suppor inplace in tensor_method_setitem

paddle/fluid/eager/auto_code_generator/eager_generator.cc

@@ -2108,6 +2108,10 @@ static std::string GenerateSingleOpBase(
             GRAD_OUTS_CONTENT_TEMPLATE, grad_output_name, grads_position);
 
       } else {
+        if (dispensable_input_name_set.count(fwd_name) &&
+            grad_ins_fwd_slotname_map.count(fwd_name)) {
+          continue;
+        }
         size_t fwd_input_position = fwd_inputs_name_pos_map.at(fwd_name);
         if (duplicable_input_name_set.count(fwd_name) &&
             !is_op_base_per_duplicable_input) {
@@ -2144,6 +2148,42 @@ static std::string GenerateSingleOpBase(
       BWD_OUTS_MAP_TEMPLATE, outs_name, outs_contents_str);
   generated_grad_function_body += outs_map_str;
   generated_grad_function_body += "\n";
+  for (auto iter : grad_outs) {
+    const std::string& grad_output_name = iter.first;
+
+    if (grad_outs_slotname_map.count(grad_output_name)) {
+      // Fwd Tensor
+      const std::string& fwd_name = grad_outs_slotname_map.at(grad_output_name);
+      if (fwd_inputs_name_pos_map.count(fwd_name)) {
+        if (dispensable_input_name_set.count(fwd_name) &&
+            grad_ins_fwd_slotname_map.count(fwd_name)) {
+          if (duplicable_input_name_set.count(fwd_name) &&
+              !is_op_base_per_duplicable_input) {
+            size_t fwd_input_position = fwd_inputs_name_pos_map.at(fwd_name);
+            const char* DISPENSABLE_GRAD_OUTS_FWD_CONTENT_TEMPLATE =
+                "  if(%s.size() > 0) %s[\"%s\"] = egr::EagerUtils::CreateVars( "
+                "this->OutputMeta()[%d].size() );\n";
+            generated_grad_function_body += paddle::string::Sprintf(
+                DISPENSABLE_GRAD_OUTS_FWD_CONTENT_TEMPLATE, fwd_name, outs_name,
+                grad_output_name, fwd_input_position);
+          } else {
+            const char* DISPENSABLE_GRAD_OUTS_FWD_CONTENT_TEMPLATE =
+                "  if(%s.initialized()) %s[\"%s\"] = "
+                "{std::make_shared<egr::EagerVariable>(egr::Controller::"
+                "Instance().GenerateUniqueName())};\n";
+            generated_grad_function_body += paddle::string::Sprintf(
+                DISPENSABLE_GRAD_OUTS_FWD_CONTENT_TEMPLATE, fwd_name, outs_name,
+                grad_output_name);
+          }
+        }
+      }
+    } else {
+      PADDLE_THROW(platform::errors::Fatal(
+          "Detected mismatched slot names."
+          "Unable to find forward slot name that matches %s",
+          grad_output_name));
+    }
+  }
 
   VLOG(6) << "Generated Outs Map";
 

paddle/fluid/pybind/eager_method.cc

@@ -703,8 +703,6 @@ static PyObject* tensor_method__setitem_eager_tensor(TensorObject* self,
     }
   });
 
-  // TODO(pangyoki) add inplace(BumpInplaceVersion) if need
-
   // 1. Check argumnets
   bool parse_index = true;
 
@@ -753,12 +751,6 @@ static PyObject* tensor_method__setitem_eager_tensor(TensorObject* self,
 
     if (PyCheckTensor(value_obj)) {
       value_tensor = reinterpret_cast<TensorObject*>(value_obj)->tensor;
-
-      // pass the stop_gradient from value to tensor
-      if (!egr::EagerUtils::autograd_meta(&value_tensor)->StopGradient() &&
-          egr::EagerUtils::autograd_meta(&self->tensor)->StopGradient()) {
-        egr::EagerUtils::autograd_meta(&self->tensor)->SetStopGradient(false);
-      }
     } else if (py::isinstance<py::array>(value_obj)) {
       paddle::experimental::Tensor value_tensor_tmp(
           std::make_shared<phi::DenseTensor>(),
@@ -858,8 +850,18 @@ static PyObject* tensor_method__setitem_eager_tensor(TensorObject* self,
     {
       // Release gil and do tracing
       py::gil_scoped_release release;
-      self->tensor = set_value_dygraph_function(self->tensor, value_tensor, {},
-                                                {}, {}, attrs);
+      // use inplace set_value_ operator
+      self->tensor = set_value__dygraph_function(self->tensor, value_tensor, {},
+                                                 {}, {}, attrs);
+    }
+    if (PyCheckTensor(value_obj)) {
+      // pass the stop_gradient from value to tensor.
+      // pass stop gradient should be done after CheckInplace in
+      // set_value__dygraph_function.
+      if (!egr::EagerUtils::autograd_meta(&value_tensor)->StopGradient() &&
+          egr::EagerUtils::autograd_meta(&self->tensor)->StopGradient()) {
+        egr::EagerUtils::autograd_meta(&self->tensor)->SetStopGradient(false);
+      }
     }
   } else {
     auto self_numpy = TensorToPyArray(*self_tensor);
@@ -1179,6 +1181,15 @@ static PyObject* tensor__inplace_version(TensorObject* self, PyObject* args,
   EAGER_CATCH_AND_THROW_RETURN_NULL
 }
 
+static PyObject* tensor__bump_inplace_version(TensorObject* self,
+                                              PyObject* args,
+                                              PyObject* kwargs) {
+  EAGER_TRY
+  self->tensor.bump_inplace_version();
+  return Py_None;
+  EAGER_CATCH_AND_THROW_RETURN_NULL
+}
+
 static PyObject* tensor_method_is_selected_rows(TensorObject* self,
                                                 PyObject* args,
                                                 PyObject* kwargs) {
@@ -1287,6 +1298,9 @@ PyMethodDef variable_methods[] = {
     /***the method of sparse tensor****/
     {"_inplace_version", (PyCFunction)(void (*)(void))tensor__inplace_version,
      METH_VARARGS | METH_KEYWORDS, NULL},
+    {"_bump_inplace_version",
+     (PyCFunction)(void (*)(void))tensor__bump_inplace_version,
+     METH_VARARGS | METH_KEYWORDS, NULL},
     {"is_selected_rows",
      (PyCFunction)(void (*)(void))tensor_method_is_selected_rows,
      METH_VARARGS | METH_KEYWORDS, NULL},

paddle/phi/api/lib/tensor.cc

@@ -364,11 +364,7 @@ void Tensor::bump_inplace_version() {
     auto &inplace_version_counter =
         std::dynamic_pointer_cast<phi::DenseTensor>(impl_)
             ->InplaceVersionCounter();
-    VLOG(3) << "yoki: before bump inplace version: "
-            << inplace_version_counter.CurrentVersion();
     inplace_version_counter.Bump();
-    VLOG(3) << "yoki: after bump inplace version: "
-            << inplace_version_counter.CurrentVersion();
   } else {
     PADDLE_THROW(phi::errors::Unimplemented(
         "bump_inplace_version is only supported on DenseTensor now."));
@@ -380,8 +376,6 @@ uint32_t Tensor::current_inplace_version() {
     auto &inplace_version_counter =
         std::dynamic_pointer_cast<phi::DenseTensor>(impl_)
             ->InplaceVersionCounter();
-    VLOG(3) << "yoki: print version: "
-            << inplace_version_counter.CurrentVersion();
     return inplace_version_counter.CurrentVersion();
   } else {
     PADDLE_THROW(phi::errors::Unimplemented(

python/paddle/fluid/tests/unittests/CMakeLists.txt

@@ -961,7 +961,6 @@ set_tests_properties(test_bicubic_interp_op PROPERTIES TIMEOUT 120)
 set_tests_properties(test_deformable_conv_op PROPERTIES TIMEOUT 120)
 set_tests_properties(test_nearest_interp_op PROPERTIES TIMEOUT 120)
 set_tests_properties(test_profiler PROPERTIES TIMEOUT 120)
-set_tests_properties(test_inplace_eager_fluid PROPERTIES TIMEOUT 120)
 set_tests_properties(test_inplace_softmax_with_cross_entropy PROPERTIES TIMEOUT 120)
 set_tests_properties(test_cross_entropy2_op PROPERTIES TIMEOUT 120)
 set_tests_properties(test_fetch_unmerged PROPERTIES TIMEOUT 120)