bug描述 Describe the Bug

复现方法：

• export GLOG_v=10 in terminal
• run following script:

import paddle
from paddle.jit import to_static

class SimpleNet(paddle.nn.Layer):
    def __init__(self):
        super(SimpleNet, self).__init__()
        self.linear = paddle.nn.Linear(10, 3)

    @to_static # 动静转换
    def forward(self, x, y):
        out = self.linear(x)
        out = out + y
        return out

net = SimpleNet()
net.train()
optimizer = paddle.optimizer.Adam(learning_rate=0.001, parameters=net.parameters())
x = paddle.rand([2, 10])
y = paddle.rand([2, 3])
out = net(x, y)
out.backward()
optimizer.step()
optimizer.clear_grad()
print(out)

报错信息：

Traceback (most recent call last):
  File "test_log.py", line 21, in <module>
    out.backward()
  File "<decorator-gen-139>", line 2, in backward
  File "/root/miniconda3/envs/eager/lib/python3.7/site-packages/paddle/fluid/wrapped_decorator.py", line 26, in __impl__
    return wrapped_func(*args, **kwargs)
  File "/root/miniconda3/envs/eager/lib/python3.7/site-packages/paddle/fluid/framework.py", line 507, in __impl__
    return func(*args, **kwargs)
  File "/root/miniconda3/envs/eager/lib/python3.7/site-packages/paddle/fluid/dygraph/varbase_patch_methods.py", line 297, in backward
    framework._dygraph_tracer())
RuntimeError: (NotFound) Variable is not initialized.
  [Hint: holder_ should not be null.] (at ../paddle/fluid/framework/variable.h:72)

原因：

dy2static为了支持cuda, run_program_op 里面传入了一个未初始化的var _cuda_graph_vec ：

Paddle/python/paddle/fluid/dygraph/dygraph_to_static/partial_program.py

Line 366 in e379455

| | self._double_grads, self._cuda_graph_vec, *attrs) |

在反向的时候，打开Glog的情况下，由于变量 cuda_graph 未初始化，var内部的 holder_ 为空指针，LOG获取varType的时候会assert fail。

Paddle/paddle/fluid/framework/variable.h

Line 137 in e379455

| | << platform::demangle(framework::ToTypeName(Type())); |

可能的修复

在上述代码中增加 holder_==nullptr 的判断分支。

其他补充信息 Additional Supplementary Information

No response