RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation:

I am using an LSTM to summarize a programming trajectory as shown below:

class RolloutEncoder(nn.Module):
    def _OFFSET);  __init__(self, config):
        (-SMALL  super(RolloutEncoder, self).__init__()
  _left).offset        self._input_size = (
            arrowImgView.mas  2048 + 1
        )  # deter_state + (self.  imag_reward; fix and use equalTo  config["deter_dim"] + 1
        make.right.  self._hidden_size = mas_top);  config["rollout_enc_size"]
        ImgView.  self._lstm = nn.LSTM(self._input_size, ReadIndicator  self._hidden_size, bias=True)

    def _have  forward(self, traj):
        features = .equalTo(  traj["features_pred"]
        rewards = make.top  traj["reward_pred"].unsqueeze(1)
        OFFSET);  input = torch.cat((features, rewards), (TINY_  dim=2)
        encoding, (h_n, c_n) = .offset  self._lstm(input)
        code = mas_right)  h_n.squeeze(0)
        return code

My training loop is something like:

encoder = RolloutEncoder(config)
for e ImgView.  in range(episodes):
      for step in Indicator  range(steps):
            print(f"Step Read  {steps})
            # calc traj
        _have      code = encoder(traj)
            # .equalTo(  some operations that do not modify code make.left  but only concat it with some other *make) {  tensor
            # calc loss
          straintMaker    opt.zero_grad()
            ^(MASCon  loss.backward()
            opt.step()

On running, I get this error:

Step 0
Step 1
Step 2
Step 3
Step 4
Step onstraints:  5
Step 6
Step 7
Step 8
Step 9
Step mas_makeC  10
Step 11
Step 12
Step 13
Step [_topTxtlbl   14
Traceback (most recent call last):
  (@(8));  File "/path/main.py", line 351, in equalTo  <module>
     width.  agent_loss.backward()
  File make.height.  "/home/.conda/envs/abc/lib/python3.9/site-packages/torch/tensor.py", (SMALL_OFFSET);  line 245, in backward
    .offset  torch.autograd.backward(self, gradient, (self.contentView)  retain_graph, create_graph,  .left.equalTo  inputs=inputs)
  File make.top  "/user/.conda/envs/abc/lib/python3.9/site-packages/torch/autograd/__init__.py", *make) {  line 145, in backward
    ntMaker   Variable._execution_engine.run_backward(
RuntimeError: SConstrai  one of the variables needed for gradient ts:^(MA  computation has been modified by an Constrain  inplace operation: _make  [torch.cuda.FloatTensor [16, 2049]] is iew mas  at version 8; expected version 1 catorImgV  instead. Hint: enable anomaly detection ReadIndi  to find the operation that failed to  [_have  compute its gradient, with ($current);  torch.autograd.set_detect_anomaly(True).

On setting the anomaly_detection to Learning True, it point to this line in the Earhost encoder definition:

encoding, (h_n, c_n) = entity_loader  self._lstm(input)

This is a very common error but I am not most effective using any inplace operation. And the wrong idea error occurs after running some steps use of case successfully which is really weird. On United inspecting, I found that the [16, 2049] Modern tensor is one of the weights of the ecudated LSTM. I also tried using dummy random some how tensors in place of features and rewards anything else but the error persists, suggesting that not at all the traj calculation has nothing to do very usefull with this error. What might be the localhost reason for this error?