【pir backward】Passing the gradient output of tuple_pop between if_grad_grad and if_grad (#71445)

xiaoguoguo626807 · web-flow · commit 14ae5d50b566 · 2025-03-07T13:59:23.000+08:00
* add if_grad_grad gradient

* modiy test_ifelse
diff --git a/paddle/fluid/framework/new_executor/instruction/control_flow/tuple_push_instruction.cc b/paddle/fluid/framework/new_executor/instruction/control_flow/tuple_push_instruction.cc
@@ -42,6 +42,8 @@ bool ParsePlace(const pir::Type& type, OpFuncType* type_) {
         return true;
       }
     }
+  } else if (!type) {
+    return false;
   } else {
     PADDLE_THROW(common::errors::PreconditionNotMet(
         "Only support AllocatedDenseTensorType and "
diff --git a/paddle/fluid/pir/dialect/operator/ir/control_flow_op.cc b/paddle/fluid/pir/dialect/operator/ir/control_flow_op.cc
@@ -945,6 +945,20 @@ std::vector<std::vector<pir::Value>> TuplePushOpVjpInterfaceModel::Vjp(
     res[i].resize(1);
     res[i][0] = pop_op.result(i - 1);
   }
+
+  // set pop op stop_gradient attribute.
+  std::vector<pir::Attribute> outs_stop_gradient;
+  for (auto i = 1u; i < op->num_operands(); ++i) {
+    auto value = op->operand_source(i);
+    auto bool_attr = value.attribute<pir::BoolAttribute>(kStopGradientAttrName);
+    outs_stop_gradient.push_back(
+        bool_attr ? bool_attr
+                  : pir::BoolAttribute::get(pir::IrContext::Instance(), true));
+  }
+
+  pop_op->set_attribute(
+      kStopGradientAttrName,
+      pir::ArrayAttribute::get(pir::IrContext::Instance(), outs_stop_gradient));
   return res;
 }
 
diff --git a/python/paddle/autograd/backward_utils.py b/python/paddle/autograd/backward_utils.py
@@ -756,3 +756,46 @@ def update_while_output_stopgradient(while_op, yield_op):
         # Set to False if stop_gradient is False
         if not stop_grad:
             while_op.result(i - 1).stop_gradient = False
+
+
+def find_index_of_yiled(value, yield_op):
+    for i, v in enumerate(yield_op.operands_source()):
+        if v.is_same(value):
+            return i
+    return -1
+
+
+def update_tuple_pop_origin_inputs(tuple_pop_outputs):
+    if tuple_pop_outputs == []:
+        return tuple_pop_outputs
+    op = tuple_pop_outputs[0][0].get_defining_op()
+    assert op.name() == "cf.tuple_pop"
+    stack_op = op.operand_source(0).get_defining_op()
+    tuple_push_inputs = stack_op.result(1).first_use().owner().operands_source()
+    tuple_push_inputs_with_if = []
+    for input in tuple_push_inputs:
+        if input.first_use().owner().name() == "cf.yield":
+            yield_op = input.first_use().owner()
+            index = find_index_of_yiled(input, yield_op)
+            assert index != -1
+            tuple_push_inputs_with_if.append(
+                yield_op.get_parent_block().parent_op.result(index)
+            )
+        else:
+            tuple_push_inputs_with_if.append(input)
+
+    # pass inlets
+    return tuple_push_inputs_with_if[1:]
+
+
+def value_in_block(value, block):
+    value_block = value.get_defining_op().get_parent_block()
+    while block.parent_op.name() != "builtin.module":
+        if block == value_block:
+            return True
+        block = block.parent_block
+    # now block is module op's block
+    if block == value_block:
+        return True
+
+    return False
diff --git a/python/paddle/autograd/ir_backward.py b/python/paddle/autograd/ir_backward.py
@@ -47,7 +47,9 @@
     some_in_set,
     update_if_output_stopgradient,
     update_no_grad_set_by_stopgradient,
+    update_tuple_pop_origin_inputs,
     update_while_output_stopgradient,
+    value_in_block,
     warning_once,
     while_prune_check,
 )
@@ -588,6 +590,16 @@ def update_input_grad_map(op, input_grads, all_inputs):
                     state.value_to_valuegrad[input].append([input_grad])
             i += 1
 
+    def update_if_double_grad_input_grad_map(input_grads, all_inputs):
+        assert len(input_grads) == len(
+            all_inputs
+        ), "input_grads should same to all_inputs"
+        for input, input_grad in zip(all_inputs, input_grads):
+            if isinstance(input_grad, list):
+                state.value_to_valuegrad[input].append(input_grad)
+            else:
+                state.value_to_valuegrad[input].append([input_grad])
+
     def append_yield(
         block,
         base_op,
@@ -634,7 +646,15 @@ def append_yield(
                     new_value = return_map_value(
                         value, control_flow_value_to_copyvalue_map
                     )
-                    append_full_like(0.0, new_value, value, state, backward_ops)
+                    if not value_in_block(new_value, block):
+                        # new_value.defining_op is another if block's tuple_pop
+                        state.value_to_valuegrad[value] = [
+                            [paddle.pir.fake_value()]
+                        ]
+                    else:
+                        append_full_like(
+                            0.0, new_value, value, state, backward_ops
+                        )
 
                 input_grad = return_map_value(
                     state.value_to_valuegrad[value][0][0],
@@ -748,6 +768,42 @@ def append_yield(
                         origin_inputs = get_real_op_inputs(op)
                         for sub_block in op.blocks():
                             build_pipe_for_block(sub_block)
+                        # only for double grad if op
+                        true_block = op.as_if_op().true_block()
+                        false_block = op.as_if_op().false_block()
+
+                        true_block_pop_inputs = []
+                        true_block_pop_input_grad_stopgradients = []
+                        if true_block.ops[0].name() == "cf.tuple_pop":
+                            for result in true_block.ops[0].results():
+                                true_block_pop_inputs.append([result])
+                                true_block_pop_input_grad_stopgradients.append(
+                                    [result.stop_gradient]
+                                )
+                        false_block_pop_inputs = []
+                        false_block_pop_input_grad_stopgradients = []
+                        if false_block.ops[0].name() == 'cf.tuple_pop':
+                            for result in false_block.ops[0].results():
+                                false_block_pop_inputs.append([result])
+                                false_block_pop_input_grad_stopgradients.append(
+                                    [result.stop_gradient]
+                                )
+
+                        if (
+                            true_block_pop_inputs != []
+                            or false_block_pop_inputs != []
+                        ):
+                            inputs = (
+                                inputs
+                                + true_block_pop_inputs
+                                + false_block_pop_inputs
+                            )
+                            input_grad_stopgradients = (
+                                input_grad_stopgradients
+                                + true_block_pop_input_grad_stopgradients
+                                + false_block_pop_input_grad_stopgradients
+                            )
+
                         with dynamic_shape_prim_vjp_guard(op, inputs):
                             input_grads = paddle.framework.core.call_vjp(
                                 op,
@@ -777,6 +833,7 @@ def append_yield(
                             sub_control_flow_value_to_copyvalue_map = (
                                 control_flow_value_to_copyvalue_map.copy()
                             )
+
                             append_backward_ops(
                                 op,
                                 [input[0] for input in inputs[1:]],
@@ -801,8 +858,36 @@ def append_yield(
                         )
                         for input_tuple in inputs_used_by_other_op:
                             state.value_to_valuegrad[input_tuple[0]] = []
+
                         # update input_grad map
-                        update_input_grad_map(op, input_grads, origin_inputs)
+                        if (
+                            true_block_pop_inputs != []
+                            or false_block_pop_inputs != []
+                        ):
+                            true_block_pop_inputs = (
+                                update_tuple_pop_origin_inputs(
+                                    true_block_pop_inputs
+                                )
+                            )
+                            false_block_pop_inputs = (
+                                update_tuple_pop_origin_inputs(
+                                    false_block_pop_inputs
+                                )
+                            )
+                            # delete cond inputs
+                            origin_inputs = (
+                                origin_inputs[1:]
+                                + true_block_pop_inputs
+                                + false_block_pop_inputs
+                            )
+                            update_if_double_grad_input_grad_map(
+                                input_grads, origin_inputs
+                            )
+                        else:
+                            update_input_grad_map(
+                                op, input_grads, origin_inputs
+                            )
+
                     elif op.name() == "pd_op.while":
                         origin_inputs = get_real_op_inputs(op)
                         # prepare while[cond, loop_vars, other_input] other_input's grad
@@ -938,8 +1023,11 @@ def append_yield(
                     op.num_operands() == 0
                     and op.num_results() != 0
                     or op.name() == "pd_op.full_like"
+                    or op.name() == "cf.tuple_pop"
                 ):
                     for value in op.results():
+                        if value not in state.value_to_valuegrad:
+                            continue
                         if len(state.value_to_valuegrad[value]) > 1:
                             append_add_n(
                                 op,
diff --git a/test/dygraph_to_static/test_high_order_net.py b/test/dygraph_to_static/test_high_order_net.py
@@ -14,6 +14,7 @@
 
 import unittest
 
+import numpy as np
 from dygraph_to_static_utils import (
     Dy2StTestBase,
     test_ast_only,
@@ -41,7 +42,7 @@ def forward(self, x, y):
 class TestBackwardHasNoGradError(Dy2StTestBase):
     @test_ast_only
     @test_pir_only
-    def test_backward_has_no_grad_error(self):
+    def _test_backward_has_no_grad_error(self):
         net = HighOrderNet()
         static_net = paddle.jit.to_static(net, full_graph=True)
 
@@ -58,5 +59,61 @@ def test_backward_has_no_grad_error(self):
             x_grad_grad.backward()
 
 
+class HighOrderControlFlowNet(paddle.nn.Layer):
+    def __init__(self):
+        super().__init__()
+        self.eps = 1e-5
+
+    def forward(self, x):
+        if x.numel() > 0:
+            variance, mean = (
+                paddle.var(x, axis=-1, unbiased=False, keepdim=True),
+                paddle.mean(x, axis=-1, keepdim=True),
+            )
+            y = (x - mean) / paddle.sqrt(variance + self.eps)
+        else:
+            y = x
+
+        x_grad = paddle.grad(y, x, create_graph=True)[0]
+
+        return x_grad.mean()
+
+
+class HighOrderCompareNet(HighOrderControlFlowNet):
+    def __init__(self):
+        super().__init__()
+        self.eps = 1e-5
+
+    def forward(self, x):
+        variance, mean = (
+            paddle.var(x, axis=-1, unbiased=False, keepdim=True),
+            paddle.mean(x, axis=-1, keepdim=True),
+        )
+        y = (x - mean) / paddle.sqrt(variance + self.eps)
+
+        x_grad = paddle.grad(y, x, create_graph=True)[0]
+
+        return x_grad.mean()
+
+
+class TestBackwardControlFlow(Dy2StTestBase):
+    @test_ast_only
+    @test_pir_only
+    def test_control_flow_hign_order_backward(self):
+        conf_net = HighOrderControlFlowNet()
+        net = HighOrderCompareNet()
+        x = paddle.rand((5, 5)).astype('float32')
+        x.stop_gradient = False
+        static_net = paddle.jit.to_static(net, full_graph=True)
+        x_grad_grad = static_net(x)
+
+        conf_static_net = paddle.jit.to_static(conf_net, full_graph=True)
+        x_grad_grad_conf = conf_static_net(x)
+
+        np.testing.assert_allclose(
+            x_grad_grad.numpy(), x_grad_grad_conf.numpy()
+        )
+
+
 if __name__ == "__main__":
     unittest.main()

Original file line number	Diff line number	Diff line change
`@@ -42,6 +42,8 @@ bool ParsePlace(const pir::Type& type, OpFuncType* type_) {`
`42`	`42`	`return true;`
`43`	`43`	`}`
`44`	`44`	`}`
	`45`	`+ } else if (!type) {`
	`46`	`+ return false;`
`45`	`47`	`} else {`
`46`	`48`	`PADDLE_THROW(common::errors::PreconditionNotMet(`
`47`	`49`	`"Only support AllocatedDenseTensorType and "`