Tensorflow tfcompile: fetching gradients - tensorflow

Question

I created a very simple tensorflow model where I fetch gradients:
# tf Graph Input
X = tf.placeholder(tf.float32, [1, 2], name="X")
Y = tf.placeholder(tf.float32, [1, 2], name="Y")
# Model parameter variables
W = tf.Variable([[1.0, 2.0], [3.0, 4.0]], name="weight")
B = tf.Variable([[5.0, 6.0]], name="bias")
# Construct a multivariate linear model
matmul = tf.matmul(X, W, name="matrixMul")
pred = tf.add(matmul, B, name="addition")
# Mean squared error
cost = tf.reduce_sum(tf.pow(pred-Y, 2) / 2 )
# Fetch gradients
grads = tf.gradients(cost, [W, B])
I exported this graph into a protobuf and now I use tfcompile for AOT compilation. I want to use the compiled graph in a C++ program and fetch the computed gradients.
The config file for tfcompile looks like:
feed {
id { node_name: "X" }
shape {
dim { size: 1 }
dim { size: 2 }
}
name: "x"
}
feed {
id { node_name: "Y" }
shape {
dim { size: 1 }
dim { size: 2 }
}
name: "y"
}
feed {
id { node_name: "weight" }
shape {
dim { size: 2 }
dim { size: 2 }
}
name: "w"
}
feed {
id { node_name: "bias" }
shape {
dim { size: 1 }
dim { size: 2 }
}
name: "b"
}
fetch {
id { node_name: "addition"}
name: "prediction"
}
fetch {
id { node_name: "gradients/matrixMul_grad/MatMul_1"}
name: "weight_grad"
}
fetch {
id { node_name: "gradients/addition_grad/Reshape"}
name: "bias_grad"
}
Finally I run this C++ code:
obj.set_arg_x_data(x.data());
obj.set_arg_y_data(y.data());
obj.set_arg_w_data(w.data());
obj.set_arg_b_data(b.data());
obj.Run();
std::cout << "result_prediction =" << std::endl << obj.result_prediction(0,0) << " " << obj.result_prediction(0,1) << std::endl;
std::cout << "result_weight_grad =" << std::endl << obj.result_weight_grad(0,0) << " " << obj.result_weight_grad(0,1) << " " << obj.result_weight_grad(1,0) << " " << obj.result_weight_grad(1,1) << std::endl;
std::cout << "result_bias_grad =" << std::endl << obj.result_bias_grad(0,0) << " " << obj.result_bias_grad(0,1) << std::endl;
For result_prediction and result_bias_grad I get the expected values.
Just for result_weight_grad I get only 0,0,0,0.
Maybe I am fetching there the wrong node:
fetch {
id { node_name: "gradients/matrixMul_grad/MatMul_1"}
name: "weight_grad"
}
Does somebody tried already to fetch computed gradients? Tensorflow only offers examples where they using tfcompile for prediction.