Exception encountered when calling layer 'tft_multi_head_attention' (type TFTMultiHeadAttention)

Question

I am trying to build a forecasting model with tft module with Temporal Fusion Transformer,I am getting below error when I am trying to train the model, since I am new to tensorflow, I can't understand fully what does it mean. I thought that simply calling the model train will train the model.

# Train Model
model.train(train_dataset = trainset,             # trainset obtain from data_objec using the dataobj.train_test_dataset() method 
            test_dataset = testset,              # testset obtain from data_objec using the dataobj.train_test_dataset() method
            loss_function = loss_fn,             # Any supported loss function defined in tft.supported_losses
            metric='MSE',              # Either 'MSE' or 'MAE'
            learning_rate=0.0001,      # Use higher lr only with valid clipnorm
            max_epochs=100,
            min_epochs=10,       
            prefill_buffers=False,     # Indicates whether to create a static dataset (requires more memory but trains faster)
            num_train_samples=200000,  # (NOT USED if prefill_buffers=False)
            num_test_samples=50000,    # (NOT USED if prefill_buffers=False)
            train_batch_size=64,       # (NOT USED if prefill_buffers=False, Batch Size specified in data object is used instead) 
            test_batch_size=128,        # (NOT USED if prefill_buffers=False, Batch Size specified in data object is used instead) 
            train_steps_per_epoch=200, # (NOT USED if prefill_buffers=True)
            test_steps_per_epoch=100,  # (NOT USED if prefill_buffers=True)
            patience=10,               # Max epochs to train without further drop in loss value (use higher patience when prefill_buffers=False)
            weighted_training=False,   # Whether to compute & optimize on the basis of weighted losses 
            model_prefix='./tft_model',
            logdir='/tmp/tft_logs',
            opt=None,                  # provide own optimizer object (default is Adam/Nadam)             
            clipnorm=0.1,              # max global norm applied. Used for stable training. Default is 'None'.
            min_delta=0.0001,          # min decrease in val. loss to be considered an improvement 
            shuffle=True)