Ml Agents release 19 training is freezing. Unity out of memory.

Hi, I'm having trouble training my 5dof robotic arm. I upgraded to release 19 branch, now I get this error and I still run out of memory:

  \ml-agents\mlagents\trainers\torch\networks.py:91: UserWarning: Creating a tensor from a list of numpy.ndarrays is extremely slow. Please consider converting the list to a single numpy.ndarray with numpy.array() before converting to a tensor. (Triggered internally at  C:\Users\builder\tkoch\workspace\pytorch\pytorch_1647970138273\work\torch\csrc\utils\tensor_new.cpp:201.)
    enc.update_normalization(torch.as_tensor(vec_input))

I don't know why this error comes out now but it may be related to the laggy (10 robot arm training at 19fps in the editor)

here is the config for the trainer:

  [INFO] Hyperparameters for behavior name arm:
          trainer_type:   ppo
          hyperparameters:
            batch_size:   1024
            buffer_size:  10240
            learning_rate:        0.0003
            beta: 0.005
            epsilon:      0.2
            lambd:        0.95
            num_epoch:    7
            learning_rate_schedule:       linear
            beta_schedule:        linear
            epsilon_schedule:     linear
          network_settings:
            normalize:    True
            hidden_units: 32
            num_layers:   2
            vis_encode_type:      simple
            memory:       None
            goal_conditioning_type:       hyper
            deterministic:        False
          reward_signals:
            extrinsic:
              gamma:      0.995
              strength:   1.0
              network_settings:
                normalize:        False
                hidden_units:     128
                num_layers:       2
                vis_encode_type:  simple
                memory:   None
                goal_conditioning_type:   hyper
                deterministic:    False
          init_path:      results\scenario1\arm\checkpoint.pt
          keep_checkpoints:       5
          checkpoint_interval:    500000
          max_steps:      30000000
          time_horizon:   250
          summary_freq:   30000
          threaded:       False
          self_play:      None
          behavioral_cloning:     None

running python 3.7.13.

Anybody have a clue ?`enter code here`