fix multi-targets bugs which this is common case in dy2static (#45277)

python/paddle/fluid/backward.py

@@ -642,12 +642,16 @@ def _addup_repetitive_outputs_(op_descs,
     return op_descs
 
 
-def _remove_no_grad_branch_(op_descs, no_grad_set, grad_op_id_to_fwd_op=None):
+def _remove_no_grad_branch_(op_descs,
+                            no_grad_set,
+                            grad_op_id_to_fwd_op=None,
+                            target_vars=[]):
     """
     Remove unnecessary grad ops
     A grad op can be removed in two cases:
         1. all outputs of the grad op are in 'no_grad_set'
         2. all grad inputs of the grad op are in 'no_grad_set'
+    NOTE: we will skip target_vars's grad name.
     """
 
     def _op_can_be_removed_(op_desc, no_grad_set):
@@ -658,11 +662,13 @@ def _remove_no_grad_branch_(op_descs, no_grad_set, grad_op_id_to_fwd_op=None):
                 name for name in op_desc.input_arg_names()
                 if name.find(core.grad_var_suffix()) != -1
         ], no_grad_set):
-            no_grad_set.update(out_arg_names)
+            no_grad_set.update(set(out_arg_names) - target_grad_var_names)
             return True
         return False
 
     # Remove ops whose outputs are all in no_grad_dict
+    target_grad_var_names = set(
+        [var.name + core.grad_var_suffix() for var in target_vars])
     op_descs = [
         op_desc for op_desc in op_descs
         if not _op_can_be_removed_(op_desc, no_grad_set)
@@ -824,6 +830,7 @@ def serialize_op_decs(op_desc):
 
 def _append_backward_ops_with_checkpoints_(block,
                                            ops,
+                                           target_vars,
                                            target_block,
                                            no_grad_dict,
                                            grad_to_var,
@@ -835,6 +842,7 @@ def _append_backward_ops_with_checkpoints_(block,
     Args:
         block(Block): the block where forward ops are
         ops(Op): the forward operators whose forward recomputation backward ops need to be added
+        target_vars(list[Tensor]): the loss vars we want to calculate gradient.
         target_block(Block): the block which is going to hold new generated grad ops
         no_grad_dict(dict):
             key(int) block index
@@ -1070,7 +1078,7 @@ def _append_backward_ops_with_checkpoints_(block,
     # 4) remove no grad branch as it is in _remove_no_grad_branch_
     grad_op_descs = _remove_no_grad_branch_(grad_op_descs,
                                             no_grad_dict[block.idx],
-                                            grad_op_id_to_fwd_op)
+                                            grad_op_id_to_fwd_op, target_vars)
     added_descs = _add_descs_to_block(grad_op_descs, target_block,
                                       grad_op_id_to_fwd_op)
     return program_stat, checkpoints_name, vars_should_be_hold, recompute_segments
@@ -1140,6 +1148,7 @@ def _rename_grad_name_(name, grad_order):
 
 def _append_backward_ops_(block,
                           ops,
+                          target_vars,
                           target_block,
                           no_grad_dict,
                           grad_to_var,
@@ -1155,6 +1164,7 @@ def _append_backward_ops_(block,
     Args:
         block(Block): the block where forward ops are
         ops(Op): the forward operators whose backward ops need to be added
+        target_vars(list[Tensor]): the loss vars we want to calculate gradient.
         target_block(Block): the block which is going to hold new generated grad ops
         no_grad_dict(dict):
             key(int)  block index
@@ -1212,6 +1222,7 @@ def _append_backward_ops_(block,
             sub_block_path = op_path_dict[op._block_attr_id("sub_block")]
             _append_backward_ops_(sub_block,
                                   sub_block_path,
+                                  target_vars,
                                   grad_sub_block,
                                   no_grad_dict,
                                   grad_to_var,
@@ -1330,7 +1341,7 @@ def _append_backward_ops_(block,
     # if all inputs of the grad op are in no_grad_set, just remove this op
     grad_op_descs = _remove_no_grad_branch_(grad_op_descs,
                                             no_grad_dict[block.idx],
-                                            grad_op_id_to_fwd_op)
+                                            grad_op_id_to_fwd_op, target_vars)
 
     # remove some backward ops
     not_need_ops = _find_not_need_ops(grad_op_descs, ops, input_grad_names_set)
@@ -1765,6 +1776,7 @@ def append_backward(loss,
                 _append_backward_ops_with_checkpoints_(
                     root_block,
                     op_path,
+                    [loss],
                     root_block,
                     no_grad_dict,
                     grad_to_var,
@@ -1774,6 +1786,7 @@ def append_backward(loss,
             _append_backward_ops_(
                 block,  # the block where forward ops are in
                 op_path,
+                [loss],
                 target_grad_block,
                 no_grad_dict,
                 grad_to_var,
@@ -2135,6 +2148,7 @@ def calc_gradient(targets, inputs, target_gradients=None, no_grad_set=None):
     grad_info_map = dict()
     _append_backward_ops_(block,
                           op_path,
+                          targets,
                           block,
                           no_grad_dict,
                           grad_to_var,