Autograd issues with GRU

[cwoolfo1]

Hello, I am making a gated recurrent unit from scratch and I am having trouble getting the autograd package to calculate the gradients correctly.

Here is the error I am getting:
MXNetError: Check failed: !AGInfo: :IsNone(*i): Cannot differentiate node because it is not in a computational graph. You need to set is_recording to true or use autograd.record() to save computational graphs for backward. If you want to differentiate the same graph twice, you need to pass retain_graph=True to backward.

Have one of yall had this issue before? How can I go about fixing it?

Attached below is the code I am using to compute the gradients. All the weights and biases have had .attach_grad() called on them when I initialized them.

def ComputeGradient(self, X):
        with autograd.record():
            y = self.forward(X)
            y.backward()

    def forward(self, X):
        HiddenStateVector = pickle.load(open(self.HiddenStateVectorPath, 'rb'))
        # Loads previous hidden state vector into memory
        
        UpdateWeightInput, UpdateWeightHidden, UpdateBias = self.GetWeights('Update')
        UpdateGate = mx.nd.sigmoid(mx.nd.dot(X,UpdateWeightInput) + mx.nd.dot(HiddenStateVector, UpdateWeightHidden) + UpdateBias)
        del UpdateWeightInput, UpdateWeightHidden, UpdateBias
        # Computes the update gate
        
        ResetWeightInput, ResetWeightHidden, ResetBias = self.GetWeights('Reset')
        ResetGate = mx.nd.sigmoid(mx.nd.dot(X, ResetWeightInput) + mx.nd.dot(HiddenStateVector, ResetWeightHidden) + ResetBias)
        del ResetWeightInput, ResetWeightHidden, ResetBias
        # Computes the reset gate
        
        HiddenWeightInput, HiddenWeightHidden, HiddenBias = self.GetWeights('Hidden')
        HiddenGate = mx.nd.tanh(mx.nd.dot(X, HiddenWeightInput) + mx.nd.dot(ResetGate * HiddenStateVector, HiddenWeightHidden) + HiddenBias)
        del HiddenWeightInput, HiddenWeightHidden, HiddenBias, ResetGate
        # Computes the hidden gate
        
        HiddenStateVector = UpdateGate * HiddenStateVector + (1-UpdateGate) * HiddenGate
        pickle.dump(HiddenStateVector, open(self.HiddenStateVectorPath, 'wb'))
        # Computes the hidden state vector and saves it to memory
        output = pickle.load(open(self.outputpath, 'rb'))

        outputBias = pickle.load(open(self.outputBiaspath, 'rb'))
        # Computes the output and output bias

        return mx.nd.dot(HiddenStateVector, output) + outputBias
    # Forward pass through the neural network