[Dy2St]Fix recurrent op eager deletion pass error in dy2st (#47105)

* Fix recurrent op eager deletion pass error in dy2st

* Polish code

* Refine error message

paddle/fluid/framework/ir/memory_optimize_pass/recurrent_op_eager_deletion_pass.cc

@@ -36,6 +36,29 @@ void RecurrentOpEagerDeletionPass::ApplyImpl(Graph *graph) const {
   std::unordered_map<size_t, OpAndGradOpPair> target_ops =
       DeviceIdToRecurrentAndRecurrentGradOp(*graph);
 
+  if (graph->IsConstructedByPartialProgram()) {
+    PADDLE_ENFORCE_LE(target_ops.size(),
+                      1,
+                      platform::errors::InvalidArgument(
+                          "Unsupported multi devices if graph is constructed "
+                          "with partial program."));
+    size_t scope_idx = 0;
+    auto &recur_ops = target_ops[scope_idx].first;
+    auto &recur_grad_ops = target_ops[scope_idx].second;
+
+    auto all_ops = graph->OriginProgram().Block(0).AllOps();
+    if (recur_ops.empty()) {
+      operators::AppendOpVariantByOpName(
+          all_ops, std::string("recurrent"), &recur_ops);
+    } else if (recur_grad_ops.empty()) {
+      operators::AppendOpVariantByOpName(
+          all_ops, std::string("recurrent_grad"), &recur_grad_ops);
+    } else {
+      PADDLE_THROW(platform::errors::InvalidArgument(
+          "One of recur_ops or recur_grad_ops should be empty."));
+    }
+  }
+
   for (auto &entry : target_ops) {
     // Prepare safe eager deletion on different devices because the garbage
     // collection may be different across devices

paddle/fluid/operators/controlflow/op_variant.cc

@@ -81,5 +81,22 @@ void AppendOpVariantByOpName(const std::vector<framework::OpDesc *> &op_descs,
   }
 }
 
+void AppendOpVariantByOpName(
+    const std::vector<framework::OpDesc *> &op_descs,
+    const std::string &candidate_op_name,
+    std::unordered_set<OpVariant, OpVariant::Hasher> *result_ops) {
+  PADDLE_ENFORCE_NOT_NULL(
+      result_ops,
+      platform::errors::Unavailable("result_ops should not be a null_ptr."));
+  for (auto *op_desc : op_descs) {
+    PADDLE_ENFORCE_NOT_NULL(
+        op_desc,
+        platform::errors::Unavailable("op_desc should not be a null_ptr."));
+    if (op_desc->Type() == candidate_op_name) {
+      result_ops->emplace(op_desc);
+    }
+  }
+}
+
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/controlflow/op_variant.h

@@ -78,5 +78,10 @@ void AppendOpVariantByOpName(const std::vector<framework::OpDesc *> &op_descs,
                              const std::string &candidate_op_name,
                              std::vector<OpVariant> *result_ops);
 
+void AppendOpVariantByOpName(
+    const std::vector<framework::OpDesc *> &op_descs,
+    const std::string &candidate_op_name,
+    std::unordered_set<OpVariant, OpVariant::Hasher> *result_ops);
+
 }  // namespace operators
 }  // namespace paddle

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

@@ -690,5 +690,45 @@ class EagerDeletionFarwardOnlyRnnAndBackwardRnnTest(
         np.testing.assert_allclose(pd_output, py_output, rtol=0.01)
 
 
+class RecurrentNet(paddle.nn.Layer):
+
+    def __init__(self):
+        super(RecurrentNet, self).__init__()
+        self.cell = paddle.nn.SimpleRNNCell(16, 32)
+        self.rnn = paddle.nn.RNN(self.cell)
+
+    def forward(self, inputs, prev_h):
+        outputs, final_states = self.rnn(inputs, prev_h)
+        return outputs, final_states
+
+
+class TestDy2StRecurrentOpBackward(unittest.TestCase):
+
+    def setUp(self):
+        paddle.disable_static()
+        paddle.seed(100)
+
+    def tearDown(self):
+        paddle.enable_static()
+
+    def test_recurrent_backward(self):
+        net = RecurrentNet()
+        inputs = paddle.rand((4, 23, 16))
+        inputs.stop_gradient = False
+        prev_h = paddle.randn((4, 32))
+        prev_h.stop_gradient = False
+
+        outputs, final_states = net(inputs, prev_h)
+        outputs.backward()
+        dy_grad = inputs.gradient()
+        inputs.clear_gradient()
+
+        net = paddle.jit.to_static(net)
+        outputs, final_states = net(inputs, prev_h)
+        outputs.backward()
+        st_grad = inputs.gradient()
+        np.testing.assert_allclose(dy_grad, st_grad)
+
+
 if __name__ == '__main__':
     unittest.main()