OOM error in HF trainer during validation

šŸ¤—Transformers
1

Hi, i have seen this issue posted multiple times. And the answer under every post was to use
eval_accumulation_steps. I have set eval_accumulation_steps to 1. and still I got OOM error during validation.

Here is the code

from transformers import AutoModelForSeq2SeqLM, DataCollatorForSeq2Seq, Seq2SeqTrainingArguments, Seq2SeqTrainer
args = Seq2SeqTrainingArguments(
   output_dir="./mymodel",
  group_by_length=True,
  per_device_train_batch_size=32,
  per_device_eval_batch_size=32,
  gradient_accumulation_steps=4,
  evaluation_strategy="steps",
  num_train_epochs=5,
  fp16=True,
  save_steps=10000,
  eval_steps=5000,
  logging_steps=5000,
  learning_rate=3e-4,
  save_total_limit=1,
  predict_with_generate=True,
  eval_accumulation_steps=1,
)

data_collator = DataCollatorForSeq2Seq(tokenizer, model=model)
metric = datasets.load_metric("sacrebleu")


def postprocess_text(preds, labels):
    preds = [pred.strip() for pred in preds]
    labels = [[label.strip()] for label in labels]#ValueError: Got a string but expected a list instead: 'NorveƧ'in rakfisk'i: DĆ¼nyanın en kokulu bal bu mu?'
    return preds, labels
def compute_metrics(eval_preds):
    preds, labels = eval_preds
    if isinstance(preds, tuple):
        preds = preds[0]
    decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
    # Replace -100 in the labels as we can't decode them.
    labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
    decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)

    # Some simple post-processing
    decoded_preds, decoded_labels = postprocess_text(decoded_preds, decoded_labels)
    result = metric.compute(predictions=decoded_preds, references=decoded_labels)
   
    result = {"bleu": result["score"]}
   
    return result


trainer = Seq2SeqTrainer(
    model,
    args,
    train_dataset=tokenized_datasets["train"],
    eval_dataset=tokenized_datasets["test"],
    data_collator=data_collator,
    tokenizer=tokenizer,
    compute_metrics=compute_metrics,
)

What should I do?