RAM OOM with a large save_steps using trainer API for MLM training

[Sumsky21]

Environment info

• transformers version: 4.17.0
• Platform: Linux-4.14.105-1-tlinux3-0013-x86_64-with-glibc2.17
• Python version: 3.9.2
• PyTorch version (GPU?): 1.10.2+cu113 (True)
• Tensorflow version (GPU?): not installed (NA)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Using GPU in script?: yes
• Using distributed or parallel set-up in script?: yes (in fact the trainer API handles it automatically, with 8 gpus)

Who can help

Electra model: @LysandreJik
Trainer: @sgugger

Information

Model I am using (Bert, XLNet ...): Electra(ForMaskedLM)

The problem arises when using:

• [ ] the official example scripts: (give details below)
• [x] my own modified scripts: (give details below)

The tasks I am working on is:

• [x] an official task: Masked LM (training from scratch)
• [ ] my own task or dataset: (give details below)

Expected behavior

I don't know if it was a bug, so I'll describe the situation first.

I'm using the trainer API to train a ElectraForMaskedLM model from scratch. I write a simple script (will be shown below) and set save_step argument as 10000. But after running the script, I observe the RAM(memory, NOT CUDA) usage continuously increases, finally casued an OOM error thus terminate the docker. At the same time, GPU, CUDA memory and CPU usage is quite normal.

To solve the problem, I've tried adjusting several training arguments, and then find that when I set a smaller save_step (like 2000), RAM usage goes up and down periodically, and the cycle length is approximately equal to the interval between two save checkpoints, like the figure below:

So is there a bug about the checkpoint saving mechanism? Or it's normal so I should have set a small save_steps?

The training script is below:

import torch
import numpy as np
from transformers import ElectraConfig, ElectraTokenizerFast, ElectraForMaskedLM
import os
os.environ["TOKENIZERS_PARALLELISM"] = "true"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# device = "cpu"
print(device)

# Tokenizer
print("Importing tokenizer and model...")
tokenizer = ElectraTokenizerFast(tokenizer_file='./models/bert-custom.json')
# Model
config = ElectraConfig(vocab_size=10000)
model = ElectraForMaskedLM(config=config).to(device)

# dataset
print("Importing dataset...")
from transformers import LineByLineTextDataset
dataset = LineByLineTextDataset(
    tokenizer=tokenizer,
    file_path="./data/in_neg_1e7.csv",
    block_size=80,
)
eval_set = LineByLineTextDataset(
    tokenizer=tokenizer,
    file_path="./data/in_eval_1e6.csv",
    block_size=80,
)

from transformers import DataCollatorForLanguageModeling
data_collator = DataCollatorForLanguageModeling(
    tokenizer=tokenizer, mlm=True, mlm_probability=0.15
)

from transformers import Trainer, TrainingArguments, EarlyStoppingCallback
print("Start Training...")
training_args = TrainingArguments(
    output_dir="./electra_save/",
    overwrite_output_dir=True,
    num_train_epochs=50,
    per_device_train_batch_size=512,
    per_device_eval_batch_size=512,
    eval_accumulation_steps=300,
    save_steps=10000,                             # the key argument
    save_total_limit=2,
    dataloader_num_workers=80,
    evaluation_strategy='steps', 
    eval_steps=10000,
    logging_steps=1000,
    metric_for_best_model='eval_loss',
    load_best_model_at_end=True,
)

trainer = Trainer(
    model=model,
    args=training_args,
    data_collator=data_collator,
    tokenizer=tokenizer,
    train_dataset=dataset,
    eval_dataset=eval_set,
    callbacks=[EarlyStoppingCallback(early_stopping_patience=10)],
)

trainer.train()

trainer.save_model(output_dir='./electra_save/')

To reproduce

Steps to reproduce the behavior:

1. run the training script above with a large dataset (I use a dataset with 10M lines of text) and a large save_steps argument
2. wait for a period of time and monitor on the memory usage
3. encounter with OOM

[Sumsky21]

Addition: Previously I've used a similar script to train BertForMaskedLM, and no memory fluctuations were observed that time: That's why I suspect it's abnormal this time.

[sgugger]

So the issue only appears with Electra, right? Did you encounter it for any other model?

[Sumsky21]

Yes. I've only tried on Bert and Electra so far (and there's no problem during Bert training, as above figure has shown).

If there's a need, I can try the similar script on RobertaforMaskedLM in next few days.

[sgugger]

That would be helpful if you could!

[github-actions[bot]]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.