静态图高阶微分计算结果不符合预期

[HydrogenSulfate]

bug描述 Describe the Bug

import paddle
from paddle import nn


class Model(nn.Layer):
    def __init__(self):
        super().__init__()
        self.fc1 = nn.Linear(2, 1)

    def forward(self, x, y):
        z = paddle.concat([x, y], axis=-1)
        z = self.fc1(z)
        z = nn.functional.silu(z)

        gx = paddle.grad(z, x, create_graph=True)[0]
        return paddle.abs(gx).pow(2).sum()

model = Model()

# data
N = 2
x = paddle.randn([N, 1])
x.stop_gradient = False
y = paddle.randn([N, 1])
y.stop_gradient = False

# dynamic
loss_dynamic = model(x, y)
loss_dynamic.backward()

grads1 = [param.grad.clone().detach() for param in model.parameters()]

model.clear_gradients()

# static
st_model = paddle.jit.to_static(model)

loss_static = st_model(x, y)
loss_static.backward()

grads2 = [param.grad.clone().detach() for param in model.parameters()]

# check
import numpy as np
for p1, p2 in zip(grads1, grads2):
    np.testing.assert_allclose(p1.numpy(), p2.numpy(), 1e-3, 1e-3)

运行命令：

FLAGS_prim_all=False FLAGS_enable_pir_in_executor=true FLAGS_enable_pir_api=True python test_run.py

结果：

AssertionError: 
Not equal to tolerance rtol=0.001, atol=0.001

Mismatched elements: 8 / 8 (100%)
Max absolute difference: 0.08932383
Max relative difference: 3.0927124
 x: array([[ 0.050079,  0.003131,  0.12114 , -0.012295],
       [-0.106463,  0.056415, -0.26742 ,  0.029972]], dtype=float32)
 y: array([[ 0.012236, -0.017195,  0.031816, -0.003915],
       [-0.068708,  0.077533, -0.178814,  0.021687]], dtype=float32)

由于silu没有二阶大算子，因此运行时预期应该报错，而不是仍然进行计算并给出错误结果；

如果把激活函数从silu换成配二阶算子的tanh，则可以通过测试。

其他补充信息 Additional Supplementary Information

No response

[SigureMo]

会在 #63914 解决～