【pir】modify Paddle detection bug (#62165)

xiaoguoguo626807 · web-flow · commit b0ae0c2bc81f · 2024-02-28T18:15:50.000+08:00
* modify if nest pop_to_push_map

* modify paddledectation

* modify utf-8 bug
diff --git a/paddle/fluid/pir/dialect/operator/ir/manual_op.cc b/paddle/fluid/pir/dialect/operator/ir/manual_op.cc
@@ -196,7 +196,7 @@ std::vector<pir::Type> AddNOp::InferMeta(
   paddle::dialect::IrTensor dense_out;
   paddle::dialect::IrMetaTensor meta_out(&dense_out);
 
-  phi::AddNInferMeta(meta_x, &meta_out);
+  phi::AddNInferMeta(meta_x, &meta_out, phi::MetaConfig(false, false));
 
   std::vector<pir::Type> argument_outputs;
   pir::Type out_dense_tensor_type = paddle::dialect::DenseTensorType::get(
@@ -358,7 +358,7 @@ std::vector<pir::Type> AddN_Op::InferMeta(
   paddle::dialect::IrTensor dense_out;
   paddle::dialect::IrMetaTensor meta_out(&dense_out);
 
-  phi::AddNInferMeta(meta_inputs, &meta_out);
+  phi::AddNInferMeta(meta_inputs, &meta_out, phi::MetaConfig(false, false));
 
   std::vector<pir::Type> argument_outputs;
   pir::Type out_dense_tensor_type = paddle::dialect::DenseTensorType::get(
@@ -548,7 +548,7 @@ std::vector<pir::Type> AddNWithKernelOp::InferMeta(
   paddle::dialect::IrTensor dense_out;
   paddle::dialect::IrMetaTensor meta_out(&dense_out);
 
-  phi::AddNInferMeta(meta_inputs, &meta_out);
+  phi::AddNInferMeta(meta_inputs, &meta_out, phi::MetaConfig(false, false));
 
   std::vector<pir::Type> argument_outputs;
   pir::Type out_dense_tensor_type = paddle::dialect::DenseTensorType::get(
diff --git a/python/paddle/autograd/backward_utils.py b/python/paddle/autograd/backward_utils.py
@@ -419,17 +419,22 @@ def remove_useless_full_like_ops(block, ops, state):
     remove ops which are not in use recursively,
 
     '''
+    remove_ops = []
+    inverse_ops = inverse_sort_op(list(ops))
     # from output to input
-    for op in inverse_sort_op(list(ops)):
-        if op.name() == 'pd_op.full_like':
+    for op in inverse_ops:
+        if op.name() == "pd_op.full_like":
             if op.result(0).use_empty():
                 full_op = op.operand_source(1).get_defining_op()
-                remove_op(block, op, state)
-                remove_op(block, full_op, state)
+                remove_ops.append(op)
+                remove_ops.append(full_op)
         elif is_control_flow(op):
             for sub_block in op.blocks():
                 remove_useless_full_like_ops(sub_block, sub_block.ops, state)
 
+    for op in remove_ops:
+        remove_op(block, op, state)
+
 
 def all_stop_gradient_true(block):
     for op in block.ops:
@@ -518,3 +523,10 @@ def get_grad_semantic_info(op):
     else:
         grad_semantic_info = op.get_input_grad_semantics()
     return grad_semantic_info
+
+
+def get_split_op(value):
+    for op in value.all_used_ops():
+        if op.name() == "builtin.split":
+            return op
+    return None
diff --git a/python/paddle/autograd/ir_backward.py b/python/paddle/autograd/ir_backward.py
@@ -29,6 +29,7 @@
     dynamic_shape_prim_vjp_guard,
     get_grad_semantic_info,
     get_real_op_inputs,
+    get_split_op,
     inverse_sort_op,
     is_control_flow,
     is_inplace_net,
@@ -90,24 +91,30 @@ def append_add_n(
     # need add sum op to accumulate gradient
     add_n_list = []
     for item in state.value_to_valuegrad[value]:
-        add_n_list.append(
-            return_map_value(item[0], bwd_value_to_block_argument_map)
-        )
+        if item[0] is not None:
+            add_n_list.append(
+                return_map_value(item[0], bwd_value_to_block_argument_map)
+            )
 
-    if value.is_dense_tensor_array_type():
-        add_n_value = paddle._pir_ops.add_n_array(add_n_list)
+    if len(add_n_list) == 0:
+        for tmp in state.value_to_valuegrad[value]:
+            state.value_to_sumvaluegrad[value].append(tmp)
+        state.value_to_valuegrad[value] = []
     else:
-        add_n_value = paddle.add_n(add_n_list)
+        if value.is_dense_tensor_array_type():
+            add_n_value = paddle._pir_ops.add_n_array(add_n_list)
+        else:
+            add_n_value = paddle.add_n(add_n_list)
 
-    add_n_op = add_n_value.get_defining_op()
-    combine_op = add_n_op.operand_source(0).get_defining_op()
-    update_bwdop_structure(
-        backward_ops, state.op_to_opgrad[op], [combine_op, add_n_op]
-    )
+        add_n_op = add_n_value.get_defining_op()
+        combine_op = add_n_op.operand_source(0).get_defining_op()
+        update_bwdop_structure(
+            backward_ops, state.op_to_opgrad[op], [combine_op, add_n_op]
+        )
 
-    for tmp in state.value_to_valuegrad[value]:
-        state.value_to_sumvaluegrad[value].append(tmp)
-    state.value_to_valuegrad[value] = [[add_n_value]]
+        for tmp in state.value_to_valuegrad[value]:
+            state.value_to_sumvaluegrad[value].append(tmp)
+        state.value_to_valuegrad[value] = [[add_n_value]]
 
 
 def update_bwdop_structure(backward_ops, op_to_opgrad_list, grad_op_list):
@@ -342,18 +349,15 @@ def make_output_with_output_grad(op):
                 value not in state.value_to_valuegrad
                 or state.value_to_valuegrad[value] == []
             ):
-                if (
-                    not value.use_empty()
-                    and value.first_use().owner().name() == "builtin.split"
-                ):
+                if not value.use_empty() and get_split_op(value) is not None:
                     # pattern case:
                     # this fwd_op's output is vectorType, it will split to
                     # Type by builtin_split op, so need get from split op's outputs.
                     (
                         split_zero_flag,
                         split_outputs,
                         split_output_grad,
-                    ) = make_output_with_output_grad(value.first_use().owner())
+                    ) = make_output_with_output_grad(get_split_op(value))
                     zero_flag[i] = all(split_zero_flag)
                     grad_values = [value[0] for value in split_output_grad]
                     state.value_to_valuegrad[value] = [grad_values]
@@ -374,6 +378,8 @@ def make_output_with_output_grad(op):
 
             outputs.append(new_value)
             grad_value = state.value_to_valuegrad[value][0]
+            if grad_value[0] is None:
+                zero_flag[i] = True
             output_grads.append(
                 return_map_value_list(
                     grad_value, bwd_value_to_block_argument_map
