support test_auto_prune_partial (#38871)

paddle/fluid/eager/api/utils/tensor_utils.cc

@@ -49,7 +49,6 @@ egr::EagerTensor CreateTensorWithValue(const pten::DDim& ddim,
   egr::EagerTensor out = egr::EagerTensor();
   out.set_tensor(std::make_shared<paddle::experimental::Tensor>(tensor));
   auto meta = EagerUtils::autograd_meta(&out);
-
   if (is_leaf) {
     auto accumulation_node = std::make_shared<GradNodeAccumulation>();
     meta->SetGradNode(accumulation_node);

paddle/fluid/eager/backward.cc

@@ -181,7 +181,9 @@ void RunBackward(const std::vector<egr::EagerTensor>& tensors,
     PADDLE_ENFORCE(edges.size() == grad_output_tensors.size() || edges.empty(),
                    paddle::platform::errors::Fatal(
                        "Number of edges should be either empty ( for leaf node "
-                       ") or the same as number of output grad tensors"));
+                       ") or the same as number of output grad tensors, but we "
+                       "got edges size is: %d, grad_output size is: %d",
+                       edges.size(), grad_output_tensors.size()));
 
     for (size_t i = 0; i < edges.size(); i++) {
       for (size_t j = 0; j < edges[i].size(); j++) {

paddle/fluid/eager/eager_tensor.h

@@ -195,7 +195,6 @@ class EagerTensor final {
     }
     tensor_->copy_(*(src.tensor_.get()), blocking);
   }
-
   /* Part 6: Operator overloading */
   EagerTensor& operator=(const EagerTensor& x) & {
     tensor_ = x.tensor_;
@@ -238,7 +237,7 @@ class EagerTensor final {
           // Contruct framework::Tensor from egr::EagerTensor
           auto tensor_dense =
               std::dynamic_pointer_cast<pten::DenseTensor>(tensor_->impl());
-          if (tensor_dense) {
+          if (tensor_dense && tensor_dense.get()) {
             paddle::experimental::SharesStorage(tensor_dense.get(),
                                                 framework_tensor);
           } else {
@@ -292,11 +291,10 @@ class EagerTensor final {
   template <typename LEGACY_TYPE, typename TYPE>
   void SetImplWithLegacyTensor() {
     const auto& framework_tensor = var_.Get<LEGACY_TYPE>();
-    if (this->initialized()) {
+    if (defined()) {
       VLOG(8) << "Sync Var to initialized tensor for: " << name();
       paddle::experimental::ReMakePtenDenseTensor(
-          framework_tensor,
-          static_cast<pten::DenseTensor*>(this->impl().get()));
+          framework_tensor, static_cast<pten::DenseTensor*>(impl().get()));
     } else {
       VLOG(8) << "Sync Var to uninitialized tensor for: " << name();
       this->set_impl(std::move(

paddle/fluid/eager/grad_node_info.cc

@@ -47,48 +47,18 @@ void GradNodeBase::AddEdges(std::vector<AutogradMeta*>* metas, size_t slot_id) {
     // adj_edges has as same rank as fwd inputs, and record it's output rank
     // from
     // its pre-ops
-    if (meta) {
+    if (meta && !meta->StopGradient()) {
       auto node = meta->GetMutableGradNode();
       if (node) {
         adj_edges_[slot_id].emplace_back(meta->GetMutableGradNode(),
                                          meta->OutRankInfo());
       } else {
-        if (!meta->StopGradient()) {
         meta->SetGradNode(std::make_shared<egr::GradNodeAccumulation>());
         adj_edges_[slot_id].emplace_back(meta->GetMutableGradNode(),
                                          meta->OutRankInfo());
       }
     }
   }
-  }
-}
-
-void GradNodeBase::AddEdges(const std::vector<AutogradMeta*>& metas,
-                            size_t slot_id) {
-  PADDLE_ENFORCE_LT(
-      slot_id, adj_edges_.size(),
-      paddle::platform::errors::InvalidArgument(
-          "Given slot id is out of range of adj_edges outter size, "
-          "adj_edges is designed to has the same size of grad "
-          "inputs's slot num."));
-  for (const auto& meta : metas) {
-    // adj_edges has as same rank as fwd inputs, and record it's output rank
-    // from
-    // its pre-ops
-    if (meta) {
-      auto node = meta->GetMutableGradNode();
-      if (node) {
-        adj_edges_[slot_id].emplace_back(meta->GetMutableGradNode(),
-                                         meta->OutRankInfo());
-      } else {
-        if (!meta->StopGradient()) {
-          meta->SetGradNode(std::make_shared<egr::GradNodeAccumulation>());
-          adj_edges_[slot_id].emplace_back(meta->GetMutableGradNode(),
-                                           meta->OutRankInfo());
-        }
-      }
-    }
-  }
 }
 
 void GradNodeBase::AddEdges(AutogradMeta* meta, size_t slot_id) {
@@ -98,19 +68,18 @@ void GradNodeBase::AddEdges(AutogradMeta* meta, size_t slot_id) {
           "Given slot id is out of range of adj_edges outter size, "
           "adj_edges is designed to has the same size of grad "
           "inputs's slot num."));
-  if (meta) {
+  if (meta && !meta->StopGradient()) {
+    VLOG(6) << "Add Edges for slot: " << slot_id;
     auto node = meta->GetMutableGradNode();
     if (node) {
       adj_edges_[slot_id].emplace_back(meta->GetMutableGradNode(),
                                        meta->OutRankInfo());
     } else {
-      if (!meta->StopGradient()) {
       meta->SetGradNode(std::make_shared<egr::GradNodeAccumulation>());
       adj_edges_[slot_id].emplace_back(meta->GetMutableGradNode(),
                                        meta->OutRankInfo());
     }
   }
-  }
 }
 
 const std::vector<GradSlotMeta>& GradNodeBase::InputMeta() const {

paddle/fluid/eager/grad_node_info.h

@@ -106,7 +106,6 @@ class GradNodeBase {
    * This one is called slot by slot
    * **/
   void AddEdges(std::vector<AutogradMeta*>* metas, size_t slot_id);
-  void AddEdges(const std::vector<AutogradMeta*>& metas, size_t slot_id);
   void AddEdges(AutogradMeta* meta, size_t slot_id);
 
   /**

paddle/fluid/eager/tests/data_structure_tests/grad_node_info_test.cc

@@ -56,15 +56,17 @@ TEST(GradNodeInfo, GradNodeBase) {
   VLOG(6) << "Test Add Edges";
   egr::Edge edge0(grad_test_node1, 1, 2);
   auto auto_grad0 = std::make_shared<egr::AutogradMeta>(edge0);
+  auto_grad0->SetStopGradient(false);
   egr::Edge edge1(grad_test_node1, 3, 4);
   auto auto_grad1 = std::make_shared<egr::AutogradMeta>(edge1);
+  auto_grad1->SetStopGradient(false);
   grad_test_node0->AddEdges(auto_grad0.get(), 0);
   CHECK_EQ(grad_test_node0->GetEdges()[0][0].GetEdgeRankInfo().first,
            size_t(1));
   CHECK_EQ(grad_test_node0->GetEdges()[0][0].GetEdgeRankInfo().second,
            size_t(2));
   std::vector<egr::AutogradMeta*> metas = {auto_grad1.get()};
-  grad_test_node0->AddEdges(metas, 1);
+  grad_test_node0->AddEdges(&metas, 1);
   CHECK_EQ(grad_test_node0->GetEdges()[1][0].GetEdgeRankInfo().first,
            size_t(3));
   CHECK_EQ(grad_test_node0->GetEdges()[1][0].GetEdgeRankInfo().second,

paddle/fluid/eager/tests/task_tests/backward_test.cc

@@ -69,9 +69,11 @@ TEST(Backward, SingleNodeEmptyGrad) {
 
     // Connect Node0 -> AccumulationNode via Edge
     auto meta = egr::AutogradMeta();
+    meta.SetStopGradient(false);
     meta.SetSingleOutRankWithSlot(0, 0);
     meta.SetGradNode(acc_node_ptr);
-    node0_ptr->AddEdges({&meta}, 0);
+    std::vector<egr::AutogradMeta*> res = {&meta};
+    node0_ptr->AddEdges(&res, 0);
   }
   std::vector<egr::EagerTensor> outs = {target_tensor};
   // Run Backward
@@ -130,9 +132,11 @@ TEST(Backward, SingleNodeCustomGrad) {
 
     // Connect Node0 -> AccumulationNode via Edge
     auto meta = egr::AutogradMeta();
+    meta.SetStopGradient(false);
     meta.SetSingleOutRankWithSlot(0, 0);
     meta.SetGradNode(acc_node_ptr);
-    node0_ptr->AddEdges({&meta}, 0);
+    std::vector<egr::AutogradMeta*> res = {&meta};
+    node0_ptr->AddEdges(&res, 0);
   }
 
   // Run Backward
@@ -188,9 +192,11 @@ TEST(Backward, LinearNodes) {
 
     // Connect Node0 -> Node1 via Edge
     auto meta0 = egr::AutogradMeta();
+    meta0.SetStopGradient(false);
     meta0.SetSingleOutRankWithSlot(0, 0);
     meta0.SetGradNode(node1_ptr);
-    node0_ptr->AddEdges({&meta0}, 0);
+    std::vector<egr::AutogradMeta*> res0 = {&meta0};
+    node0_ptr->AddEdges(&res0, 0);
 
     // Connect Tensor and AccumulationNode via AutoGradMeta
     auto acc_node_ptr = std::make_shared<egr::GradNodeAccumulation>();
@@ -204,9 +210,11 @@ TEST(Backward, LinearNodes) {
 
     // Connect Node1 -> AccumulationNode via Edge
     auto meta1 = egr::AutogradMeta();
+    meta1.SetStopGradient(false);
     meta1.SetSingleOutRankWithSlot(0, 0);
     meta1.SetGradNode(acc_node_ptr);
-    node1_ptr->AddEdges({&meta1}, 0);
+    std::vector<egr::AutogradMeta*> res1 = {&meta1};
+    node1_ptr->AddEdges(&res1, 0);
   }
 
   // Use Empty Grad Tensor
@@ -283,15 +291,19 @@ TEST(Backward, WithAccumulation) {
 
     // Connect Node0 -> Node2 via Edge
     auto meta0 = egr::AutogradMeta();
+    meta0.SetStopGradient(false);
     meta0.SetSingleOutRankWithSlot(0, 0);
     meta0.SetGradNode(node2_ptr);
-    node0_ptr->AddEdges({&meta0}, 0);
+    std::vector<egr::AutogradMeta*> res0 = {&meta0};
+    node0_ptr->AddEdges(&res0, 0);
 
     // Connect Node1 -> Node2 via Edge
     auto meta1 = egr::AutogradMeta();
+    meta1.SetStopGradient(false);
     meta1.SetSingleOutRankWithSlot(0, 0);
     meta1.SetGradNode(node2_ptr);
-    node1_ptr->AddEdges({&meta1}, 0);
+    std::vector<egr::AutogradMeta*> res1 = {&meta1};
+    node1_ptr->AddEdges(&res1, 0);
 
     // Connect Tensor and AccumulationNode via AutoGradMeta
     auto acc_node_ptr = std::make_shared<egr::GradNodeAccumulation>();
@@ -305,9 +317,11 @@ TEST(Backward, WithAccumulation) {
 
     // Connect Node2 -> AccumulationNode via Edge
     auto meta2 = egr::AutogradMeta();
+    meta2.SetStopGradient(false);
     meta2.SetSingleOutRankWithSlot(0, 0);
     meta2.SetGradNode(acc_node_ptr);
-    node2_ptr->AddEdges({&meta2}, 0);
+    std::vector<egr::AutogradMeta*> res2 = {&meta2};
+    node2_ptr->AddEdges(&res2, 0);
   }
 
   RunBackward(target_tensors, grad_tensors);

paddle/fluid/eager/tests/task_tests/cross_batch_accumulation_test.cc

@@ -62,8 +62,10 @@ TEST(CrossBatchAccumulation, SingleScaleNode) {
 
     auto meta = AutogradMeta();
     meta.SetSingleOutRankWithSlot(0, 0);
+    meta.SetStopGradient(false);
     meta.SetGradNode(acc_node_ptr);
-    scale_node_ptr->AddEdges({&meta}, 0);
+    std::vector<egr::AutogradMeta*> res = {&meta};
+    scale_node_ptr->AddEdges(&res, 0);
 
     AutogradMeta* auto_grad_meta1 = EagerUtils::autograd_meta(&leaf_tensor);
     auto_grad_meta1->SetGradNode(

paddle/fluid/eager/tests/task_tests/hook_test.cc

@@ -105,9 +105,11 @@ TEST(RetainGrad, HookBeforeRetainGrad) {
   // Connect ScaleNode -> AccumulationNode via Edge
   {
     auto meta = AutogradMeta();
+    meta.SetStopGradient(false);
     meta.SetSingleOutRankWithSlot(0, 0);
     meta.SetGradNode(acc_node_ptr);
-    scale_node_ptr->AddEdges({&meta}, 0);
+    std::vector<egr::AutogradMeta*> res = {&meta};
+    scale_node_ptr->AddEdges(&res, 0);
   }
 
   // Retain Grad for leaf tensor1
@@ -180,9 +182,11 @@ TEST(RetainGrad, HookAfterRetainGrad) {
   // Connect ScaleNode -> AccumulationNode via Edge
   {
     auto meta = AutogradMeta();
+    meta.SetStopGradient(false);
     meta.SetSingleOutRankWithSlot(0, 0);
     meta.SetGradNode(acc_node_ptr);
-    scale_node_ptr->AddEdges({&meta}, 0);
+    std::vector<egr::AutogradMeta*> res = {&meta};
+    scale_node_ptr->AddEdges(&res, 0);
   }
 
   // Retain Grad for leaf tensor1

paddle/fluid/pybind/eager_method.cc

@@ -234,6 +234,44 @@ static PyObject* eager_tensor__zero_grads(EagerTensorObject* self,
   EAGER_CATCH_AND_THROW_RETURN_NULL
 }
 
+static PyObject* eager_tensor__share_buffer_to(EagerTensorObject* self,
+                                               PyObject* args,
+                                               PyObject* kwargs) {
+  EAGER_SYNC_TRY
+  egr::EagerTensor* src_ptr =
+      &(reinterpret_cast<EagerTensorObject*>(PyTuple_GET_ITEM(args, 0))
+            ->eager_tensor);
+  PADDLE_ENFORCE_EQ(self->eager_tensor.initialized(), true,
+                    platform::errors::InvalidArgument(
+                        "Tensor %s has not been initialized! please initialize "
+                        "src tensor before share_buffer_with to other.",
+                        self->eager_tensor.name()));
+  src_ptr->set_impl(self->eager_tensor.impl());
+  Py_INCREF(Py_None);
+  return Py_None;
+  EAGER_CATCH_AND_THROW_RETURN_NULL
+}
+
+static PyObject* eager_tensor__is_shared_buffer_with(EagerTensorObject* self,
+                                                     PyObject* args,
+                                                     PyObject* kwargs) {
+  EAGER_SYNC_TRY
+  egr::EagerTensor src_tensor =
+      CastPyArg2EagerTensor(PyTuple_GET_ITEM(args, 0), 0);
+  PADDLE_ENFORCE_EQ(src_tensor.initialized(), true,
+                    platform::errors::InvalidArgument(
+                        "Tensor %s has not been initialized! please initialize "
+                        "src tensor before share_buffer_with to other.",
+                        src_tensor.name()));
+  bool res = false;
+  if (!self->eager_tensor.defined() || !src_tensor.defined()) {
+    return ToPyObject(res);
+  }
+  res = (self->eager_tensor.impl().get() == src_tensor.impl().get());
+  return ToPyObject(res);
+  EAGER_CATCH_AND_THROW_RETURN_NULL
+}
+
 static PyObject* eager_tensor_method_detach(EagerTensorObject* self,
                                             PyObject* args, PyObject* kwargs) {
   EAGER_SYNC_TRY
@@ -278,6 +316,12 @@ PyMethodDef variable_methods[] = {
      METH_VARARGS | METH_KEYWORDS, NULL},
     {"_zero_grads", (PyCFunction)(void (*)(void))eager_tensor__zero_grads,
      METH_VARARGS | METH_KEYWORDS, NULL},
+    {"_is_shared_buffer_to",
+     (PyCFunction)(void (*)(void))eager_tensor__share_buffer_to,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"_share_buffer_with",
+     (PyCFunction)(void (*)(void))eager_tensor__is_shared_buffer_with,
+     METH_VARARGS | METH_KEYWORDS, NULL},
     {"detach", (PyCFunction)(void (*)(void))eager_tensor_method_detach,
      METH_VARARGS | METH_KEYWORDS, NULL},
     {NULL, NULL, 0, NULL}};

python/paddle/fluid/tests/unittests/test_egr_python_api.py

@@ -645,6 +645,33 @@ class EagerTensorPropertiesTestCase(unittest.TestCase):
                 self.assertTrue(tensor3.stop_gradient, True)
                 self.assertTrue(tensor3.place.is_cpu_place())
 
+        def test_share_buffer_to():
+            arr = np.ones([4, 16, 16, 32]).astype('float32')
+            arr1 = np.zeros([4, 16]).astype('float32')
+            arr2 = np.ones([4, 16, 16, 32]).astype('float32') + np.ones(
+                [4, 16, 16, 32]).astype('float32')
+            tensor = None
+            tensor2 = None
+            tensor = paddle.to_tensor(arr, core.VarDesc.VarType.FP32,
+                                      core.CPUPlace())
+            tensor3 = core.eager.EagerTensor()
+            if core.is_compiled_with_cuda():
+                tensor2 = paddle.to_tensor(arr2, core.VarDesc.VarType.FP32,
+                                           core.CUDAPlace(0))
+            else:
+                tensor2 = paddle.to_tensor(arr2, core.VarDesc.VarType.FP32,
+                                           core.CPUPlace())
+            self.assertTrue(np.array_equal(tensor.numpy(), arr1))
+            self.assertTrue(np.array_equal(tensor2.numpy(), arr2))
+            tensor2._share_buffer_to(tensor)
+            self.assertTrue(np.array_equal(tensor.numpy(), arr2))
+            self.assertTrue(np.array_equal(tensor2.numpy(), arr2))
+            self.assertTrue(tensor._is_shared_buffer_with(tensor2))
+            self.assertTrue(tensor2._is_shared_buffer_with(tensor))
+            tensor._share_buffer_to(tensor3)
+            self.assertTrue(np.array_equal(tensor3.numpy(), arr2))
+            self.assertTrue(tensor3._is_shared_buffer_with(tensor))
+
     def test_properties(self):
         print("Test_properties")
         with _test_eager_guard():

python/paddle/fluid/tests/unittests/test_imperative_auto_prune.py

@@ -15,6 +15,7 @@
 import unittest
 import paddle.fluid as fluid
 import numpy as np
+from paddle.fluid.framework import _test_eager_guard
 
 
 class AutoPruneLayer0(fluid.Layer):
@@ -145,7 +146,7 @@ class MyLayer2(fluid.Layer):
 
 
 class TestImperativeAutoPrune(unittest.TestCase):
-    def test_auto_prune(self):
+    def func_auto_prune(self):
         with fluid.dygraph.guard():
             case1 = AutoPruneLayer0(input_size=5)
             value1 = np.arange(25).reshape(5, 5).astype("float32")
@@ -157,7 +158,12 @@ class TestImperativeAutoPrune(unittest.TestCase):
             self.assertTrue(case1.linear2.weight._grad_ivar() is not None)
             self.assertTrue(case1.linear1.weight._grad_ivar() is not None)
 
-    def test_auto_prune2(self):
+    def test_auto_prune(self):
+        with _test_eager_guard():
+            self.func_auto_prune()
+        self.func_auto_prune()
+
+    def func_auto_prune2(self):
         with fluid.dygraph.guard():
             case2 = AutoPruneLayer1(input_size=5)
             value1 = np.arange(25).reshape(5, 5).astype("float32")
@@ -170,6 +176,11 @@ class TestImperativeAutoPrune(unittest.TestCase):
             self.assertTrue(case2.linear2.weight._grad_ivar() is None)
             self.assertTrue(case2.linear1.weight._grad_ivar() is not None)
 
+    def test_auto_prune2(self):
+        with _test_eager_guard():
+            self.func_auto_prune2()
+        self.func_auto_prune2()
+
     def test_auto_prune3(self):
         with fluid.dygraph.guard():
             case3 = AutoPruneLayer3(input_size=784)