include <ThreadPool.h>
 #include <gflags/gflags.h>
 
 #include <functional>
 #include <future>  // NOLINT
 #include <memory>
 #include <string>
+#include <thread>  // NOLINT
 #include <unordered_map>
 #include <unordered_set>
 #include <utility>
 #include <vector>
 
-#include <thread>  // NOLINT
-
-#include <ThreadPool.h>
 #include "paddle/fluid/framework/generator.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/rw_lock.h"
@@ -89,26 +88,17 @@ class UniformInitializer : public Initializer {
     min_ = std::stof(attrs[2]);
     max_ = std::stof(attrs[3]);
 
-    if (seed_ == 0) {
-      seed_ = std::random_device()();
-    }
-
-    random_engine_.seed(seed_);
     dist_ = std::uniform_real_distribution<float>(min_, max_);
+    random_engine_ = framework::GetCPURandomEngine(seed_);
   }
 
-  float GetValue() override {
-    return framework::Generator::GetInstance()->is_init_py
-               ? dist_(framework::Generator::GetInstance()->GetCPUEngine())
-               : dist_(random_engine_);
-    // return dist_(random_engine_);
-  }
+  float GetValue() override { return dist_(*random_engine_); }
 
  private:
   float min_;
   float max_;
 
-  std::minstd_rand random_engine_;
+  std::shared_ptr<std::mt19937_64> random_engine_;
   std::uniform_real_distribution<float> dist_;
 };
 
@@ -139,26 +129,18 @@ class GaussianInitializer : public Initializer {
     mean_ = std::stof(attrs[2]);
     std_ = std::stof(attrs[3]);
 
-    if (seed_ == 0) {
-      seed_ = std::random_device()();
-    }
+    random_engine_ = framework::GetCPURandomEngine(seed_);
 
-    random_engine_.seed(seed_);
     dist_ = std::normal_distribution<float>(mean_, std_);
   }
 
-  float GetValue() override {
-    return framework::Generator::GetInstance()->is_init_py
-               ? dist_(framework::Generator::GetInstance()->GetCPUEngine())
-               : dist_(random_engine_);
-    // return dist_(random_engine_);
-  }
+  float GetValue() override { return dist_(*random_engine_); }
 
  private:
   float std_;
   float mean_;
 
-  std::minstd_rand random_engine_;
+  std::shared_ptr<std::mt19937_64> random_engine_;
   std::normal_distribution<float> dist_;
 };
 
diff --git a/paddle/fluid/operators/dropout_op.h b/paddle/fluid/operators/dropout_op.h
index bce4c7ca19a60..9d9eb4a82a075 100644
--- a/paddle/fluid/operators/dropout_op.h
+++ b/paddle/fluid/operators/dropout_op.h
@@ -55,30 +55,22 @@ class CPUDropoutKernel : public framework::OpKernel<T> {
         std::memset(mask_data, 0, size * sizeof(*mask_data));  // NOLINT
         return;
       }
-
-      bool init_generator_py = framework::Generator::GetInstance()->is_init_py;
-
+      // std::minstd_rand engine;
       // NOTE: fixed seed should only be used in unittest or for debug.
       // Guarantee to use random seed in training.
-      std::random_device rnd;
-      std::minstd_rand engine;
-      int seed_data;
+      int seed_data = 0;
       if (seed) {
         seed_data = *(seed->data<int>());
       } else {
         seed_data =
-            context.Attr<bool>("fix_seed") ? context.Attr<int>("seed") : rnd();
+            context.Attr<bool>("fix_seed") ? context.Attr<int>("seed") : 0;
       }
-      engine.seed(seed_data);
+      auto engine = framework::GetCPURandomEngine(seed_data);
 
       std::uniform_real_distribution<float> dist(0, 1);
 
       for (size_t i = 0; i < size; ++i) {
-        float cur_random =
-            init_generator_py
-                ? dist(framework::Generator::GetInstance()->GetCPUEngine())
-                : dist(engine);
-        if (cur_random < dropout_prob) {
+        if (dist(*engine) < dropout_prob) {
           mask_data[i] = 0;
           y_data[i] = 0;
         } else {
diff --git a/paddle/fluid/operators/gaussian_random_op.cc b/paddle/fluid/operators/gaussian_random_op.cc
index 111d4ad449007..4f128463375b9 100644
--- a/paddle/fluid/operators/gaussian_random_op.cc
+++ b/paddle/fluid/operators/gaussian_random_op.cc
@@ -39,26 +39,14 @@ class CPUGaussianRandomKernel : public framework::OpKernel<T> {
     tensor->Resize(shape);
     int64_t size = tensor->numel();
     T* data = tensor->mutable_data<T>(context.GetPlace());
+    unsigned int seed = static_cast<unsigned int>(context.Attr<int>("seed"));
+    auto engine = framework::GetCPURandomEngine(seed);
 
-    if (framework::Generator::GetInstance()->is_init_py) {
-      std::mt19937_64& gen_engine =
-          framework::Generator::GetInstance()->GetCPUEngine();
-      for (int64_t i = 0; i < size; ++i) {
-        data[i] = dist(gen_engine);
-      }
-    } else {
-      unsigned int seed = static_cast<unsigned int>(context.Attr<int>("seed"));
-      std::minstd_rand engine;
-      if (seed == 0) {
-        seed = std::random_device()();
-      }
-      engine.seed(seed);
-      for (int64_t i = 0; i < size; ++i) {
-        data[i] = dist(engine);
-      }
+    for (int64_t i = 0; i < size; ++i) {
+      data[i] = dist(*engine);
     }
   }
-};
+};  // namespace operators
 
 template <typename T>
 class CPUGaussianRandomBatchSizeLikeKernel : public framework::OpKernel<T> {
diff --git a/paddle/fluid/operators/math/sampler.cc b/paddle/fluid/operators/math/sampler.cc
index 86feaa72d5fa6..a4bdc923eecc3 100644
--- a/paddle/fluid/operators/math/sampler.cc
+++ b/paddle/fluid/operators/math/sampler.cc
@@ -13,11 +13,14 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/math/sampler.h"
+
 #include <glog/logging.h>
+
 #include <iostream>
 #include <queue>
 #include <utility>
 #include <vector>
+
 #include "paddle/fluid/framework/generator.h"
 
 namespace paddle {
@@ -28,22 +31,17 @@ Sampler::~Sampler() {}
 
 UniformSampler::UniformSampler(int64_t range, unsigned int seed)
     : Sampler(range, seed), inv_range_(1.0 / (range + 1)) {
-  random_engine_ = std::make_shared<std::mt19937_64>(seed_);
+  random_engine_ = framework::GetCPURandomEngine(seed_);
   dist_ = std::make_shared<std::uniform_int_distribution<>>(0, range);
 }
 
-int64_t UniformSampler::Sample() const {
-  return framework::Generator::GetInstance()->is_init_py
-             ? (*dist_)(framework::Generator::GetInstance()->GetCPUEngine())
-             : (*dist_)(*random_engine_);
-  // return (*dist_)(*random_engine_);
-}
+int64_t UniformSampler::Sample() const { return (*dist_)(*random_engine_); }
 
 float UniformSampler::Probability(int64_t value) const { return inv_range_; }
 
 LogUniformSampler::LogUniformSampler(int64_t range, unsigned int seed)
     : Sampler(range, seed), log_range_(log(range + 1)) {
-  random_engine_ = std::make_shared<std::mt19937_64>(seed_);
+  random_engine_ = framework::GetCPURandomEngine(seed_);
   dist_ = std::make_shared<std::uniform_real_distribution<>>(0, 1);
 }
 
@@ -52,10 +50,7 @@ int64_t LogUniformSampler::Sample() const {
   // inverse_transform_sampling method
   // More details:
   // https://wanghaoshuang.github.io/2017/11/Log-uniform-distribution-sampler/
-  auto cur_random =
-      framework::Generator::GetInstance()->is_init_py
-          ? (*dist_)(framework::Generator::GetInstance()->GetCPUEngine())
-          : (*dist_)(*random_engine_);
+  auto cur_random = (*dist_)(*random_engine_);
   const int64_t value = static_cast<int64_t>(exp(cur_random * log_range_)) - 1;
   // Mathematically, value should be <= range_, but might not be due to some
   // floating point roundoff, so we mod by range_.
@@ -74,7 +69,7 @@ CustomSampler::CustomSampler(int64_t range, const float *probabilities,
                              const int *alias, const float *alias_probabilities,
                              unsigned int seed)
     : Sampler(range, seed) {
-  random_engine_ = std::make_shared<std::mt19937>(seed_);
+  random_engine_ = framework::GetCPURandomEngine(seed_);
   real_dist_ = std::make_shared<std::uniform_real_distribution<>>(0, 1);
   int_dist_ = std::make_shared<std::uniform_int_distribution<>>(0, range);
 
@@ -84,14 +79,8 @@ CustomSampler::CustomSampler(int64_t range, const float *probabilities,
 }
 
 int64_t CustomSampler::Sample() const {
-  auto index =
-      framework::Generator::GetInstance()->is_init_py
-          ? (*int_dist_)(framework::Generator::GetInstance()->GetCPUEngine())
-          : (*int_dist_)(*random_engine_);
-  auto p =
-      framework::Generator::GetInstance()->is_init_py
-          ? (*real_dist_)(framework::Generator::GetInstance()->GetCPUEngine())
-          : (*real_dist_)(*random_engine_);
+  auto index = (*int_dist_)(*random_engine_);
+  auto p = (*real_dist_)(*random_engine_);
   if (p > alias_probs_[index]) {
     int alias = alias_[index];
 
diff --git a/paddle/fluid/operators/math/sampler.h b/paddle/fluid/operators/math/sampler.h
index 3fa5a7ae336a9..480576ef9dc8c 100644
--- a/paddle/fluid/operators/math/sampler.h
+++ b/paddle/fluid/operators/math/sampler.h
@@ -26,8 +26,8 @@ namespace math {
 // TODO(wanghaoshuang): Support for GPU
 
 /**
-* Sample integers from [0, range).
-*/
+ * Sample integers from [0, range).
+ */
 class Sampler {
  public:
   explicit Sampler(int64_t range, unsigned int seed = 0UL) : range_(range) {
@@ -117,7 +117,7 @@ class CustomSampler : public Sampler {
   const int* alias_;
   const float* probs_;
   const int exceptional_val = -1;
-  std::shared_ptr<std::mt19937> random_engine_;
+  std::shared_ptr<std::mt19937_64> random_engine_;
   std::shared_ptr<std::uniform_real_distribution<>> real_dist_;
   std::shared_ptr<std::uniform_int_distribution<>> int_dist_;
 };
diff --git a/paddle/fluid/operators/mkldnn/gaussian_random_mkldnn_op.cc b/paddle/fluid/operators/mkldnn/gaussian_random_mkldnn_op.cc
index d0ecca78ae8b2..98200caca8cf6 100644
--- a/paddle/fluid/operators/mkldnn/gaussian_random_mkldnn_op.cc
+++ b/paddle/fluid/operators/mkldnn/gaussian_random_mkldnn_op.cc
@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include <string>
+
 #include "paddle/fluid/framework/generator.h"
 #include "paddle/fluid/operators/fill_constant_op.h"
 #include "paddle/fluid/operators/mean_op.h"
@@ -35,23 +36,11 @@ class GaussianMKLDNNKernel : public paddle::framework::OpKernel<T> {
     T* data = tensor->mutable_data<T>(context.GetPlace());
     int64_t size = tensor->numel();
     std::normal_distribution<T> dist(mean, std);
+    unsigned int seed = static_cast<unsigned int>(context.Attr<int>("seed"));
+    auto engine = framework::GetCPURandomEngine(seed);
 
-    if (framework::Generator::GetInstance()->is_init_py) {
-      std::mt19937_64& gen_engine =
-          framework::Generator::GetInstance()->GetCPUEngine();
-      for (int64_t i = 0; i < size; ++i) {
-        data[i] = dist(gen_engine);
-      }
-    } else {
-      unsigned int seed = static_cast<unsigned int>(context.Attr<int>("seed"));
-      std::minstd_rand engine;
-      if (seed == 0) {
-        seed = std::random_device()();
-      }
-      engine.seed(seed);
-      for (int64_t i = 0; i < size; ++i) {
-        data[i] = dist(engine);
-      }
+    for (int64_t i = 0; i < size; ++i) {
+      data[i] = dist(*engine);
     }
 
     tensor->set_layout(DataLayout::kMKLDNN);
diff --git a/paddle/fluid/operators/randint_op.cc b/paddle/fluid/operators/randint_op.cc
index 662fe3bcb3b3b..b3a2e14331955 100644
--- a/paddle/fluid/operators/randint_op.cc
+++ b/paddle/fluid/operators/randint_op.cc
@@ -46,22 +46,11 @@ class CPURandintKernel : public framework::OpKernel<T> {
 
     std::uniform_int_distribution<T> dist(ctx.Attr<int>("low"),
                                           ctx.Attr<int>("high") - 1);
+    unsigned int seed = static_cast<unsigned int>(ctx.Attr<int>("seed"));
+    auto engine = framework::GetCPURandomEngine(seed);
 
-    if (framework::Generator::GetInstance()->is_init_py) {
-      std::mt19937_64& gen_engine =
-          framework::Generator::GetInstance()->GetCPUEngine();
-      for (int64_t i = 0; i < size; ++i) data[i] = dist(gen_engine);
-    } else {
-      unsigned int seed = static_cast<unsigned int>(ctx.Attr<int>("seed"));
-      std::minstd_rand engine;
-      if (seed == 0) {
-        seed = std::random_device()();
-      }
-      engine.seed(seed);
-
-      for (int64_t i = 0; i < size; ++i) {
-        data[i] = dist(engine);
-      }
+    for (int64_t i = 0; i < size; ++i) {
+      data[i] = dist(*engine);
     }
   }
 };
diff --git a/paddle/fluid/operators/randperm_op.h b/paddle/fluid/operators/randperm_op.h
index 0eb028ad80684..02aabb9a7b569 100644
--- a/paddle/fluid/operators/randperm_op.h
+++ b/paddle/fluid/operators/randperm_op.h
@@ -19,6 +19,7 @@ limitations under the License. */
 #include <ctime>
 #include <string>
 #include <vector>
+
 #include "paddle/fluid/framework/generator.h"
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/framework/tensor_util.h"
@@ -29,20 +30,12 @@ namespace operators {
 
 template <typename T>
 static inline void random_permate(T* data_ptr, int num, unsigned int seed) {
+  auto engine = framework::GetCPURandomEngine(seed);
   for (int i = 0; i < num; ++i) {
     data_ptr[i] = static_cast<T>(i);
   }
-  if (framework::Generator::GetInstance()->is_init_py) {
-    std::shuffle(data_ptr, data_ptr + num,
-                 framework::Generator::GetInstance()->GetCPUEngine());
 
-  } else {
-    if (seed == 0) {
-      seed = std::random_device()();
-    }
-    std::srand(seed);
-    std::random_shuffle(data_ptr, data_ptr + num);
-  }
+  std::shuffle(data_ptr, data_ptr + num, *engine);
 }
 
 template <typename DeviceContext, typename T>
diff --git a/paddle/fluid/operators/sampling_id_op.h b/paddle/fluid/operators/sampling_id_op.h
index a09220b1ccd13..9bec08f593afe 100644
--- a/paddle/fluid/operators/sampling_id_op.h
+++ b/paddle/fluid/operators/sampling_id_op.h
@@ -51,20 +51,15 @@ class SamplingIdKernel : public framework::OpKernel<T> {
     framework::TensorToVector(*input, context.device_context(), &ins_vector);
 
     unsigned int seed = static_cast<unsigned int>(context.Attr<int>("seed"));
-    std::minstd_rand engine;
-    if (seed == 0) {
-      seed = std::random_device()();
-    }
-    engine.seed(seed);
+
     std::uniform_real_distribution<T> dist(
         static_cast<T>(context.Attr<float>("min")),
         static_cast<T>(context.Attr<float>("max")));
 
+    auto engine = framework::GetCPURandomEngine(seed);
     std::vector<int64_t> ids(batch_size);
     for (int i = 0; i < batch_size; ++i) {
-      T r = framework::Generator::GetInstance()->is_init_py
-                ? dist(framework::Generator::GetInstance()->GetCPUEngine())
-                : dist(engine);
+      T r = dist(*engine);
       int idx = width - 1;
       for (int j = 0; j < width; ++j) {
         if ((r -= ins_vector[i * width + j]) < 0) {
diff --git a/paddle/fluid/operators/truncated_gaussian_random_op.cc b/paddle/fluid/operators/truncated_gaussian_random_op.cc
index 3aa9ff544af63..419f0f7a2a578 100644
--- a/paddle/fluid/operators/truncated_gaussian_random_op.cc
+++ b/paddle/fluid/operators/truncated_gaussian_random_op.cc
@@ -14,6 +14,7 @@ limitations under the License. */
 
 #include <limits>
 #include <random>
+
 #include "paddle/fluid/framework/generator.h"
 #include "paddle/fluid/framework/op_registry.h"
 
@@ -167,22 +168,10 @@ class CPUTruncatedGaussianRandomKernel : public framework::OpKernel<T> {
     TruncatedNormal<T> truncated_normal(mean, std);
     int64_t size = tensor->numel();
 
-    if (framework::Generator::GetInstance()->is_init_py) {
-      std::mt19937_64& gen_engine =
-          framework::Generator::GetInstance()->GetCPUEngine();
-      for (int64_t i = 0; i < size; ++i) {
-        data[i] = truncated_normal(dist(gen_engine));
-      }
-    } else {
-      unsigned int seed = static_cast<unsigned int>(context.Attr<int>("seed"));
-      std::minstd_rand engine;
-      if (seed == 0) {
-        seed = std::random_device()();
-      }
-      engine.seed(seed);
-      for (int64_t i = 0; i < size; ++i) {
-        data[i] = truncated_normal(dist(engine));
-      }
+    unsigned int seed = static_cast<unsigned int>(context.Attr<int>("seed"));
+    auto engine = framework::GetCPURandomEngine(seed);
+    for (int64_t i = 0; i < size; ++i) {
+      data[i] = truncated_normal(dist(*engine));
     }
   }
 };
diff --git a/paddle/fluid/operators/uniform_random_op.cc b/paddle/fluid/operators/uniform_random_op.cc
index a4487cde27799..9cffe09a33abf 100644
--- a/paddle/fluid/operators/uniform_random_op.cc
+++ b/paddle/fluid/operators/uniform_random_op.cc
@@ -12,7 +12,9 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 #include "paddle/fluid/operators/uniform_random_op.h"
+
 #include <string>
+
 #include "paddle/fluid/framework/generator.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/operator.h"
@@ -62,34 +64,12 @@ class CPUUniformRandomKernel : public framework::OpKernel<T> {
     std::uniform_real_distribution<T> dist(
         static_cast<T>(ctx.Attr<float>("min")),
         static_cast<T>(ctx.Attr<float>("max")));
-    auto gen_ptr = framework::Generator::GetInstance();
-    if (gen_ptr->is_init_py) {
-      std::mt19937_64 &gen_engine = gen_ptr->GetCPUEngine();
-      // auto gen_engine = gen_ptr_->GetCPUEngine();
-      // std::uniform_real_distribution<T> dist(
-      //    static_cast<T>(ctx.Attr<float>("min")),
-      //    static_cast<T>(ctx.Attr<float>("max")));
+    unsigned int seed = static_cast<unsigned int>(ctx.Attr<int>("seed"));
+    auto engine = framework::GetCPURandomEngine(seed);
 
-      for (int64_t i = 0; i < size; ++i) {
-        data[i] = dist(gen_engine);
-      }
-    } else {
-      unsigned int seed = static_cast<unsigned int>(ctx.Attr<int>("seed"));
-      std::minstd_rand engine;
-      if (seed == 0) {
-        seed = std::random_device()();
-      }
-      engine.seed(seed);
-      // std::uniform_real_distribution<T> dist(
-      //    static_cast<T>(ctx.Attr<float>("min")),
-      //    static_cast<T>(ctx.Attr<float>("max")));
-      // int64_t size = tensor->numel();
-      for (int64_t i = 0; i < size; ++i) {
-        data[i] = dist(engine);
-      }
+    for (int64_t i = 0; i < size; ++i) {
+      data[i] = dist(*engine);
     }
-    // std::mt19937_64 &engine = gen_ptr->GetCPUEngine();
-    // auto engine = gen_ptr_->GetCPUEngine();
 
     unsigned int diag_num =
         static_cast<unsigned int>(ctx.Attr<int>("diag_num"));
@@ -139,12 +119,12 @@ class UniformRandomOp : public framework::OperatorWithKernel {
     if (ctx->HasInputs("ShapeTensorList")) {
       // top prority shape
       auto inputs_name = ctx->Inputs("ShapeTensorList");
-      PADDLE_ENFORCE_GT(
-          inputs_name.size(), 0,
-          platform::errors::InvalidArgument(
-              "Input(ShapeTensorList)'size of Op(uniform_random) can't be zero."
-              "Please check the Attr(shape)'s size of"
-              "Op(fluid.layers.uniform_random).)"));
+      PADDLE_ENFORCE_GT(inputs_name.size(), 0,
+                        platform::errors::InvalidArgument(
+                            "Input(ShapeTensorList)'size of "
+                            "Op(uniform_random) can't be zero."
+                            "Please check the Attr(shape)'s size of"
+                            "Op(fluid.layers.uniform_random).)"));
       auto out_dims = std::vector<int>(inputs_name.size(), -1);
       ctx->SetOutputDim("Out", framework::make_ddim(out_dims));
 
diff --git a/paddle/fluid/operators/uniform_random_op.cu b/paddle/fluid/operators/uniform_random_op.cu
index c024bb87b09c0..4df1e0ffeb975 100644
--- a/paddle/fluid/operators/uniform_random_op.cu
+++ b/paddle/fluid/operators/uniform_random_op.cu
@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 #include <thrust/random.h>
 #include <thrust/transform.h>
+
 #include "paddle/fluid/framework/generator.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/operator.h"
@@ -88,15 +89,12 @@ class GPUUniformRandomKernel : public framework::OpKernel<T> {
     }
     T* data = tensor->mutable_data<T>(context.GetPlace());
     unsigned int seed = static_cast<unsigned int>(context.Attr<int>("seed"));
-    if (framework::Generator::GetInstance()->is_init_py) {
-      seed = static_cast<unsigned int>(
-          framework::Generator::GetInstance()->GetCurrentSeed());
-    } else {
-      if (seed == 0) {
-        std::random_device rd;
-        seed = rd();
-      }
+
+    if (seed == 0) {
+      std::random_device rd;
+      seed = rd();
     }
+
     T min = static_cast<T>(context.Attr<float>("min"));
     T max = static_cast<T>(context.Attr<float>("max"));
     unsigned int diag_num =
diff --git a/paddle/fluid/pybind/generator_py.cc b/paddle/fluid/pybind/generator_py.cc
index 3bccd5fb2dd92..90b7f50105253 100644
--- a/paddle/fluid/pybind/generator_py.cc
+++ b/paddle/fluid/pybind/generator_py.cc
@@ -29,23 +29,36 @@ namespace py = pybind11;
 
 namespace paddle {
 namespace pybind {
-void BindGenerator(py::module* m) {
-  py::class_<framework::GeneratorState>(*m, "GeneratorState", "");
-  py::class_<std::mt19937_64>(*m, "mt19937_64", "");
+void BindGenerator(py::module* m_ptr) {
+  auto& m = *m_ptr;
+  py::class_<framework::GeneratorState,
+             std::shared_ptr<framework::GeneratorState>>(m, "GeneratorState")
+      .def("current_seed",
+           [](std::shared_ptr<framework::GeneratorState>& self) {
+             return self->current_seed;
+           });
+  py::class_<std::mt19937_64>(m, "mt19937_64", "");
   py::class_<framework::Generator, std::shared_ptr<framework::Generator>>(
-      *m, "Generator")
-      .def(py::init([]() { return framework::Generator::GetInstanceX(); }),
-           py::return_value_policy::reference)
-      .def("get_state", &framework::Generator::GetState,
-           py::return_value_policy::move)
+      m, "Generator")
+      .def("__init__",
+           [](framework::Generator& self) {
+             new (&self) framework::Generator();
+           })
+      .def("get_state", &framework::Generator::GetState)
       .def("set_state", &framework::Generator::SetState)
-      .def("manual_seed", &framework::Generator::SetCurrentSeed)
+      .def("manual_seed",
+           [](std::shared_ptr<framework::Generator>& self, uint64_t seed) {
+             self->SetCurrentSeed(seed);
+             return self;
+           })
       .def("seed", &framework::Generator::Seed)
       .def("initial_seed", &framework::Generator::GetCurrentSeed)
       .def("random", &framework::Generator::Random64)
-      .def("get_cpu_engine", &framework::Generator::GetCPUEngine,
-           py::return_value_policy::move)
-      .def("set_cpu_engine", &framework::Generator::SetCPUEngine);
+      //  .def("get_cpu_engine", &framework::Generator::GetCPUEngine)
+      //  .def("set_cpu_engine", &framework::Generator::SetCPUEngine)
+      .def_property("_is_init_py", &framework::Generator::GetIsInitPy,
+                    &framework::Generator::SetIsInitPy);
+  m.def("default_cpu_generator", &framework::DefaultCPUGenerator);
 }  // end Generator
 }  // end namespace pybind
-}  // end namespace paddle
+}  // namespace paddle
diff --git a/python/paddle/dataset/tests/test_sentiment.py b/python/paddle/dataset/tests/test_sentiment.py
index bb9830132e987..3540ea06b075e 100644
--- a/python/paddle/dataset/tests/test_sentiment.py
+++ b/python/paddle/dataset/tests/test_sentiment.py
@@ -42,9 +42,11 @@ def test_sort_files(self):
     def test_data_set(self):
         data_set = st.load_sentiment_data()
         last_label = -1
+
         for each in st.test():
             self.assertNotEqual(each[1], last_label)
             last_label = each[1]
+
         self.assertEqual(len(data_set), st.NUM_TOTAL_INSTANCES)
         self.assertEqual(len(list(st.train())), st.NUM_TRAINING_INSTANCES)
         self.assertEqual(
diff --git a/python/paddle/fluid/contrib/tests/test_weight_decay_extend.py b/python/paddle/fluid/contrib/tests/test_weight_decay_extend.py
index 2b331308de5ee..a5f08ca969ac4 100644
--- a/python/paddle/fluid/contrib/tests/test_weight_decay_extend.py
+++ b/python/paddle/fluid/contrib/tests/test_weight_decay_extend.py
@@ -92,9 +92,11 @@ def run_program(self, place, feed_list):
         return param_sum
 
     def check_weight_decay(self, place, model):
+        paddle.manual_seed(1)
+        paddle.framework.random._manual_program_seed(1)
         main_prog = fluid.framework.Program()
         startup_prog = fluid.framework.Program()
-        startup_prog.random_seed = 1
+
         with prog_scope_guard(main_prog=main_prog, startup_prog=startup_prog):
             data = fluid.layers.data(
                 name="words", shape=[1], dtype="int64", lod_level=1)
@@ -113,9 +115,11 @@ def check_weight_decay(self, place, model):
         return param_sum
 
     def check_weight_decay2(self, place, model):
+        paddle.manual_seed(1)
+        paddle.framework.random._manual_program_seed(1)
         main_prog = fluid.framework.Program()
         startup_prog = fluid.framework.Program()
-        startup_prog.random_seed = 1
+
         with prog_scope_guard(main_prog=main_prog, startup_prog=startup_prog):
             data = fluid.layers.data(
                 name="words", shape=[1], dtype="int64", lod_level=1)
diff --git a/python/paddle/fluid/generator.py b/python/paddle/fluid/generator.py
index e11b2e484dce1..98924f801413b 100644
--- a/python/paddle/fluid/generator.py
+++ b/python/paddle/fluid/generator.py
@@ -17,44 +17,28 @@
 
 __all__ = ['Generator']
 
-default_rng_seed_val = 34342423252
 
-
-class Generator(object):
+class Generator(core.Generator):
     """Generator class"""
 
-    def __init__(self, device="CPU"):
-        """init"""
-        self.device = device
-        seed_in = default_rng_seed_val
-        if self.device == "CPU":
-            self.generator = core.Generator()
-            # self.generator.manual_seed(seed_in)
-        else:
-            raise ValueError(
-                "generator class with device %s does not exist, currently only support generator with device 'CPU' "
-                % device)
-
-    def get_state(self):
-        return self.generator.get_state()
-
-    def set_state(self, state):
-        self.generator.set_state(state)
+    def __init__(self, place=None):
+        """
+        Create a generator object which manages the random number generation. ( Experimental Feature )
 
-    def manual_seed(self, seed):
-        self.generator.manual_seed(seed)
+        Parameters:
+            place(CPUPlace|CUDAPinnedPlace|CUDAPlace, optional): The place to allocate Tensor. Can be  
+                CPUPlace, CUDAPinnedPlace, CUDAPlace. Default: None, means global place.
 
-    def seed(self):
-        return self.generator.seed()
+        Returns:
+            Generator: A generator object.
 
-    def initial_seed(self):
-        return self.generator.initial_seed()
-
-    def random(self):
-        return self.generator.random()
-
-    def get_cpu_engine(self):
-        return self.generator.get_cpu_engine()
-
-    def set_cpu_engine(self, cpu_engine):
-        self.generator.set_cpu_engine(cpu_engine)
+        """
+        self.place = place
+        if not place:
+            place = core.CPUPlace()
+        if isinstance(place, core.CPUPlace):
+            super(Generator, self).__init__()
+        else:
+            raise ValueError(
+                "Generator class with %s does is not supported yet, currently only support generator with CPUPlace "
+                % place)
diff --git a/python/paddle/fluid/tests/unittests/dygraph_to_static/test_bmn.py b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_bmn.py
index dd58a49bb55c2..af7e73c41464d 100644
--- a/python/paddle/fluid/tests/unittests/dygraph_to_static/test_bmn.py
+++ b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_bmn.py
@@ -15,6 +15,7 @@
 import math
 import numpy as np
 import unittest
+import paddle
 from paddle.jit import to_static
 import paddle.fluid as fluid
 from paddle.fluid import ParamAttr
@@ -560,8 +561,8 @@ def train_bmn(args, place, to_static):
     loss_data = []
 
     with fluid.dygraph.guard(place):
-        fluid.default_main_program().random_seed = SEED
-        fluid.default_startup_program().random_seed = SEED
+        paddle.manual_seed(SEED)
+        paddle.framework.random._manual_program_seed(SEED)
         global local_random
         local_random = np.random.RandomState(SEED)
 
diff --git a/python/paddle/fluid/tests/unittests/dygraph_to_static/test_lac.py b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_lac.py
index 0e2bac9fa5b5c..4d735b565ddbc 100644
--- a/python/paddle/fluid/tests/unittests/dygraph_to_static/test_lac.py
+++ b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_lac.py
@@ -21,6 +21,7 @@
 import os
 os.environ["CUDA_VISIBLE_DEVICES"] = "2"
 
+import paddle
 import paddle.fluid as fluid
 from paddle.fluid.dygraph import to_variable
 from paddle.fluid.dygraph import Embedding, Linear, GRUUnit
@@ -448,8 +449,8 @@ def do_train(args, to_static):
     place = fluid.CUDAPlace(0) if fluid.is_compiled_with_cuda(
     ) else fluid.CPUPlace()
     with fluid.dygraph.guard(place):
-        fluid.default_startup_program().random_seed = SEED
-        fluid.default_main_program().random_seed = SEED
+        paddle.manual_seed(SEED)
+        paddle.framework.random._manual_program_seed(SEED)
 
         reader = get_random_input_data(args.batch_size, args.vocab_size,
                                        args.num_labels)
diff --git a/python/paddle/fluid/tests/unittests/dygraph_to_static/test_mobile_net.py b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_mobile_net.py
index 5ec3de5871dd6..a377075062b26 100644
--- a/python/paddle/fluid/tests/unittests/dygraph_to_static/test_mobile_net.py
+++ b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_mobile_net.py
@@ -14,6 +14,7 @@
 
 import time
 import numpy as np
+import paddle
 import paddle.fluid as fluid
 from paddle.fluid.initializer import MSRA
 from paddle.fluid.param_attr import ParamAttr
@@ -447,8 +448,8 @@ def train_mobilenet(args, to_static):
     with fluid.dygraph.guard(args.place):
 
         np.random.seed(SEED)
-        fluid.default_startup_program().random_seed = SEED
-        fluid.default_main_program().random_seed = SEED
+        paddle.manual_seed(SEED)
+        paddle.framework.random._manual_program_seed(SEED)
 
         if args.model == "MobileNetV1":
             net = MobileNetV1(class_dim=args.class_dim, scale=1.0)
diff --git a/python/paddle/fluid/tests/unittests/dygraph_to_static/test_ptb_lm.py b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_ptb_lm.py
index 790319936ac01..df2b69297bb4d 100644
--- a/python/paddle/fluid/tests/unittests/dygraph_to_static/test_ptb_lm.py
+++ b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_ptb_lm.py
@@ -19,7 +19,7 @@
 import unittest
 
 import numpy as np
-
+import paddle
 import paddle.fluid as fluid
 from paddle.fluid.dygraph.dygraph_to_static import ProgramTranslator
 from paddle.fluid.dygraph.base import to_variable
@@ -218,8 +218,8 @@ def train(place):
     batch_num = 200
 
     with fluid.dygraph.guard(place):
-        fluid.default_startup_program().random_seed = SEED
-        fluid.default_main_program().random_seed = SEED
+        paddle.manual_seed(SEED)
+        paddle.framework.random._manual_program_seed(SEED)
         ptb_model = PtbModel(
             hidden_size=hidden_size,
             vocab_size=vocab_size,
diff --git a/python/paddle/fluid/tests/unittests/dygraph_to_static/test_reinforcement_learning.py b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_reinforcement_learning.py
index 4813930159744..1d211197ebd48 100644
--- a/python/paddle/fluid/tests/unittests/dygraph_to_static/test_reinforcement_learning.py
+++ b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_reinforcement_learning.py
@@ -16,6 +16,7 @@
 import math
 import itertools
 import numpy as np
+import paddle
 import paddle.fluid as fluid
 import paddle.fluid.dygraph.nn as nn
 from paddle.fluid.dygraph import to_variable, Layer
@@ -64,8 +65,8 @@ def train(args, place, to_static):
     env.seed(SEED)
 
     with fluid.dygraph.guard(place):
-        fluid.default_main_program().random_seed = SEED
-        fluid.default_startup_program().random_seed = SEED
+        paddle.manual_seed(SEED)
+        paddle.framework.random._manual_program_seed(SEED)
         local_random = np.random.RandomState(SEED)
 
         policy = Policy()
diff --git a/python/paddle/fluid/tests/unittests/dygraph_to_static/test_resnet.py b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_resnet.py
index 46eb2b42e9265..6556b2f03bd53 100644
--- a/python/paddle/fluid/tests/unittests/dygraph_to_static/test_resnet.py
+++ b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_resnet.py
@@ -215,8 +215,8 @@ def train(to_static):
     """
     with fluid.dygraph.guard(place):
         np.random.seed(SEED)
-        fluid.default_startup_program().random_seed = SEED
-        fluid.default_main_program().random_seed = SEED
+        paddle.manual_seed(SEED)
+        paddle.framework.random._manual_program_seed(SEED)
 
         train_reader = paddle.batch(
             reader_decorator(paddle.dataset.flowers.train(use_xmap=False)),
diff --git a/python/paddle/fluid/tests/unittests/dygraph_to_static/test_se_resnet.py b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_se_resnet.py
index 30cba78fec19c..38e4d5ad5480b 100644
--- a/python/paddle/fluid/tests/unittests/dygraph_to_static/test_se_resnet.py
+++ b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_se_resnet.py
@@ -331,8 +331,8 @@ def train(train_reader, to_static):
     np.random.seed(SEED)
 
     with fluid.dygraph.guard(place):
-        fluid.default_startup_program().random_seed = SEED
-        fluid.default_main_program().random_seed = SEED
+        paddle.manual_seed(SEED)
+        paddle.framework.random._manual_program_seed(SEED)
         se_resnext = SeResNeXt()
         optimizer = optimizer_setting(train_parameters, se_resnext.parameters())
 
diff --git a/python/paddle/fluid/tests/unittests/dygraph_to_static/test_sentiment.py b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_sentiment.py
index fd5a58be26be4..2aa3396fb7f85 100644
--- a/python/paddle/fluid/tests/unittests/dygraph_to_static/test_sentiment.py
+++ b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_sentiment.py
@@ -15,6 +15,7 @@
 import unittest
 import numpy as np
 
+import paddle
 import paddle.fluid as fluid
 from paddle.fluid.dygraph.nn import Conv2D, Linear, Embedding
 from paddle.fluid.dygraph import to_variable, ProgramTranslator, declarative
@@ -285,8 +286,8 @@ def train(args, to_static):
 
     with fluid.dygraph.guard(place):
         np.random.seed(SEED)
-        fluid.default_startup_program().random_seed = SEED
-        fluid.default_main_program().random_seed = SEED
+        paddle.manual_seed(SEED)
+        paddle.framework.random._manual_program_seed(SEED)
 
         train_reader = fake_data_reader(args.class_num, args.vocab_size,
                                         args.batch_size, args.padding_size)
diff --git a/python/paddle/fluid/tests/unittests/dygraph_to_static/test_simnet.py b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_simnet.py
index 552a6307f3337..14b9ac2e99584 100644
--- a/python/paddle/fluid/tests/unittests/dygraph_to_static/test_simnet.py
+++ b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_simnet.py
@@ -108,8 +108,8 @@ def train(conf_dict, to_static):
         place = fluid.CPUPlace()
 
     with fluid.dygraph.guard(place):
-        fluid.default_startup_program().random_seed = SEED
-        fluid.default_main_program().random_seed = SEED
+        paddle.manual_seed(SEED)
+        paddle.framework.random._manual_program_seed(SEED)
 
         conf_dict['dict_size'] = len(vocab)
         conf_dict['seq_len'] = args.seq_len
diff --git a/python/paddle/fluid/tests/unittests/dygraph_to_static/test_transformer.py b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_transformer.py
index 7aa465949eb70..4fc8d27d30cb8 100644
--- a/python/paddle/fluid/tests/unittests/dygraph_to_static/test_transformer.py
+++ b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_transformer.py
@@ -18,6 +18,7 @@
 import unittest
 
 import numpy as np
+import paddle
 import paddle.fluid as fluid
 
 import transformer_util as util
@@ -31,10 +32,11 @@
 
 
 def train_static(args, batch_generator):
+    paddle.manual_seed(SEED)
+    paddle.framework.random._manual_program_seed(SEED)
     train_prog = fluid.Program()
     startup_prog = fluid.Program()
-    train_prog.random_seed = SEED
-    startup_prog.random_seed = SEED
+
     with fluid.program_guard(train_prog, startup_prog):
         with fluid.unique_name.guard():
             # define input and reader
@@ -128,8 +130,8 @@ def train_static(args, batch_generator):
 def train_dygraph(args, batch_generator):
     with fluid.dygraph.guard(place):
         if SEED is not None:
-            fluid.default_main_program().random_seed = SEED
-            fluid.default_startup_program().random_seed = SEED
+            paddle.manual_seed(SEED)
+            paddle.framework.random._manual_program_seed(SEED)
         # define data loader
         train_loader = fluid.io.DataLoader.from_generator(capacity=10)
         train_loader.set_batch_generator(batch_generator, places=place)
@@ -220,7 +222,8 @@ def train_dygraph(args, batch_generator):
 
 def predict_dygraph(args, batch_generator):
     with fluid.dygraph.guard(place):
-        fluid.default_main_program().random_seed = SEED
+        paddle.manual_seed(SEED)
+        paddle.framework.random._manual_program_seed(SEED)
 
         # define data loader
         test_loader = fluid.io.DataLoader.from_generator(capacity=10)
@@ -291,7 +294,8 @@ def predict_dygraph(args, batch_generator):
 def predict_static(args, batch_generator):
     test_prog = fluid.Program()
     with fluid.program_guard(test_prog):
-        test_prog.random_seed = SEED
+        paddle.manual_seed(SEED)
+        paddle.framework.random._manual_program_seed(SEED)
 
         # define input and reader
         input_field_names = util.encoder_data_input_fields + util.fast_decoder_data_input_fields
diff --git a/python/paddle/fluid/tests/unittests/dygraph_to_static/test_tsm.py b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_tsm.py
index 13a97fb7478db..bedca412157f0 100644
--- a/python/paddle/fluid/tests/unittests/dygraph_to_static/test_tsm.py
+++ b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_tsm.py
@@ -20,7 +20,7 @@
 import sys
 import time
 import unittest
-
+import paddle
 import paddle.fluid as fluid
 from paddle.fluid.dygraph import declarative, ProgramTranslator, to_variable
 from paddle.fluid.dygraph.nn import Conv2D, BatchNorm, Linear, Pool2D
@@ -272,8 +272,8 @@ def train(args, fake_data_reader, to_static):
     random.seed(0)
     np.random.seed(0)
     with fluid.dygraph.guard(place):
-        fluid.default_startup_program().random_seed = 1000
-        fluid.default_main_program().random_seed = 1000
+        paddle.manual_seed(1000)
+        paddle.framework.random._manual_program_seed(1000)
 
         video_model = TSM_ResNet("TSM", train_config, 'Train')
 
diff --git a/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py b/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py
index ec6b81f138321..9c3ed13cbb000 100644
--- a/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py
+++ b/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py
@@ -17,6 +17,7 @@
 import multiprocessing
 import os
 import unittest
+import paddle
 import paddle.fluid as fluid
 import paddle.fluid.core as core
 from paddle.fluid import compiler
@@ -64,10 +65,11 @@ def run_executor(exe, binary, feed, fetch_list):
                 feed_data_reader, FeedDataReader
             ), "feed_data_reader must be type of FeedDataReader"
 
+        paddle.manual_seed(1)
+        paddle.framework.random._manual_program_seed(1)
         main = fluid.Program()
         startup = fluid.Program()
-        startup.random_seed = 1
-        main.random_seed = 1
+
         with fluid.program_guard(main, startup):
             feed_dict, loss = cls.build_model(feed_dict, get_data_from_feeder,
                                               main, method, optimizer)
diff --git a/python/paddle/fluid/tests/unittests/test_buffer_shared_memory_reuse_pass.py b/python/paddle/fluid/tests/unittests/test_buffer_shared_memory_reuse_pass.py
index 671efd8c72155..43d485a0a6d24 100644
--- a/python/paddle/fluid/tests/unittests/test_buffer_shared_memory_reuse_pass.py
+++ b/python/paddle/fluid/tests/unittests/test_buffer_shared_memory_reuse_pass.py
@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import paddle
 import paddle.fluid as fluid
 from paddle.fluid.framework import Parameter
 import numpy as np
@@ -44,10 +45,10 @@ def setUp(self):
 
     def build_program_and_scope(self):
         self.place = fluid.CUDAPlace(0) if self.use_cuda else fluid.CPUPlace()
+        paddle.manual_seed(1)
+        paddle.framework.random._manual_program_seed(1)
         startup_program = fluid.Program()
         main_program = fluid.Program()
-        startup_program.random_seed = 1
-        main_program.random_seed = 1
 
         scope = fluid.Scope()
         with fluid.program_guard(main_program, startup_program):
diff --git a/python/paddle/fluid/tests/unittests/test_compiled_program.py b/python/paddle/fluid/tests/unittests/test_compiled_program.py
index 8430f39578047..751fed2e56126 100644
--- a/python/paddle/fluid/tests/unittests/test_compiled_program.py
+++ b/python/paddle/fluid/tests/unittests/test_compiled_program.py
@@ -16,6 +16,7 @@
 
 import unittest
 import numpy as np
+import paddle
 import paddle.fluid as fluid
 from paddle.fluid import core
 from test_imperative_base import new_program_scope
@@ -29,8 +30,8 @@ def setUp(self):
         self.label = np.random.randint(
             low=0, high=10, size=[16, 1], dtype=np.int64)
         with new_program_scope():
-            fluid.default_startup_program().random_seed = self.seed
-            fluid.default_main_program().random_seed = self.seed
+            paddle.manual_seed(self.seed)
+            paddle.framework.random._manual_program_seed(self.seed)
             place = fluid.CUDAPlace(0) if core.is_compiled_with_cuda(
             ) else fluid.CPUPlace()
             exe = fluid.Executor(place)
@@ -46,8 +47,8 @@ def setUp(self):
 
     def test_compiled_program_base(self):
         with new_program_scope():
-            fluid.default_startup_program().random_seed = self.seed
-            fluid.default_main_program().random_seed = self.seed
+            paddle.manual_seed(self.seed)
+            paddle.framework.random._manual_program_seed(self.seed)
             place = fluid.CUDAPlace(0) if core.is_compiled_with_cuda(
             ) else fluid.CPUPlace()
             exe = fluid.Executor(place)
@@ -64,8 +65,8 @@ def test_compiled_program_base(self):
 
     def test_compiled_program_with_data_parallel(self):
         with new_program_scope():
-            fluid.default_startup_program().random_seed = self.seed
-            fluid.default_main_program().random_seed = self.seed
+            paddle.manual_seed(self.seed)
+            paddle.framework.random._manual_program_seed(self.seed)
             place = fluid.CUDAPlace(0) if core.is_compiled_with_cuda(
             ) else fluid.CPUPlace()
             exe = fluid.Executor(place)
diff --git a/python/paddle/fluid/tests/unittests/test_decoupled_py_reader.py b/python/paddle/fluid/tests/unittests/test_decoupled_py_reader.py
index a16f21c0f97c0..cc0f3745bbf7b 100644
--- a/python/paddle/fluid/tests/unittests/test_decoupled_py_reader.py
+++ b/python/paddle/fluid/tests/unittests/test_decoupled_py_reader.py
@@ -34,10 +34,10 @@ def random_reader():
 
 
 def simple_fc_net(places, use_legacy_py_reader, use_double_buffer):
+    paddle.manual_seed(1)
+    paddle.framework.random._manual_program_seed(1)
     startup_prog = fluid.Program()
     main_prog = fluid.Program()
-    startup_prog.random_seed = 1
-    main_prog.random_seed = 1
 
     with fluid.unique_name.guard():
         with fluid.program_guard(main_prog, startup_prog):
diff --git a/python/paddle/fluid/tests/unittests/test_dygraph_multi_forward.py b/python/paddle/fluid/tests/unittests/test_dygraph_multi_forward.py
index ae4355ec412c8..88b496c1d89e6 100644
--- a/python/paddle/fluid/tests/unittests/test_dygraph_multi_forward.py
+++ b/python/paddle/fluid/tests/unittests/test_dygraph_multi_forward.py
@@ -27,6 +27,8 @@
 from paddle.fluid.dygraph.base import to_variable
 from test_imperative_base import new_program_scope
 
+SEED = 123123111
+
 
 class SimpleImgConvPool(fluid.dygraph.Layer):
     def __init__(self,
@@ -105,12 +107,11 @@ def forward(self, inputs):
 
 class TestDygraphMultiForward(unittest.TestCase):
     def test_mnist_forward_float32(self):
-        seed = 90
         epoch_num = 1
-        with fluid.dygraph.guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
 
+        with fluid.dygraph.guard():
+            paddle.manual_seed(SEED)
+            paddle.framework.random._manual_program_seed(SEED)
             mnist = MNIST()
             sgd = SGDOptimizer(
                 learning_rate=1e-3, parameter_list=mnist.parameters())
@@ -142,9 +143,8 @@ def test_mnist_forward_float32(self):
                             dy_param_init_value[param.name] = param.numpy()
 
         with new_program_scope():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
-
+            paddle.manual_seed(SEED)
+            paddle.framework.random._manual_program_seed(SEED)
             exe = fluid.Executor(fluid.CPUPlace(
             ) if not core.is_compiled_with_cuda() else fluid.CUDAPlace(0))
 
diff --git a/python/paddle/fluid/tests/unittests/test_eager_deletion_padding_rnn.py b/python/paddle/fluid/tests/unittests/test_eager_deletion_padding_rnn.py
index 6c0bb97bf6f14..e0c0277270b40 100644
--- a/python/paddle/fluid/tests/unittests/test_eager_deletion_padding_rnn.py
+++ b/python/paddle/fluid/tests/unittests/test_eager_deletion_padding_rnn.py
@@ -18,6 +18,7 @@
 
 import unittest
 import numpy as np
+import paddle
 import paddle.fluid as fluid
 import paddle.fluid.core as core
 import paddle.fluid.layers as layers
@@ -465,9 +466,9 @@ def set_customed_config(self):
         pass
 
     def _prepare_program(self, config, parallel=True):
+        paddle.manual_seed(config.random_seed)
         self.main_program = fluid.Program()
         self.startup_program = fluid.Program()
-        self.startup_program.random_seed = config.random_seed
         with fluid.program_guard(self.main_program, self.startup_program):
             with fluid.unique_name.guard():
                 res_vars = lm_model(
diff --git a/python/paddle/fluid/tests/unittests/test_embedding_id_stop_gradient.py b/python/paddle/fluid/tests/unittests/test_embedding_id_stop_gradient.py
index 5a562dc14650a..c18b7c5b044e7 100644
--- a/python/paddle/fluid/tests/unittests/test_embedding_id_stop_gradient.py
+++ b/python/paddle/fluid/tests/unittests/test_embedding_id_stop_gradient.py
@@ -13,6 +13,7 @@
 # limitations under the License.
 
 import numpy as np
+import paddle
 import paddle.fluid as fluid
 import six
 import unittest
@@ -37,13 +38,13 @@ def test_check_grad(self):
             self.assertTrue(np.array_equal(grad_value1, grad_value2))
 
     def run_program(self, place, stop_gradient=False):
+        np.random.seed(1)
+        paddle.manual_seed(1)
+        paddle.framework.random._manual_program_seed(1)
+
         startup_program = fluid.Program()
         main_program = fluid.Program()
 
-        np.random.seed(1)
-        startup_program.random_seed = 1
-        main_program.random_seed = 1
-
         scope = fluid.Scope()
         with fluid.program_guard(main_program, startup_program):
             with fluid.scope_guard(scope):
diff --git a/python/paddle/fluid/tests/unittests/test_fc_op.py b/python/paddle/fluid/tests/unittests/test_fc_op.py
index e5a7e6c702aec..ec30cb70c5790 100644
--- a/python/paddle/fluid/tests/unittests/test_fc_op.py
+++ b/python/paddle/fluid/tests/unittests/test_fc_op.py
@@ -13,6 +13,7 @@
 # limitations under the License.
 
 import unittest
+import paddle
 import numpy as np
 from op_test import OpTest
 import paddle.fluid as fluid
@@ -135,31 +136,32 @@ def config(self):
 
 class TestFcOp_NumFlattenDims_NegOne(unittest.TestCase):
     def test_api(self):
-        startup_program = Program()
-        main_program = Program()
-        startup_program.random_seed = SEED
-        main_program.random_seed = SEED
-
-        with program_guard(main_program, startup_program):
-            input = np.random.random([2, 2, 25]).astype("float32")
-            x = fluid.layers.data(
-                name="x",
-                shape=[2, 2, 25],
-                append_batch_size=False,
-                dtype="float32")
-
-            out_1 = fluid.layers.fc(input=x, size=1, num_flatten_dims=-1)
-            out_2 = fluid.layers.fc(input=x, size=1, num_flatten_dims=2)
-
-        place = fluid.CPUPlace() if not core.is_compiled_with_cuda(
-        ) else fluid.CUDAPlace(0)
-        exe = fluid.Executor(place=place)
-        exe.run(startup_program)
-        res_1, res_2 = exe.run(main_program,
-                               feed={"x": input},
-                               fetch_list=[out_1, out_2])
-
-        assert np.array_equal(res_1, res_2)
+        def run_program(num_flatten_dims):
+            paddle.manual_seed(SEED)
+            startup_program = Program()
+            main_program = Program()
+
+            with program_guard(main_program, startup_program):
+                input = np.random.random([2, 2, 25]).astype("float32")
+                x = fluid.layers.data(
+                    name="x",
+                    shape=[2, 2, 25],
+                    append_batch_size=False,
+                    dtype="float32")
+
+                out = fluid.layers.fc(input=x,
+                                      size=1,
+                                      num_flatten_dims=num_flatten_dims)
+
+            place = fluid.CPUPlace() if not core.is_compiled_with_cuda(
+            ) else fluid.CUDAPlace(0)
+            exe = fluid.Executor(place=place)
+            exe.run(startup_program)
+            out = exe.run(main_program, feed={"x": input}, fetch_list=[out])
+
+        res_1 = run_program(-1)
+        res_2 = run_program(2)
+        self.assertTrue(np.array_equal(res_1, res_2))
 
 
 class TestFCOpError(unittest.TestCase):
diff --git a/python/paddle/fluid/tests/unittests/test_fuse_all_reduce_pass.py b/python/paddle/fluid/tests/unittests/test_fuse_all_reduce_pass.py
index 06f8da84a28d2..47671ab3a85e8 100644
--- a/python/paddle/fluid/tests/unittests/test_fuse_all_reduce_pass.py
+++ b/python/paddle/fluid/tests/unittests/test_fuse_all_reduce_pass.py
@@ -11,6 +11,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+
 from simple_nets import simple_fc_net, fc_with_batchnorm, init_data, bow_net
 from fake_reader import fake_imdb_reader
 from parallel_executor_test_base import TestParallelExecutorBase
diff --git a/python/paddle/fluid/tests/unittests/test_fuse_bn_act_pass.py b/python/paddle/fluid/tests/unittests/test_fuse_bn_act_pass.py
index 62eef67a5695f..921dbdbc6d4e1 100644
--- a/python/paddle/fluid/tests/unittests/test_fuse_bn_act_pass.py
+++ b/python/paddle/fluid/tests/unittests/test_fuse_bn_act_pass.py
@@ -19,8 +19,6 @@
 
 class TestFuseBatchNormActPass(unittest.TestCase):
     def build_program(self, main_program, startup_program, use_cuda, seed=1):
-        main_program.random_seed = seed
-        startup_program.random_seed = seed
         with fluid.program_guard(main_program, startup_program):
             x = fluid.layers.data(name='x', shape=[1, 28, 28], dtype='float32')
             y = fluid.layers.data(name="y", shape=[1], dtype='int64')
@@ -59,6 +57,8 @@ def build_program(self, main_program, startup_program, use_cuda, seed=1):
         return x, y, loss
 
     def check(self, place, use_cuda):
+        paddle.manual_seed(1)
+        paddle.framework.random._manual_program_seed(1)
         main_program = fluid.Program()
         startup_program = fluid.Program()
         x, y, loss = self.build_program(main_program, startup_program, use_cuda)
diff --git a/python/paddle/fluid/tests/unittests/test_fuse_optimizer_pass.py b/python/paddle/fluid/tests/unittests/test_fuse_optimizer_pass.py
index b47bcd2a032a3..a22daeedd09e9 100644
--- a/python/paddle/fluid/tests/unittests/test_fuse_optimizer_pass.py
+++ b/python/paddle/fluid/tests/unittests/test_fuse_optimizer_pass.py
@@ -11,6 +11,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+
 from simple_nets import simple_fc_net, fc_with_batchnorm, init_data, bow_net
 from fake_reader import fake_imdb_reader
 from parallel_executor_test_base import TestParallelExecutorBase
diff --git a/python/paddle/fluid/tests/unittests/test_gaussian_random_op.py b/python/paddle/fluid/tests/unittests/test_gaussian_random_op.py
index 6b08c4250f61c..9ab8440407390 100644
--- a/python/paddle/fluid/tests/unittests/test_gaussian_random_op.py
+++ b/python/paddle/fluid/tests/unittests/test_gaussian_random_op.py
@@ -16,7 +16,7 @@
 
 import unittest
 import numpy as np
-
+import paddle
 import paddle.fluid as fluid
 import paddle.fluid.core as core
 from paddle.fluid.op import Operator
@@ -37,6 +37,7 @@ def setUp(self):
             "seed": 10,
             "use_mkldnn": self.use_mkldnn
         }
+        paddle.manual_seed(10)
 
         self.outputs = {'Out': np.zeros((123, 92), dtype='float32')}
 
diff --git a/python/paddle/fluid/tests/unittests/test_generator.py b/python/paddle/fluid/tests/unittests/test_generator.py
index 6cc43d3d54982..8b1f420358d31 100644
--- a/python/paddle/fluid/tests/unittests/test_generator.py
+++ b/python/paddle/fluid/tests/unittests/test_generator.py
@@ -16,6 +16,7 @@
 from __future__ import print_function
 import os
 import unittest
+import paddle
 import paddle.fluid.generator as generator
 import time  # temp for debug
 
@@ -34,10 +35,11 @@ def test_basic_generator(self):
         st = gen.get_state()
         gen.set_state(st)
         gen.random()
-        gen.set_cpu_engine(gen.get_cpu_engine())
 
     def test_basic_generator_error(self):
-        self.assertRaises(ValueError, generator.Generator, device="CUDA")
+        if paddle.fluid.core.is_compiled_with_cuda():
+            self.assertRaises(
+                ValueError, generator.Generator, place=paddle.CUDAPlace(0))
 
 
 if __name__ == "__main__":
diff --git a/python/paddle/fluid/tests/unittests/test_generator_dataloader.py b/python/paddle/fluid/tests/unittests/test_generator_dataloader.py
index 4f0beb8c0dcd5..7c1ff41f7e767 100644
--- a/python/paddle/fluid/tests/unittests/test_generator_dataloader.py
+++ b/python/paddle/fluid/tests/unittests/test_generator_dataloader.py
@@ -35,10 +35,10 @@ def random_reader():
 
 
 def simple_fc_net(places, use_legacy_py_reader, use_double_buffer):
+    paddle.manual_seed(1)
+    paddle.framework.random._manual_program_seed(1)
     startup_prog = fluid.Program()
     main_prog = fluid.Program()
-    startup_prog.random_seed = 1
-    main_prog.random_seed = 1
 
     with fluid.unique_name.guard():
         with fluid.program_guard(main_prog, startup_prog):
diff --git a/python/paddle/fluid/tests/unittests/test_hsigmoid_op.py b/python/paddle/fluid/tests/unittests/test_hsigmoid_op.py
index 5777bb3c6f5e3..5c9867e681524 100644
--- a/python/paddle/fluid/tests/unittests/test_hsigmoid_op.py
+++ b/python/paddle/fluid/tests/unittests/test_hsigmoid_op.py
@@ -16,6 +16,7 @@
 
 import unittest
 import numpy as np
+import paddle
 import paddle.fluid.core as core
 import paddle.fluid as fluid
 from paddle.fluid import Program, program_guard
@@ -266,8 +267,8 @@ def hs_net_conf(self, is_sparse):
 
     def training_test(self, is_sparse):
         with fluid.program_guard(fluid.Program(), fluid.Program()):
+            paddle.manual_seed(1)
             start_up = fluid.default_startup_program()
-            start_up.random_seed = 1  # Fix random seed
             x = np.arange(6).reshape(6)
             path_table = np.array([(1, 2, -1), (1, 2, -1)]).astype('int64')
             path_code = np.array([(1, 0, -1), (0, 0, -1)]).astype('int64')
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_auto_mixed_precision.py b/python/paddle/fluid/tests/unittests/test_imperative_auto_mixed_precision.py
index a4f3858d6fb24..fdf7adbfb45f0 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_auto_mixed_precision.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_auto_mixed_precision.py
@@ -121,6 +121,7 @@ def test_minimize(self):
 
         def run_simple_conv(inp_np, use_scaler=True):
             paddle.manual_seed(10)
+            paddle.framework.random._manual_program_seed(10)
             with fluid.dygraph.guard():
                 model = SimpleConv(
                     num_channels=3,
@@ -204,6 +205,7 @@ def train_resnet(self, enable_amp=True):
 
         with fluid.dygraph.guard():
             paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
 
             resnet = ResNet(use_cudnn=True)
             optimizer = optimizer_setting(
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_deepcf.py b/python/paddle/fluid/tests/unittests/test_imperative_deepcf.py
index af71d9d27b9a3..cc6c2f97a9334 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_deepcf.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_deepcf.py
@@ -206,11 +206,10 @@ def test_deefcf(self):
         else:
             (users_np, items_np, labels_np, num_users, num_items,
              matrix) = get_data()
-
+        paddle.manual_seed(seed)
+        paddle.framework.random._manual_program_seed(seed)
         startup = fluid.Program()
-        startup.random_seed = seed
         main = fluid.Program()
-        main.random_seed = seed
 
         scope = fluid.core.Scope()
         with new_program_scope(main=main, startup=startup, scope=scope):
@@ -244,8 +243,8 @@ def test_deefcf(self):
                     sys.stderr.write('static loss %s\n' % static_loss)
 
         with fluid.dygraph.guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
 
             deepcf = DeepCF(num_users, num_items, matrix)
             adam = fluid.optimizer.AdamOptimizer(
@@ -269,8 +268,8 @@ def test_deefcf(self):
                     sys.stderr.write('dynamic loss: %s %s\n' % (slice, dy_loss))
 
         with fluid.dygraph.guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
 
             deepcf2 = DeepCF(num_users, num_items, matrix)
             adam2 = fluid.optimizer.AdamOptimizer(
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_double_grad.py b/python/paddle/fluid/tests/unittests/test_imperative_double_grad.py
index 227cd5d4acb29..720c9f95c251e 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_double_grad.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_double_grad.py
@@ -312,6 +312,7 @@ def model_f(input):
 
         with fluid.dygraph.guard():
             paddle.manual_seed(123)
+            paddle.framework.random._manual_program_seed(123)
             a = fluid.dygraph.to_variable(value)
             a.stop_gradient = False
 
@@ -328,6 +329,7 @@ def model_f(input):
 
         with fluid.dygraph.guard():
             paddle.manual_seed(123)
+            paddle.framework.random._manual_program_seed(123)
             a = fluid.dygraph.to_variable(value)
             a.stop_gradient = False
 
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_gan.py b/python/paddle/fluid/tests/unittests/test_imperative_gan.py
index 80bdf2ea8a898..b752b439f0fa9 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_gan.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_gan.py
@@ -56,13 +56,11 @@ def forward(self, inputs):
 class TestDygraphGAN(unittest.TestCase):
     def test_gan_float32(self):
         seed = 90
-
+        paddle.manual_seed(1)
+        paddle.framework.random._manual_program_seed(1)
         startup = fluid.Program()
-        startup.random_seed = seed
         discriminate_p = fluid.Program()
         generate_p = fluid.Program()
-        discriminate_p.random_seed = seed
-        generate_p.random_seed = seed
 
         scope = fluid.core.Scope()
         with new_program_scope(
@@ -133,8 +131,8 @@ def test_gan_float32(self):
 
         dy_params = dict()
         with fluid.dygraph.guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(1)
+            paddle.framework.random._manual_program_seed(1)
 
             discriminator = Discriminator()
             generator = Generator()
@@ -177,10 +175,9 @@ def test_gan_float32(self):
 
         dy_params2 = dict()
         with fluid.dygraph.guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
             fluid.set_flags({'FLAGS_sort_sum_gradient': True})
-
+            paddle.manual_seed(1)
+            paddle.framework.random._manual_program_seed(1)
             discriminator2 = Discriminator()
             generator2 = Generator()
             sgd2 = SGDOptimizer(
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_gnn.py b/python/paddle/fluid/tests/unittests/test_imperative_gnn.py
index 01f3c02774698..4db6f2d0da1d5 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_gnn.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_gnn.py
@@ -61,12 +61,10 @@ def forward(self, x, adj):
 
 class TestDygraphGNN(unittest.TestCase):
     def test_gnn_float32(self):
-        seed = 90
-
+        paddle.manual_seed(90)
+        paddle.framework.random._manual_program_seed(90)
         startup = fluid.Program()
-        startup.random_seed = seed
         main = fluid.Program()
-        main.random_seed = seed
 
         scope = fluid.core.Scope()
         with new_program_scope(main=main, startup=startup, scope=scope):
@@ -114,8 +112,8 @@ def test_gnn_float32(self):
                 scope.find_var(model.gc.weight.name).get_tensor())
 
         with fluid.dygraph.guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(90)
+            paddle.framework.random._manual_program_seed(90)
 
             features = np.ones([1, 100, 50], dtype=np.float32)
             # Use selected rows when it's supported.
@@ -140,8 +138,8 @@ def test_gnn_float32(self):
             model_gc_weight_value = model.gc.weight.numpy()
 
         with fluid.dygraph.guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(90)
+            paddle.framework.random._manual_program_seed(90)
 
             features2 = np.ones([1, 100, 50], dtype=np.float32)
             # Use selected rows when it's supported.
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_lod_tensor_to_selected_rows.py b/python/paddle/fluid/tests/unittests/test_imperative_lod_tensor_to_selected_rows.py
index 6349d71760934..f0fea2d7eb75c 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_lod_tensor_to_selected_rows.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_lod_tensor_to_selected_rows.py
@@ -15,6 +15,7 @@
 from __future__ import print_function
 
 import unittest
+import paddle
 import paddle.fluid as fluid
 import paddle.fluid.core as core
 from paddle.fluid.dygraph.nn import Embedding
@@ -94,8 +95,8 @@ def simple_net_float32(self, is_sparse, dtype):
 
             for is_sort_sum_gradient in [True, False]:
                 with fluid.dygraph.guard(place):
-                    fluid.default_startup_program().random_seed = seed
-                    fluid.default_main_program().random_seed = seed
+                    paddle.manual_seed(seed)
+                    paddle.framework.random._manual_program_seed(seed)
 
                     simple_net = SimpleNet(
                         hidden_size=hidden_size,
@@ -139,8 +140,8 @@ def simple_net_float32(self, is_sparse, dtype):
                     dy_loss_value = dy_loss.numpy()
 
                 with new_program_scope():
-                    fluid.default_startup_program().random_seed = seed
-                    fluid.default_main_program().random_seed = seed
+                    paddle.manual_seed(seed)
+                    paddle.framework.random._manual_program_seed(seed)
 
                     simple_net = SimpleNet(
                         hidden_size=hidden_size,
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_ocr_attention_model.py b/python/paddle/fluid/tests/unittests/test_imperative_ocr_attention_model.py
index 499a4311f6e17..5400b785d2929 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_ocr_attention_model.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_ocr_attention_model.py
@@ -16,6 +16,7 @@
 import unittest
 import numpy as np
 import six
+import paddle
 import paddle.fluid as fluid
 from paddle.fluid import core
 from paddle.fluid.dygraph.nn import Conv2D, Pool2D, Linear, BatchNorm, Embedding, GRUUnit
@@ -401,9 +402,9 @@ def test_while_op(self):
                 dtype='int64').reshape([1, Config.max_length])))
 
         with fluid.dygraph.guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
             fluid.set_flags({'FLAGS_sort_sum_gradient': True})
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
             ocr_attention = OCRAttention()
 
             if Config.learning_rate_decay == "piecewise_decay":
@@ -453,8 +454,8 @@ def test_while_op(self):
                         dy_param_value[param.name] = param.numpy()
 
         with new_program_scope():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
             exe = fluid.Executor(fluid.CPUPlace(
             ) if not core.is_compiled_with_cuda() else fluid.CUDAPlace(0))
             ocr_attention = OCRAttention()
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_optimizer.py b/python/paddle/fluid/tests/unittests/test_imperative_optimizer.py
index a7783afc5cff3..7876675bcc6a1 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_optimizer.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_optimizer.py
@@ -74,8 +74,8 @@ def _check_exception(self, exception_message, place=None):
 
         with fluid.dygraph.guard(place):
             try:
-                fluid.default_startup_program().random_seed = seed
-                fluid.default_main_program().random_seed = seed
+                paddle.manual_seed(seed)
+                paddle.framework.random._manual_program_seed(seed)
                 mlp = MLP()
                 optimizer = self.get_optimizer_dygraph(
                     parameter_list=mlp.parameters())
@@ -91,8 +91,8 @@ def _check_mlp(self, place=None):
             ) else fluid.CUDAPlace(0)
 
         with fluid.dygraph.guard(place):
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
 
             mlp = MLP()
             optimizer = self.get_optimizer_dygraph(
@@ -132,8 +132,8 @@ def _check_mlp(self, place=None):
                     dy_param_value[param.name] = param.numpy()
 
         with new_program_scope():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
 
             if place == None:
                 place = fluid.CPUPlace() if not core.is_compiled_with_cuda(
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_optimizer_v2.py b/python/paddle/fluid/tests/unittests/test_imperative_optimizer_v2.py
index 9f75c92b185ed..ca8fb4c220f5e 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_optimizer_v2.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_optimizer_v2.py
@@ -74,8 +74,8 @@ def _check_exception(self, exception_message, place=None):
 
         with fluid.dygraph.guard(place):
             try:
-                fluid.default_startup_program().random_seed = seed
-                fluid.default_main_program().random_seed = seed
+                paddle.manual_seed(seed)
+                paddle.framework.random._manual_program_seed(seed)
                 mlp = MLP()
                 optimizer = self.get_optimizer_dygraph(
                     parameter_list=mlp.parameters())
@@ -91,8 +91,8 @@ def _check_mlp(self, place=None):
             ) else fluid.CUDAPlace(0)
 
         with fluid.dygraph.guard(place):
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
 
             mlp = MLP()
             optimizer = self.get_optimizer_dygraph(
@@ -132,8 +132,8 @@ def _check_mlp(self, place=None):
                     dy_param_value[param.name] = param.numpy()
 
         with new_program_scope():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
 
             if place == None:
                 place = fluid.CPUPlace() if not core.is_compiled_with_cuda(
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_ptb_rnn.py b/python/paddle/fluid/tests/unittests/test_imperative_ptb_rnn.py
index bd629f5f4a69a..fa23ff8e7c29f 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_ptb_rnn.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_ptb_rnn.py
@@ -15,6 +15,7 @@
 from __future__ import print_function
 
 import unittest
+import paddle
 import paddle.fluid as fluid
 import paddle.fluid.core as core
 from paddle.fluid.dygraph.nn import Embedding
@@ -225,8 +226,8 @@ def ptb_rnn_cpu_float32(self, is_sparse):
         traced_layer = None
 
         with fluid.dygraph.guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
             # TODO: marsyang1993 Change seed to
             ptb_model = PtbModel(
                 hidden_size=hidden_size,
@@ -293,8 +294,8 @@ def ptb_rnn_cpu_float32(self, is_sparse):
             dy_last_hidden_value = last_hidden.numpy()
 
         with new_program_scope():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
             ptb_model = PtbModel(
                 hidden_size=hidden_size,
                 vocab_size=vocab_size,
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_ptb_rnn_sorted_gradient.py b/python/paddle/fluid/tests/unittests/test_imperative_ptb_rnn_sorted_gradient.py
index 526c1706e2d08..0487f8dd9a640 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_ptb_rnn_sorted_gradient.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_ptb_rnn_sorted_gradient.py
@@ -15,6 +15,7 @@
 from __future__ import print_function
 
 import unittest
+import paddle
 import paddle.fluid as fluid
 import paddle.fluid.core as core
 from paddle.fluid.dygraph.nn import Embedding
@@ -43,9 +44,10 @@ def ptb_rnn_sort_gradient_cpu_float32(self, is_sparse):
         batch_num = 200
 
         with fluid.dygraph.guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
             fluid.set_flags({'FLAGS_sort_sum_gradient': True})
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
+
             # TODO: marsyang1993 Change seed to
             ptb_model = PtbModel(
                 hidden_size=hidden_size,
@@ -93,8 +95,9 @@ def ptb_rnn_sort_gradient_cpu_float32(self, is_sparse):
             dy_last_hidden_value = last_hidden.numpy()
 
         with new_program_scope():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
+
             ptb_model = PtbModel(
                 hidden_size=hidden_size,
                 vocab_size=vocab_size,
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_reinforcement.py b/python/paddle/fluid/tests/unittests/test_imperative_reinforcement.py
index 735ec4d3f1ea8..0076c61e58407 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_reinforcement.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_reinforcement.py
@@ -64,8 +64,8 @@ def test_mnist_float32(self):
         mask = np.array(mask_list).astype("float32")
 
         with fluid.dygraph.guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
 
             policy = Policy(input_size=4)
 
@@ -105,8 +105,8 @@ def test_mnist_float32(self):
                 dy_param_value[param.name] = param.numpy()
 
         with new_program_scope():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
 
             exe = fluid.Executor(fluid.CPUPlace(
             ) if not core.is_compiled_with_cuda() else fluid.CUDAPlace(0))
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_resnet.py b/python/paddle/fluid/tests/unittests/test_imperative_resnet.py
index 815437072fde2..e8a2298c17d00 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_resnet.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_resnet.py
@@ -251,8 +251,8 @@ def test_resnet_float32(self):
         traced_layer = None
 
         with fluid.dygraph.guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
 
             resnet = ResNet()
             optimizer = optimizer_setting(
@@ -334,8 +334,8 @@ def test_resnet_float32(self):
                     dy_param_value[param.name] = param.numpy()
 
         with new_program_scope():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
 
             exe = fluid.Executor(fluid.CPUPlace(
             ) if not core.is_compiled_with_cuda() else fluid.CUDAPlace(0))
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_resnet_sorted_gradient.py b/python/paddle/fluid/tests/unittests/test_imperative_resnet_sorted_gradient.py
index d26d6f25aa8ff..13b12da3318ca 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_resnet_sorted_gradient.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_resnet_sorted_gradient.py
@@ -77,9 +77,10 @@ def test_resnet_sort_gradient_float32(self):
         batch_size = train_parameters["batch_size"]
         batch_num = 10
         with fluid.dygraph.guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
             fluid.set_flags({'FLAGS_sort_sum_gradient': True})
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
+
             resnet = ResNet()
             optimizer = optimizer_setting(
                 train_parameters, parameter_list=resnet.parameters())
@@ -136,8 +137,8 @@ def test_resnet_sort_gradient_float32(self):
                     dy_param_value[param.name] = param.numpy()
 
         with new_program_scope():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
 
             exe = fluid.Executor(fluid.CPUPlace(
             ) if not core.is_compiled_with_cuda() else fluid.CUDAPlace(0))
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_save_load.py b/python/paddle/fluid/tests/unittests/test_imperative_save_load.py
index eb9dc926c8207..e20f5d6a72278 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_save_load.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_save_load.py
@@ -219,8 +219,8 @@ def setUp(self):
         batch_num = 200
 
         with fluid.dygraph.guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
             # TODO: marsyang1993 Change seed to
             ptb_model = PtbModel(
                 hidden_size=hidden_size,
@@ -305,8 +305,8 @@ def testLoadAndSetVarBase(self):
         batch_num = 200
 
         with fluid.dygraph.guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
             # TODO: marsyang1993 Change seed to
             ptb_model = PtbModel(
                 hidden_size=hidden_size,
@@ -414,8 +414,8 @@ def testSetVariable(self):
         batch_num = 200
 
         with fluid.dygraph.guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
             # TODO: marsyang1993 Change seed to
             ptb_model = PtbModel(
                 hidden_size=hidden_size,
@@ -521,8 +521,8 @@ def testSetNumpy(self):
         batch_num = 200
 
         with fluid.dygraph.guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
             # TODO: marsyang1993 Change seed to
             ptb_model = PtbModel(
                 hidden_size=hidden_size,
@@ -634,8 +634,6 @@ def testSetVariableBeforeTrain(self):
         batch_num = 200
 
         with fluid.dygraph.guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
             # TODO: marsyang1993 Change seed to
             ptb_model = PtbModel(
                 hidden_size=hidden_size,
@@ -713,8 +711,8 @@ def testLoadAndSetVarBaseBeforeTrain(self):
         batch_num = 200
 
         with fluid.dygraph.guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
             # TODO: marsyang1993 Change seed to
             ptb_model = PtbModel(
                 hidden_size=hidden_size,
@@ -804,9 +802,10 @@ def testSetNumpyBeforeTrain(self):
         batch_num = 200
 
         with fluid.dygraph.guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
             # TODO: marsyang1993 Change seed to
+
             ptb_model = PtbModel(
                 hidden_size=hidden_size,
                 vocab_size=vocab_size,
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_save_load_v2.py b/python/paddle/fluid/tests/unittests/test_imperative_save_load_v2.py
index 4ab35a21aff43..5e3e8f224e01f 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_save_load_v2.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_save_load_v2.py
@@ -219,8 +219,8 @@ def setUp(self):
         batch_num = 200
 
         with fluid.dygraph.guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
             # TODO: marsyang1993 Change seed to
             ptb_model = PtbModel(
                 hidden_size=hidden_size,
@@ -305,8 +305,8 @@ def testLoadAndSetVarBase(self):
         batch_num = 200
 
         with fluid.dygraph.guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
             # TODO: marsyang1993 Change seed to
             ptb_model = PtbModel(
                 hidden_size=hidden_size,
@@ -414,8 +414,8 @@ def testSetVariable(self):
         batch_num = 200
 
         with fluid.dygraph.guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
             # TODO: marsyang1993 Change seed to
             ptb_model = PtbModel(
                 hidden_size=hidden_size,
@@ -521,8 +521,8 @@ def testSetNumpy(self):
         batch_num = 200
 
         with fluid.dygraph.guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
             # TODO: marsyang1993 Change seed to
             ptb_model = PtbModel(
                 hidden_size=hidden_size,
@@ -634,8 +634,8 @@ def testSetVariableBeforeTrain(self):
         batch_num = 200
 
         with fluid.dygraph.guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
             # TODO: marsyang1993 Change seed to
             ptb_model = PtbModel(
                 hidden_size=hidden_size,
@@ -713,8 +713,8 @@ def testLoadAndSetVarBaseBeforeTrain(self):
         batch_num = 200
 
         with fluid.dygraph.guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
             # TODO: marsyang1993 Change seed to
             ptb_model = PtbModel(
                 hidden_size=hidden_size,
@@ -804,8 +804,8 @@ def testSetNumpyBeforeTrain(self):
         batch_num = 200
 
         with fluid.dygraph.guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
             # TODO: marsyang1993 Change seed to
             ptb_model = PtbModel(
                 hidden_size=hidden_size,
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_se_resnext.py b/python/paddle/fluid/tests/unittests/test_imperative_se_resnext.py
index 283addaf6283a..a04e1e4e5aafe 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_se_resnext.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_se_resnext.py
@@ -308,8 +308,8 @@ def test_se_resnext_float32(self):
         batch_num = 1
         epoch_num = 1
         with fluid.dygraph.guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
 
             se_resnext = SeResNeXt()
             optimizer = optimizer_setting(
@@ -367,8 +367,8 @@ def test_se_resnext_float32(self):
                         dy_param_value[param.name] = param.numpy()
 
         with new_program_scope():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
 
             exe = fluid.Executor(fluid.CPUPlace(
             ) if not core.is_compiled_with_cuda() else fluid.CUDAPlace(0))
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_selected_rows_to_lod_tensor.py b/python/paddle/fluid/tests/unittests/test_imperative_selected_rows_to_lod_tensor.py
index 3765cb784d652..794f59e48507e 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_selected_rows_to_lod_tensor.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_selected_rows_to_lod_tensor.py
@@ -15,6 +15,7 @@
 from __future__ import print_function
 
 import unittest
+import paddle
 import paddle.fluid as fluid
 import paddle.fluid.core as core
 from paddle.fluid.dygraph.nn import Embedding
@@ -101,8 +102,8 @@ def simple_net_float(self, is_sparse, dtype):
             for is_sort_sum_gradient in [True, False]:
                 traced_layer = None
                 with fluid.dygraph.guard(place):
-                    fluid.default_startup_program().random_seed = seed
-                    fluid.default_main_program().random_seed = seed
+                    paddle.manual_seed(seed)
+                    paddle.framework.random._manual_program_seed(seed)
 
                     simple_net = SimpleNet(
                         hidden_size=hidden_size,
@@ -145,8 +146,8 @@ def simple_net_float(self, is_sparse, dtype):
                     dy_loss_value = dy_loss.numpy()
 
                 with new_program_scope():
-                    fluid.default_startup_program().random_seed = seed
-                    fluid.default_main_program().random_seed = seed
+                    paddle.manual_seed(seed)
+                    paddle.framework.random._manual_program_seed(seed)
 
                     simple_net = SimpleNet(
                         hidden_size=hidden_size,
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_star_gan_with_gradient_penalty.py b/python/paddle/fluid/tests/unittests/test_imperative_star_gan_with_gradient_penalty.py
index d603a7d6ca0de..e94157fa047ee 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_star_gan_with_gradient_penalty.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_star_gan_with_gradient_penalty.py
@@ -468,8 +468,8 @@ def build_optimizer(layer, cfg, loss=None):
 
 class DyGraphTrainModel(object):
     def __init__(self, cfg):
-        fluid.default_startup_program().random_seed = cfg.seed
-        fluid.default_main_program().random_seed = cfg.seed
+        paddle.manual_seed(1)
+        paddle.framework.random._manual_program_seed(1)
 
         self.generator = Generator(cfg)
         self.discriminator = Discriminator(cfg)
@@ -529,12 +529,12 @@ def create_data_layer():
                 shape=[None, cfg.c_dim], dtype='float32', name='label_trg')
             return image_real, label_org, label_trg
 
+        paddle.manual_seed(cfg.seed)
+        paddle.framework.random._manual_program_seed(cfg.seed)
         self.gen_program = fluid.Program()
         gen_startup_program = fluid.Program()
 
         with fluid.program_guard(self.gen_program, gen_startup_program):
-            self.gen_program.random_seed = cfg.seed
-            gen_startup_program.random_seed = cfg.seed
             with fluid.unique_name.guard():
                 image_real, label_org, label_trg = create_data_layer()
                 generator = Generator(cfg)
@@ -546,8 +546,6 @@ def create_data_layer():
         self.dis_program = fluid.Program()
         dis_startup_program = fluid.Program()
         with fluid.program_guard(self.dis_program, dis_startup_program):
-            self.dis_program.random_seed = cfg.seed
-            dis_startup_program.random_seed = cfg.seed
             with fluid.unique_name.guard():
                 image_real, label_org, label_trg = create_data_layer()
                 generator = Generator(cfg)
diff --git a/python/paddle/fluid/tests/unittests/test_imperative_transformer_sorted_gradient.py b/python/paddle/fluid/tests/unittests/test_imperative_transformer_sorted_gradient.py
index c59ce44ec96a8..9f58ef881e4e4 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative_transformer_sorted_gradient.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative_transformer_sorted_gradient.py
@@ -15,6 +15,7 @@
 from __future__ import print_function
 
 import unittest
+import paddle
 import paddle.fluid as fluid
 from paddle.fluid import Embedding, LayerNorm, Linear, Layer
 from paddle.fluid.dygraph import to_variable, guard
@@ -949,9 +950,9 @@ def transformer_sort_gradient_float32(self, is_sparse):
         seed = 90
 
         with guard():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
             fluid.set_flags({'FLAGS_sort_sum_gradient': True})
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
             transformer = TransFormer(
                 ModelHyperParams.src_vocab_size,
                 ModelHyperParams.trg_vocab_size,
@@ -1034,8 +1035,8 @@ def transformer_sort_gradient_float32(self, is_sparse):
             dy_token_num_value = dy_token_num.numpy()
 
         with new_program_scope():
-            fluid.default_startup_program().random_seed = seed
-            fluid.default_main_program().random_seed = seed
+            paddle.manual_seed(seed)
+            paddle.framework.random._manual_program_seed(seed)
             transformer = TransFormer(
                 ModelHyperParams.src_vocab_size,
                 ModelHyperParams.trg_vocab_size,
diff --git a/python/paddle/fluid/tests/unittests/test_ir_memory_optimize_ifelse_op.py b/python/paddle/fluid/tests/unittests/test_ir_memory_optimize_ifelse_op.py
index c5228fcf12274..eaa7e711a29c7 100644
--- a/python/paddle/fluid/tests/unittests/test_ir_memory_optimize_ifelse_op.py
+++ b/python/paddle/fluid/tests/unittests/test_ir_memory_optimize_ifelse_op.py
@@ -37,10 +37,10 @@ def check_network_convergence(self,
                                   use_cuda=True,
                                   use_mem_opt=False,
                                   iter_num=5):
+        paddle.manual_seed(100)
+        paddle.framework.random._manual_program_seed(100)
         prog = Program()
         startup_prog = Program()
-        prog.random_seed = 100
-        startup_prog.random_seed = 100
         with program_guard(prog, startup_prog):
             image = layers.data(name='x', shape=[784], dtype='float32')
 
diff --git a/python/paddle/fluid/tests/unittests/test_jit_save_load.py b/python/paddle/fluid/tests/unittests/test_jit_save_load.py
index 2b79659b9c695..87b6e76a6d0ab 100644
--- a/python/paddle/fluid/tests/unittests/test_jit_save_load.py
+++ b/python/paddle/fluid/tests/unittests/test_jit_save_load.py
@@ -18,7 +18,7 @@
 import pickle
 import unittest
 import numpy as np
-
+import paddle
 from paddle.static import InputSpec
 import paddle.fluid as fluid
 from paddle.fluid.dygraph import Linear
@@ -80,7 +80,7 @@ def forward(self, x):
 
 def train(layer, input_size=784, label_size=1):
     # create optimizer
-    adam = fluid.optimizer.SGDOptimizer(
+    sgd = fluid.optimizer.SGDOptimizer(
         learning_rate=0.01, parameter_list=layer.parameters())
     # create data loader
     train_loader = fluid.io.DataLoader.from_generator(capacity=5)
@@ -97,7 +97,7 @@ def train(layer, input_size=784, label_size=1):
         avg_loss = fluid.layers.mean(loss)
 
         avg_loss.backward()
-        adam.minimize(avg_loss)
+        sgd.minimize(avg_loss)
         layer.clear_gradients()
     return [img], layer, avg_loss
 
@@ -108,7 +108,8 @@ def setUp(self):
         # enable dygraph mode
         fluid.enable_dygraph()
         # config seed
-        fluid.default_main_program().random_seed = SEED
+        paddle.manual_seed(SEED)
+        paddle.framework.random._manual_program_seed(SEED)
 
     def train_and_save_model(self, model_path=None, configs=None):
         layer = LinearNet(784, 1)
@@ -149,8 +150,8 @@ def load_and_finetune(self, train_layer, load_train_layer):
         train_layer.train()
         load_train_layer.train()
         # train & compare
-        _, _, train_loss = train(train_layer)
-        _, _, load_train_loss = train(load_train_layer)
+        img0, _, train_loss = train(train_layer)
+        img1, _, load_train_loss = train(load_train_layer)
         self.assertTrue(
             np.array_equal(train_loss.numpy(), load_train_loss.numpy()))
 
@@ -293,7 +294,8 @@ def setUp(self):
         # enable dygraph mode
         fluid.enable_dygraph()
         # config seed
-        fluid.default_main_program().random_seed = SEED
+        paddle.manual_seed(SEED)
+        paddle.framework.random._manual_program_seed(SEED)
 
     def basic_save_load(self, layer, model_path, configs):
         # 1. train & save
@@ -385,7 +387,8 @@ def setUp(self):
         # enable dygraph mode
         fluid.enable_dygraph()
         # config seed
-        fluid.default_main_program().random_seed = SEED
+        paddle.manual_seed(SEED)
+        paddle.framework.random._manual_program_seed(SEED)
         # train and save base model
         self.train_and_save_orig_model()
 
@@ -426,7 +429,8 @@ def setUp(self):
         # enable dygraph mode
         fluid.enable_dygraph()
         # config seed
-        fluid.default_main_program().random_seed = SEED
+        paddle.manual_seed(SEED)
+        paddle.framework.random._manual_program_seed(SEED)
 
     def train_and_save(self):
         train_layer = LinearNetReturnHidden(8, 8)
diff --git a/python/paddle/fluid/tests/unittests/test_layers.py b/python/paddle/fluid/tests/unittests/test_layers.py
index 1992a3bb39807..eb79a80da99fa 100644
--- a/python/paddle/fluid/tests/unittests/test_layers.py
+++ b/python/paddle/fluid/tests/unittests/test_layers.py
@@ -57,8 +57,8 @@ def _get_place(self, force_to_use_cpu=False):
     @contextlib.contextmanager
     def static_graph(self):
         with new_program_scope():
-            fluid.default_startup_program().random_seed = self.seed
-            fluid.default_main_program().random_seed = self.seed
+            paddle.manual_seed(self.seed)
+            paddle.framework.random._manual_program_seed(self.seed)
             yield
 
     def get_static_graph_result(self,
@@ -77,8 +77,8 @@ def get_static_graph_result(self,
     def dynamic_graph(self, force_to_use_cpu=False):
         with fluid.dygraph.guard(
                 self._get_place(force_to_use_cpu=force_to_use_cpu)):
-            fluid.default_startup_program().random_seed = self.seed
-            fluid.default_main_program().random_seed = self.seed
+            paddle.manual_seed(self.seed)
+            paddle.framework.random._manual_program_seed(self.seed)
             yield
 
 
@@ -1034,7 +1034,7 @@ def test_nce(self):
             static_rlt2 = self.get_static_graph_result(
                 feed=feed_dict, fetch_list=[nce_loss2])[0]
 
-        with self.dynamic_graph(force_to_use_cpu=True):
+        with self.dynamic_graph():
             words = []
             for i in range(window_size):
                 words.append(base.to_variable(inp_word[i]))
@@ -1070,7 +1070,7 @@ def test_nce(self):
         self.assertTrue(np.allclose(static_rlt2, static_rlt))
         self.assertTrue(np.allclose(dy_rlt_value, static_rlt))
 
-        with self.dynamic_graph(force_to_use_cpu=True):
+        with self.dynamic_graph():
             custom_weight = np.random.randn(dict_size, 128).astype("float32")
             weight_attr = fluid.ParamAttr(
                 initializer=fluid.initializer.NumpyArrayInitializer(
@@ -1996,13 +1996,13 @@ def test_accuracy(self):
             exe = fluid.Executor(place)
 
             exe.run(fluid.default_startup_program())
-            x = np.random.rand(3, 32, 32).astype("float32")
-            y = np.array([[1], [0], [1]])
+            # x = np.random.rand(3, 32, 32).astype("float32")
+            # y = np.array([[1], [0], [1]])
             static_out = exe.run(feed={"input": x,
                                        "label": y},
                                  fetch_list=result[0])
 
-        with self.dynamic_graph():
+        with self.dynamic_graph(force_to_use_cpu=True):
             data = base.to_variable(x)
             label = base.to_variable(y)
             fc_out = fluid.layers.fc(data, size=10)
diff --git a/python/paddle/fluid/tests/unittests/test_manual_seed.py b/python/paddle/fluid/tests/unittests/test_manual_seed.py
index 747026622e465..a1d6eb915ce78 100644
--- a/python/paddle/fluid/tests/unittests/test_manual_seed.py
+++ b/python/paddle/fluid/tests/unittests/test_manual_seed.py
@@ -15,30 +15,33 @@
 from __future__ import print_function
 import unittest
 
+import paddle
 import paddle.fluid as fluid
 from paddle.framework import manual_seed
 from paddle.fluid.framework import Program, default_main_program, default_startup_program
+import numpy as np
 
 
 class TestManualSeed(unittest.TestCase):
     def test_manual_seed(self):
-        local_program = Program()
-        local_main_prog = default_main_program()
-        local_start_prog = default_startup_program()
-
-        self.assertEqual(0, local_program.random_seed)
-        self.assertEqual(0, local_main_prog.random_seed)
-        self.assertEqual(0, local_start_prog.random_seed)
-
-        manual_seed(102)
-        global_program1 = Program()
-        global_program2 = Program()
-        global_main_prog = default_main_program()
-        global_start_prog = default_startup_program()
-        self.assertEqual(102, global_program1.random_seed)
-        self.assertEqual(102, global_program2.random_seed)
-        self.assertEqual(102, global_main_prog.random_seed)
-        self.assertEqual(102, global_start_prog.random_seed)
+        fluid.enable_dygraph()
+
+        gen = paddle.manual_seed(12312321111)
+        x = fluid.layers.gaussian_random([10], dtype="float32")
+        st1 = gen.get_state()
+        x1 = fluid.layers.gaussian_random([10], dtype="float32")
+        gen.set_state(st1)
+        x2 = fluid.layers.gaussian_random([10], dtype="float32")
+        gen.manual_seed(12312321111)
+        x3 = fluid.layers.gaussian_random([10], dtype="float32")
+        x_np = x.numpy()
+        x1_np = x1.numpy()
+        x2_np = x2.numpy()
+        x3_np = x3.numpy()
+
+        if not fluid.core.is_compiled_with_cuda():
+            self.assertTrue(np.allclose(x1_np, x2_np))
+            self.assertTrue(np.allclose(x_np, x3_np))
 
 
 if __name__ == '__main__':
diff --git a/python/paddle/fluid/tests/unittests/test_py_func_op.py b/python/paddle/fluid/tests/unittests/test_py_func_op.py
index 6045f2d713627..32d8f73552f71 100644
--- a/python/paddle/fluid/tests/unittests/test_py_func_op.py
+++ b/python/paddle/fluid/tests/unittests/test_py_func_op.py
@@ -147,10 +147,8 @@ def test_main(use_cuda, use_py_func_op, use_parallel_executor):
 
     with fluid.program_guard(fluid.Program(), fluid.Program()):
         with fluid.scope_guard(fluid.core.Scope()):
-            fluid.default_main_program().random_seed = 1
-            fluid.default_startup_program().random_seed = 1
+            gen = paddle.manual_seed(1)
             np.random.seed(1)
-
             img = fluid.layers.data(name='image', shape=[784], dtype='float32')
             label = fluid.layers.data(name='label', shape=[1], dtype='int64')
             loss = simple_fc_net(img, label, use_py_func_op)
@@ -189,17 +187,17 @@ def setUp(self):
         self.use_parallel_executor = False
 
     def test_loss_diff(self):
-        losses = []
         for use_cuda in [True, False]:
+            losses = []
             for use_py_func_op in [True, False]:
                 L = test_main(use_cuda, use_py_func_op,
                               self.use_parallel_executor)
                 if L is not None:
                     losses.append(L)
 
-        for idx in six.moves.range(len(losses) - 1):
-            max_diff = np.max(np.abs(losses[idx] - losses[0]))
-            self.assertAlmostEqual(max_diff, 0, delta=1e-3)
+                for idx in six.moves.range(len(losses) - 1):
+                    max_diff = np.max(np.abs(losses[idx] - losses[0]))
+                    self.assertAlmostEqual(max_diff, 0, delta=1e-3)
 
 
 class TestPyFuncOpUseParallelExecutor(TestPyFuncOpUseExecutor):
diff --git a/python/paddle/fluid/tests/unittests/test_random_seed.py b/python/paddle/fluid/tests/unittests/test_random_seed.py
index 2933abe46c1b8..343508bf619b6 100644
--- a/python/paddle/fluid/tests/unittests/test_random_seed.py
+++ b/python/paddle/fluid/tests/unittests/test_random_seed.py
@@ -26,27 +26,31 @@
 
 
 class TestGeneratorSeed(unittest.TestCase):
-    """
-    Test cases for cpu generator seed.
-    """
+    #     """
+    #     Test cases for cpu generator seed.
+    #     """
 
     def test_generator_uniform_random_dygraph(self):
         """Test Generator seed."""
-        gen = generator.Generator()
 
         fluid.enable_dygraph()
 
-        gen.manual_seed(12312321111)
+        gen = paddle.manual_seed(12312321111)
         x = fluid.layers.uniform_random([10], dtype="float32", min=0.0, max=1.0)
+
         st1 = gen.get_state()
         x1 = fluid.layers.uniform_random(
             [10], dtype="float32", min=0.0, max=1.0)
+
         gen.set_state(st1)
+        print(gen.get_state())
         x2 = fluid.layers.uniform_random(
             [10], dtype="float32", min=0.0, max=1.0)
-        gen.manual_seed(12312321111)
+
+        paddle.manual_seed(12312321111)
         x3 = fluid.layers.uniform_random(
             [10], dtype="float32", min=0.0, max=1.0)
+
         x_np = x.numpy()
         x1_np = x1.numpy()
         x2_np = x2.numpy()
@@ -57,11 +61,9 @@ def test_generator_uniform_random_dygraph(self):
             self.assertTrue(np.allclose(x_np, x3_np))
 
     def test_generator_uniform_random_static(self):
-
         fluid.disable_dygraph()
 
-        gen = generator.Generator()
-        gen.manual_seed(123123143)
+        gen = paddle.manual_seed(123123143)
 
         startup_program = fluid.Program()
         train_program = fluid.Program()
@@ -93,11 +95,9 @@ def test_generator_uniform_random_static(self):
                 self.assertTrue(not np.allclose(out1_res2, out1_res1))
 
     def test_gen_dropout_dygraph(self):
-        gen = generator.Generator()
-
         fluid.enable_dygraph()
 
-        gen.manual_seed(111111111)
+        gen = paddle.manual_seed(111111111)
         st = gen.get_state()
         # x = np.arange(1,101).reshape(2,50).astype("float32")
         x = fluid.layers.uniform_random(
@@ -110,8 +110,7 @@ def test_gen_dropout_dygraph(self):
         y1 = fluid.layers.dropout(x1, 0.5)
         y_np = y.numpy()
         y1_np = y1.numpy()
-        #print(y_np)
-        #print(y1_np)
+
         if not core.is_compiled_with_cuda():
             print(">>>>>>> dropout dygraph >>>>>>>")
             self.assertTrue(np.allclose(y_np, y1_np))
@@ -119,8 +118,7 @@ def test_gen_dropout_dygraph(self):
     def test_gen_dropout_static(self):
         fluid.disable_dygraph()
 
-        gen = generator.Generator()
-        gen.manual_seed(123123143)
+        gen = paddle.manual_seed(123123143)
 
         startup_program = fluid.Program()
         train_program = fluid.Program()
@@ -137,19 +135,16 @@ def test_gen_dropout_static(self):
             out2 = exe.run(train_program, feed={}, fetch_list=[y_1])
         out1_np = np.array(out1[0])
         out2_np = np.array(out2[0])
-        # print(out1_np)
-        # print(out2_np)
+
         if not core.is_compiled_with_cuda():
             print(">>>>>>> dropout static >>>>>>>")
             self.assertTrue(np.allclose(out1_np, out2_np))
 
     def test_generator_gaussian_random_dygraph(self):
         """Test Generator seed."""
-        gen = generator.Generator()
-
         fluid.enable_dygraph()
 
-        gen.manual_seed(12312321111)
+        gen = paddle.manual_seed(12312321111)
         x = fluid.layers.gaussian_random([10], dtype="float32")
         st1 = gen.get_state()
         x1 = fluid.layers.gaussian_random([10], dtype="float32")
@@ -168,11 +163,9 @@ def test_generator_gaussian_random_dygraph(self):
             self.assertTrue(np.allclose(x_np, x3_np))
 
     def test_generator_gaussian_random_static(self):
-
         fluid.disable_dygraph()
 
-        gen = generator.Generator()
-        gen.manual_seed(123123143)
+        gen = paddle.manual_seed(123123143)
 
         startup_program = fluid.Program()
         train_program = fluid.Program()
@@ -210,7 +203,7 @@ def test_generator_randint_dygraph(self):
 
         fluid.enable_dygraph()
 
-        gen.manual_seed(12312321111)
+        gen = paddle.manual_seed(12312321111)
         x = paddle.randint(low=10, shape=[10], dtype="int32")
         st1 = gen.get_state()
         x1 = paddle.randint(low=10, shape=[10], dtype="int32")
@@ -228,12 +221,64 @@ def test_generator_randint_dygraph(self):
             self.assertTrue(np.allclose(x1_np, x2_np))
             self.assertTrue(np.allclose(x_np, x3_np))
 
-    def test_generator_ranint_static(self):
+    def test_generator_uniform_random_static(self):
+        fluid.disable_dygraph()
+
+        gen = paddle.manual_seed(123123143)
+
+        startup_program = fluid.Program()
+        train_program = fluid.Program()
+        with fluid.program_guard(train_program, startup_program):
+            # example 1:
+            # attr shape is a list which doesn't contain tensor Variable.
+            result_1 = fluid.layers.uniform_random(shape=[3, 4])
+            result_2 = fluid.layers.uniform_random(shape=[3, 4])
+
+            exe = fluid.Executor(fluid.CPUPlace())
+            exe.run(startup_program)
+            out1 = exe.run(train_program,
+                           feed={},
+                           fetch_list=[result_1, result_2])
+            #gen.set_state(cur_state)
+            gen.manual_seed(123123143)
+            out2 = exe.run(train_program,
+                           feed={},
+                           fetch_list=[result_1, result_2])
 
+            out1_res1 = np.array(out1[0])
+            out1_res2 = np.array(out1[1])
+            out2_res1 = np.array(out2[0])
+            out2_res2 = np.array(out2[1])
+
+            if not core.is_compiled_with_cuda():
+                self.assertTrue(np.allclose(out1_res1, out2_res1))
+                self.assertTrue(np.allclose(out1_res2, out2_res2))
+                self.assertTrue(not np.allclose(out1_res2, out1_res1))
+
+    def test_generator_randint_dygraph(self):
+        """Test Generator seed."""
+        fluid.enable_dygraph()
+
+        gen = paddle.manual_seed(12312321111)
+        x = paddle.randint(low=1)
+        st1 = gen.get_state()
+        x1 = paddle.randint(low=1)
+        gen.set_state(st1)
+        x2 = paddle.randint(low=1)
+        gen.manual_seed(12312321111)
+        x3 = paddle.randint(low=1)
+        x_np = x.numpy()
+        x1_np = x1.numpy()
+        x2_np = x2.numpy()
+        x3_np = x3.numpy()
+        if not core.is_compiled_with_cuda():
+            self.assertTrue(np.allclose(x1_np, x2_np))
+            self.assertTrue(np.allclose(x_np, x3_np))
+
+    def test_generator_ranint_static(self):
         fluid.disable_dygraph()
 
-        gen = generator.Generator()
-        gen.manual_seed(123123143)
+        gen = paddle.manual_seed(123123143)
 
         startup_program = fluid.Program()
         train_program = fluid.Program()
@@ -267,11 +312,10 @@ def test_generator_ranint_static(self):
 
     def test_generator_randperm_dygraph(self):
         """Test Generator seed."""
-        gen = generator.Generator()
 
         fluid.enable_dygraph()
 
-        gen.manual_seed(12312321111)
+        gen = paddle.manual_seed(12312321111)
         x = paddle.randperm(10)
         st1 = gen.get_state()
         x1 = paddle.randperm(10)
@@ -284,9 +328,6 @@ def test_generator_randperm_dygraph(self):
         x2_np = x2.numpy()
         x3_np = x3.numpy()
 
-        # print("## {}".format(x1_np))
-        # print("## {}".format(x2_np))
-
         if not core.is_compiled_with_cuda():
             print(">>>>>>> randperm dygraph >>>>>>>")
             self.assertTrue(np.allclose(x1_np, x2_np))
@@ -296,8 +337,7 @@ def test_generator_randperm_static(self):
 
         fluid.disable_dygraph()
 
-        gen = generator.Generator()
-        gen.manual_seed(123123143)
+        paddle.manual_seed(123123143)
 
         startup_program = fluid.Program()
         train_program = fluid.Program()
@@ -312,8 +352,8 @@ def test_generator_randperm_static(self):
             out1 = exe.run(train_program,
                            feed={},
                            fetch_list=[result_1, result_2])
-            #gen.set_state(cur_state)
-            gen.manual_seed(123123143)
+
+            paddle.manual_seed(123123143)
             out2 = exe.run(train_program,
                            feed={},
                            fetch_list=[result_1, result_2])
@@ -331,7 +371,7 @@ def test_generator_randperm_static(self):
 
     def test_generator_sampling_id_dygraph(self):
         """Test Generator seed."""
-        gen = generator.Generator()
+        gen = paddle.manual_seed(12312321111)
 
         fluid.enable_dygraph()
 
@@ -339,14 +379,17 @@ def test_generator_sampling_id_dygraph(self):
         x = fluid.layers.uniform_random(
             [10, 10], dtype="float32", min=0.0, max=1.0)
         y = fluid.layers.sampling_id(x)
+
         st1 = gen.get_state()
         x1 = fluid.layers.uniform_random(
             [10, 10], dtype="float32", min=0.0, max=1.0)
         y1 = fluid.layers.sampling_id(x)
+
         gen.set_state(st1)
         x2 = fluid.layers.uniform_random(
             [10, 10], dtype="float32", min=0.0, max=1.0)
         y2 = fluid.layers.sampling_id(x)
+
         gen.manual_seed(12312321111)
         x3 = fluid.layers.uniform_random(
             [10, 10], dtype="float32", min=0.0, max=1.0)
@@ -357,9 +400,6 @@ def test_generator_sampling_id_dygraph(self):
         x2_np = y2.numpy()
         x3_np = y3.numpy()
 
-        print("## {}".format(x1_np))
-        print("## {}".format(x2_np))
-
         if not core.is_compiled_with_cuda():
             print(">>>>>>> sampling id dygraph >>>>>>>")
             self.assertTrue(np.allclose(x1_np, x2_np))
@@ -369,8 +409,7 @@ def test_generator_randperm_static(self):
 
         fluid.disable_dygraph()
 
-        gen = generator.Generator()
-        gen.manual_seed(123123143)
+        paddle.manual_seed(123123143)
 
         startup_program = fluid.Program()
         train_program = fluid.Program()
@@ -386,8 +425,8 @@ def test_generator_randperm_static(self):
             out1 = exe.run(train_program,
                            feed={},
                            fetch_list=[result_1, result_2])
-            #gen.set_state(cur_state)
-            gen.manual_seed(123123143)
+
+            paddle.manual_seed(123123143)
             out2 = exe.run(train_program,
                            feed={},
                            fetch_list=[result_1, result_2])
@@ -406,8 +445,7 @@ def test_generator_randperm_static(self):
     def test_gen_TruncatedNormal_initializer(self):
         fluid.disable_dygraph()
 
-        gen = generator.Generator()
-        gen.manual_seed(123123143)
+        gen = paddle.manual_seed(123123143)
         cur_state = gen.get_state()
 
         startup_program = fluid.Program()
@@ -432,9 +470,7 @@ def test_gen_TruncatedNormal_initializer(self):
             out1 = exe.run(train_program,
                            feed={},
                            fetch_list=[result_1, result_2])
-            #gen.set_state(cur_state)
 
-        #gen.set_state(cur_state)    
         gen.manual_seed(123123143)
         with fluid.program_guard(train_program, startup_program):
             exe.run(startup_program)
@@ -447,11 +483,6 @@ def test_gen_TruncatedNormal_initializer(self):
         out2_res1 = np.array(out2[0])
         out2_res2 = np.array(out2[1])
 
-        print(out1_res1)
-        print(out1_res2)
-        print(out2_res1)
-        print(out2_res2)
-
         if not core.is_compiled_with_cuda():
             print(">>>>>>> sampling id static >>>>>>>")
             self.assertTrue(np.allclose(out1_res1, out2_res1))
diff --git a/python/paddle/fluid/tests/unittests/test_regularizer.py b/python/paddle/fluid/tests/unittests/test_regularizer.py
index 58b407f8bc1f4..44087c5421a5e 100644
--- a/python/paddle/fluid/tests/unittests/test_regularizer.py
+++ b/python/paddle/fluid/tests/unittests/test_regularizer.py
@@ -169,9 +169,10 @@ def run_program(self, place, feed_list):
         return param_sum
 
     def check_l2decay_regularizer(self, place, model):
+        paddle.manual_seed(1)
+        paddle.framework.random._manual_program_seed(1)
         main_prog = fluid.framework.Program()
         startup_prog = fluid.framework.Program()
-        startup_prog.random_seed = 1
         with self.scope_prog_guard(
                 main_prog=main_prog, startup_prog=startup_prog):
             data = fluid.layers.data(
@@ -188,9 +189,11 @@ def check_l2decay_regularizer(self, place, model):
         return param_sum
 
     def check_l2decay(self, place, model):
+        paddle.manual_seed(1)
+        paddle.framework.random._manual_program_seed(1)
         main_prog = fluid.framework.Program()
         startup_prog = fluid.framework.Program()
-        startup_prog.random_seed = 1
+
         with self.scope_prog_guard(
                 main_prog=main_prog, startup_prog=startup_prog):
             data = fluid.layers.data(
@@ -243,7 +246,8 @@ def test_repeated_regularization(self):
         with fluid.dygraph.guard():
             input = fluid.dygraph.to_variable(
                 np.random.randn(3, 5).astype('float32'))
-            fluid.default_main_program().random_seed = 1
+            paddle.manual_seed(1)
+            paddle.framework.random._manual_program_seed(1)
 
             linear1 = fluid.dygraph.Linear(
                 5, 2, param_attr=fc_param_attr, bias_attr=fc_param_attr)
diff --git a/python/paddle/fluid/tests/unittests/test_transformer_api.py b/python/paddle/fluid/tests/unittests/test_transformer_api.py
index c8d1e77134036..5fea9f69a18c8 100644
--- a/python/paddle/fluid/tests/unittests/test_transformer_api.py
+++ b/python/paddle/fluid/tests/unittests/test_transformer_api.py
@@ -211,7 +211,8 @@ def ffn(src, encoder_layer, ffn_fc1_act="relu"):
 class TestTransformer(unittest.TestCase):
     def test_multi_head_attention(self):
         def multihead_attention_test_helper(self_attention, cache):
-            paddle.framework.manual_seed(2020)
+            paddle.manual_seed(2020)
+            paddle.framework.random._manual_program_seed(2020)
             # self_attention|cross_attention, cache|No cache
             with fluid.dygraph.guard(fluid.CPUPlace()):
 
@@ -275,6 +276,7 @@ def test_transformer_encoder_layer(self):
 
         with fluid.dygraph.guard(fluid.CPUPlace()):
             paddle.framework.manual_seed(2020)
+            paddle.framework.random._manual_program_seed(2020)
 
             ffn_fc1_act = "relu"
             # 1.generate basic params
diff --git a/python/paddle/fluid/tests/unittests/test_uniform_random_op.py b/python/paddle/fluid/tests/unittests/test_uniform_random_op.py
index 158462a1e6e10..a04aaaef0d41b 100644
--- a/python/paddle/fluid/tests/unittests/test_uniform_random_op.py
+++ b/python/paddle/fluid/tests/unittests/test_uniform_random_op.py
@@ -21,6 +21,7 @@
 from op_test import OpTest
 import paddle
 import paddle.fluid.core as core
+import paddle
 from paddle.fluid.op import Operator
 import paddle.fluid as fluid
 from paddle.fluid import Program, program_guard
@@ -234,16 +235,16 @@ def test_check_output(self):
     def check_with_place(self, place):
         scope = core.Scope()
         out = scope.var("X").get_selected_rows()
-
+        paddle.manual_seed(10)
         op = Operator(
             "uniform_random",
             Out="X",
-            shape=[4, 784],
+            shape=[100, 784],
             min=-5.0,
             max=10.0,
             seed=10)
         op.run(scope, place)
-        self.assertEqual(out.get_tensor().shape(), [4, 784])
+        self.assertEqual(out.get_tensor().shape(), [100, 784])
         hist, prob = output_hist(np.array(out.get_tensor()))
         self.assertTrue(
             np.allclose(
@@ -255,19 +256,19 @@ class TestUniformRandomOpSelectedRowsWithDiagInit(
     def check_with_place(self, place):
         scope = core.Scope()
         out = scope.var("X").get_selected_rows()
-
+        paddle.manual_seed(10)
         op = Operator(
             "uniform_random",
             Out="X",
-            shape=[4, 784],
+            shape=[100, 784],
             min=-5.0,
             max=10.0,
             seed=10,
-            diag_num=4,
+            diag_num=100,
             diag_step=784,
             diag_val=1.0)
         op.run(scope, place)
-        self.assertEqual(out.get_tensor().shape(), [4, 784])
+        self.assertEqual(out.get_tensor().shape(), [100, 784])
         hist, prob = output_hist_diag(np.array(out.get_tensor()))
         self.assertTrue(
             np.allclose(
@@ -276,6 +277,7 @@ def check_with_place(self, place):
 
 class TestUniformRandomOpApi(unittest.TestCase):
     def test_api(self):
+        paddle.manual_seed(10)
         x = fluid.layers.data('x', shape=[16], dtype='float32', lod_level=1)
         y = fluid.layers.fc(x,
                             size=16,
@@ -347,12 +349,15 @@ def test_attr_tensor_int32_API(self):
 
 class TestUniformRandomOp_API_seed(unittest.TestCase):
     def test_attr_tensor_API(self):
+        _seed = 10
+        gen = paddle.manual_seed(_seed)
+        gen._is_init_py = False
         startup_program = fluid.Program()
         train_program = fluid.Program()
         with fluid.program_guard(train_program, startup_program):
             _min = 5
             _max = 10
-            _seed = 10
+
             ret = fluid.layers.nn.uniform_random(
                 [2, 3, 2], min=_min, max=_max, seed=_seed)
             ret_2 = fluid.layers.nn.uniform_random(
@@ -386,8 +391,8 @@ def check_with_place(self, place):
         scope = core.Scope()
         out = scope.var("X").get_selected_rows()
         shape_tensor = scope.var("Shape").get_tensor()
-        shape_tensor.set(np.array([4, 784]).astype("int64"), place)
-
+        shape_tensor.set(np.array([100, 784]).astype("int64"), place)
+        paddle.manual_seed(10)
         op = Operator(
             "uniform_random",
             ShapeTensor="Shape",
@@ -396,7 +401,7 @@ def check_with_place(self, place):
             max=10.0,
             seed=10)
         op.run(scope, place)
-        self.assertEqual(out.get_tensor().shape(), [4, 784])
+        self.assertEqual(out.get_tensor().shape(), [100, 784])
         hist, prob = output_hist(np.array(out.get_tensor()))
         self.assertTrue(
             np.allclose(
@@ -418,10 +423,10 @@ def check_with_place(self, place):
         scope = core.Scope()
         out = scope.var("X").get_selected_rows()
         shape_1 = scope.var("shape1").get_tensor()
-        shape_1.set(np.array([4]).astype("int64"), place)
+        shape_1.set(np.array([100]).astype("int64"), place)
         shape_2 = scope.var("shape2").get_tensor()
         shape_2.set(np.array([784]).astype("int64"), place)
-
+        paddle.manual_seed(10)
         op = Operator(
             "uniform_random",
             ShapeTensorList=["shape1", "shape2"],
@@ -430,7 +435,7 @@ def check_with_place(self, place):
             max=10.0,
             seed=10)
         op.run(scope, place)
-        self.assertEqual(out.get_tensor().shape(), [4, 784])
+        self.assertEqual(out.get_tensor().shape(), [100, 784])
         hist, prob = output_hist(np.array(out.get_tensor()))
         self.assertTrue(
             np.allclose(
@@ -455,21 +460,21 @@ def test_errors(self):
 
             def test_Variable():
                 x1 = fluid.create_lod_tensor(
-                    np.zeros((4, 784)), [[1, 1, 1, 1]], fluid.CPUPlace())
+                    np.zeros((100, 784)), [[10, 10, 10, 70]], fluid.CPUPlace())
                 fluid.layers.uniform_random_batch_size_like(x1)
 
             self.assertRaises(TypeError, test_Variable)
 
             def test_shape():
                 x1 = fluid.layers.data(
-                    name='x2', shape=[4, 784], dtype='float32')
+                    name='x2', shape=[100, 784], dtype='float32')
                 fluid.layers.uniform_random_batch_size_like(x1, shape="shape")
 
             self.assertRaises(TypeError, test_shape)
 
             def test_dtype():
                 x2 = fluid.layers.data(
-                    name='x2', shape=[4, 784], dtype='float32')
+                    name='x2', shape=[100, 784], dtype='float32')
                 fluid.layers.uniform_random_batch_size_like(x2, 'int32')
 
             self.assertRaises(TypeError, test_dtype)
@@ -495,20 +500,20 @@ def test_errors(self):
 
             def test_Variable():
                 x1 = fluid.create_lod_tensor(
-                    np.zeros((4, 784)), [[1, 1, 1, 1]], fluid.CPUPlace())
+                    np.zeros((100, 784)), [[10, 10, 10, 70]], fluid.CPUPlace())
                 paddle.tensor.random.uniform(x1)
 
             self.assertRaises(TypeError, test_Variable)
 
             def test_Variable2():
-                x1 = np.zeros((4, 784))
+                x1 = np.zeros((100, 784))
                 paddle.tensor.random.uniform(x1)
 
             self.assertRaises(TypeError, test_Variable2)
 
             def test_dtype():
                 x2 = fluid.layers.data(
-                    name='x2', shape=[4, 784], dtype='float32')
+                    name='x2', shape=[100, 784], dtype='float32')
                 paddle.tensor.random.uniform(x2, 'int32')
 
             self.assertRaises(TypeError, test_dtype)
diff --git a/python/paddle/fluid/tests/unittests/test_while_op.py b/python/paddle/fluid/tests/unittests/test_while_op.py
index 207ff66a0f877..ee01bfb21f820 100644
--- a/python/paddle/fluid/tests/unittests/test_while_op.py
+++ b/python/paddle/fluid/tests/unittests/test_while_op.py
@@ -15,6 +15,7 @@
 from __future__ import print_function
 
 import unittest
+import paddle
 import paddle.fluid.layers as layers
 from paddle.fluid.executor import Executor
 import paddle.fluid.core as core
diff --git a/python/paddle/framework/random.py b/python/paddle/framework/random.py
index 1bb13294805ef..2555d24464112 100644
--- a/python/paddle/framework/random.py
+++ b/python/paddle/framework/random.py
@@ -14,28 +14,50 @@
 
 # TODO: define random api
 import paddle.fluid as fluid
+from paddle.fluid import core
 
 __all__ = ['manual_seed']
 
 
 def manual_seed(seed):
     """
-	:alias_main: paddle.manual_seed
-	:alias: paddle.manual_seed,paddle.framework.random.manual_seed
 
-    Set global manual seed for program
+    Sets the seed for global default generator, which manages the random number generation.
 
     Args:
-        manual_seed(int): random seed for program
+        seed(int): The random seed to set. It is recommend to set a large int number.
 
     Returns:
-        None.
+        Generator: The global default generator object.
 
     Examples:
         .. code-block:: python
 
-            from paddle.framework import manual_seed
-            manual_seed(102)
+            import paddle
+            gen = paddle.manual_seed(102)
+
+    """
+    #TODO(zhiqiu): 1. remove program.random_seed when all random-related op upgrade
+    # 2. support gpu generator by global device 
+
+    seed = int(seed)
+
+    core.default_cpu_generator()._is_init_py = True
+    return core.default_cpu_generator().manual_seed(seed)
+
+
+def _manual_program_seed(seed):
+    """
+    Sets global seed for generating random numbers.
+  
+    NOTE(zhiqiu): This is the original implemention of manual_seed. Keeps it temporally 
+    since CUDA generator is not developed, so we need it in the unittest.
+
+    Args:
+        seed(int): The random seed to set. It is recommend to set a large int number.
+    
+    Returns:
+        None
     """
     fluid.default_main_program().random_seed = seed
     fluid.default_startup_program().random_seed = seed
diff --git a/python/paddle/incubate/hapi/tests/test_model.py b/python/paddle/incubate/hapi/tests/test_model.py
index 8e0c051ee8c39..7fc471aa1e2ee 100644
--- a/python/paddle/incubate/hapi/tests/test_model.py
+++ b/python/paddle/incubate/hapi/tests/test_model.py
@@ -22,6 +22,7 @@
 import shutil
 import tempfile
 
+import paddle
 from paddle import fluid
 from paddle.nn import Conv2d, Pool2D, Linear, ReLU, Sequential, Softmax
 from paddle.fluid.dygraph.base import to_variable
@@ -170,8 +171,8 @@ def setUpClass(cls):
             cls.test_dataset, places=cls.device, batch_size=64)
 
         seed = 333
-        fluid.default_startup_program().random_seed = seed
-        fluid.default_main_program().random_seed = seed
+        paddle.manual_seed(seed)
+        paddle.framework.random._manual_program_seed(seed)
 
         dy_lenet = LeNetDygraph()
         cls.init_param = dy_lenet.state_dict()
@@ -222,8 +223,8 @@ def test_prepare_context(self):
     def fit(self, dynamic, num_replicas=None, rank=None):
         fluid.enable_dygraph(self.device) if dynamic else None
         seed = 333
-        fluid.default_startup_program().random_seed = seed
-        fluid.default_main_program().random_seed = seed
+        paddle.manual_seed(seed)
+        paddle.framework.random._manual_program_seed(seed)
 
         net = LeNet(classifier_activation=None)
         optim_new = fluid.optimizer.Adam(
@@ -327,8 +328,8 @@ def forward(self, x):
 
 class TestModelFunction(unittest.TestCase):
     def set_seed(self, seed=1024):
-        fluid.default_startup_program().random_seed = seed
-        fluid.default_main_program().random_seed = seed
+        paddle.manual_seed(seed)
+        paddle.framework.random._manual_program_seed(seed)
 
     def test_train_batch(self, dynamic=True):
         dim = 20
diff --git a/python/paddle/incubate/hapi/tests/test_text.py b/python/paddle/incubate/hapi/tests/test_text.py
index bdc637997b0cb..c4fef0d749ce7 100644
--- a/python/paddle/incubate/hapi/tests/test_text.py
+++ b/python/paddle/incubate/hapi/tests/test_text.py
@@ -20,6 +20,7 @@
 
 import numpy as np
 
+import paddle
 import paddle.fluid as fluid
 from paddle.fluid.dygraph import Embedding, Linear, Layer
 from paddle.fluid.layers import BeamSearchDecoder
@@ -87,15 +88,18 @@ def _calc_output(self, place, mode="test", dygraph=True):
             fluid.enable_dygraph(place)
         else:
             fluid.disable_dygraph()
-        fluid.default_main_program().random_seed = self._random_seed
-        fluid.default_startup_program().random_seed = self._random_seed
-        layer = self.model_cls(**self.attrs) if isinstance(
-            self.attrs, dict) else self.model_cls(*self.attrs)
-        model = Model(layer, inputs=self.make_inputs())
-        model.prepare()
-        if self.param_states:
-            model.load(self.param_states, optim_state=None)
-        return model.test_batch(self.inputs)
+        gen = paddle.manual_seed(self._random_seed)
+        gen._is_init_py = False
+        paddle.framework.random._manual_program_seed(self._random_seed)
+        scope = fluid.core.Scope()
+        with fluid.scope_guard(scope):
+            layer = self.model_cls(**self.attrs) if isinstance(
+                self.attrs, dict) else self.model_cls(*self.attrs)
+            model = Model(layer, inputs=self.make_inputs())
+            model.prepare()
+            if self.param_states:
+                model.load(self.param_states, optim_state=None)
+            return model.test_batch(self.inputs)
 
     def check_output_with_place(self, place, mode="test"):
         dygraph_output = self._calc_output(place, mode, dygraph=True)
@@ -129,12 +133,9 @@ def setUp(self):
 
     @staticmethod
     def model_init(model, input_size, hidden_size):
-        model.lstm = RNN(
-            BasicLSTMCell(
-                input_size,
-                hidden_size,
-                param_attr=fluid.ParamAttr(name="lstm_weight"),
-                bias_attr=fluid.ParamAttr(name="lstm_bias")))
+        model.lstm = RNN(BasicLSTMCell(
+            input_size,
+            hidden_size, ))
 
     @staticmethod
     def model_forward(model, inputs):