Huggingface longformer memory issues

I am building Huggingface Longformer based classifier. My main code below

model = LongformerForSequenceClassification.from_pretrained('/mnt/longformer_official/',
                                                           attention_window = 512)
tokenizer = LongformerTokenizerFast.from_pretrained('/mnt/longformer_official/', max_length = 4000)

train_df_tuning_dataset_tokenized =, batched = True, batch_size = len(train_df_tuning_dataset))

training_args = TrainingArguments(

num_train_epochs = 5,# changed this from 5
per_device_train_batch_size = 4,#4,#8,#adding on 18 march from huggingface example notebook
gradient_accumulation_steps = 16,#16,  #8  adding it back 18 march even though missing in huggingface example notebook as otherwise memory issues

per_device_eval_batch_size= 16,#16
evaluation_strategy = "epoch",

save_strategy = "epoch",#adding on 18 march from huggingface example notebook
learning_rate=2e-5,#adding on 18 march from huggingface example notebook

disable_tqdm = False, 
optim="adamw_torch",#removing on 18 march from huggingface example notebook
run_name = 'longformer-classification-16March2022'

#class weights
class CustomTrainer(Trainer):
    def compute_loss(self, model, inputs, return_outputs=False):
        labels = inputs.get("labels")
        # forward pass
        outputs = model(**inputs)
        logits = outputs.get("logits")
        # compute custom loss (suppose one has 3 labels with different weights)
        loss_fct = nn.CrossEntropyLoss(weight=torch.tensor([1.0, 0.5243])).to(device)
        loss = loss_fct(logits.view(-1, self.model.config.num_labels), labels.view(-1)).to(device)
        return (loss, outputs) if return_outputs else loss

trainer = CustomTrainer(

When I try max_length=1500 in the tokenizer, the code runs fine. It fails when run with max_length=4000
I even tried setting these parameters as
per_device_train_batch_size = 1, gradient_accumulation_steps = 1, per_device_eval_batch_size = 1

My questions:

  1. is it okay to set per_device_train_batch_size = 1, gradient_accumulation_steps = 1, per_device_eval_batch_size = 1?

  2. The error that I get is as below. Is there any way around this other than getting more memory?

    RuntimeError: CUDA out of memory. Tried to allocate 720.00 MiB (GPU 0; 14.76 GiB total capacity; 12.77 GiB already allocated; 111.75 MiB free; 13.69 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF