Using Huggingface Trainer in Colab -> Disk Full

Hello everyone!

I thought I’d post this here first, as I am not sure if it is a bug or if I am doing something wrong.
I’m using the huggingface library to train an XLM-R token classifier. I originally wrote the training routine myself, which worked quite well, but I wanted to switch to the trainer for more advanced features like early stopping and easier setting of training arguments.

To prototype my code, I usually run it on a free google colab account. While the training process works, I’ve had the code crash several times, because the disk space of the Compute Environment runs out. This is NOT my google drive space, but a separate disk of around 60GB space. I have observed, that during training the used space keeps on growing, but I have no idea where or what exactly is writing data. Once the disk is full, this results in the code crashing:

The following are my training parameters/callbacks defined:

## Define Callbacks
class PrinterCallback(TrainerCallback):
    def on_train_begin(self, args, state, control, **kwargs):
        print('\033[1m'+ '=' * 25 + " Model Training " + '=' * 25 + '\033[0m')
    def on_epoch_begin(self, args, state, control, **kwargs):
        print('\n'+ '\033[1m'+ '=' * 25 +' Epoch {:} / {:} '.format(int(trainer.state.epoch) + 1, int(trainer.state.num_train_epochs)) + '=' * 25)

## Training parameters
# training arguments

training_args = TrainingArguments(
    output_dir='./checkpoints',           # output directory
    num_train_epochs=5,              # total # of training epochs
    per_device_train_batch_size=32,    # batch size per device during training
    per_device_eval_batch_size=32,     # batch size for evaluation
    warmup_steps=0,                # number of warmup steps for learning rate scheduler
    weight_decay=0,                   # strength of weight decay
    learning_rate=2e-5,               #2e-5 
    logging_dir='./logs',             # directory for storing logs
    evaluation_strategy= "epoch",     #"steps", "epoch", or "no"
    load_best_model_at_end=False,      #loads the model with the best evaluation score

## Start training

# initialize huggingface trainer
trainer = Trainer(


Any idea what is going wrong here?

Edit: Here is the Error as text from another run; apparently Torch is continuously writing something to disk, but why and what is it?

OSError                                   Traceback (most recent call last)
/usr/local/lib/python3.7/dist-packages/torch/ in save(obj, f, pickle_module, pickle_protocol, _use_new_zipfile_serialization)
    371             with _open_zipfile_writer(opened_file) as opened_zipfile:
--> 372                 _save(obj, opened_zipfile, pickle_module, pickle_protocol)
    373                 return

6 frames
/usr/local/lib/python3.7/dist-packages/torch/ in _save(obj, zip_file, pickle_module, pickle_protocol)
    490         num_bytes = storage.size() * storage.element_size()
--> 491         zip_file.write_record(name, storage.data_ptr(), num_bytes)

OSError: [Errno 28] No space left on device

During handling of the above exception, another exception occurred:

RuntimeError                              Traceback (most recent call last)
<ipython-input-36-3435b262f1ae> in <module>()
----> 1 trainer.train()

/usr/local/lib/python3.7/dist-packages/transformers/ in train(self, resume_from_checkpoint, trial, **kwargs)
   1170                     self.control = self.callback_handler.on_step_end(self.args, self.state, self.control)
-> 1172                     self._maybe_log_save_evaluate(tr_loss, model, trial, epoch)
   1174                 if self.control.should_epoch_stop or self.control.should_training_stop:

/usr/local/lib/python3.7/dist-packages/transformers/ in _maybe_log_save_evaluate(self, tr_loss, model, trial, epoch)
   1268         if self.control.should_save:
-> 1269             self._save_checkpoint(model, trial, metrics=metrics)
   1270             self.control = self.callback_handler.on_save(self.args, self.state, self.control)

/usr/local/lib/python3.7/dist-packages/transformers/ in _save_checkpoint(self, model, trial, metrics)
   1317         elif self.is_world_process_zero() and not self.deepspeed:
   1318             # deepspeed.save_checkpoint above saves model/optim/sched
-> 1319   , os.path.join(output_dir, ""))
   1320             with warnings.catch_warnings(record=True) as caught_warnings:
   1321       , os.path.join(output_dir, ""))

/usr/local/lib/python3.7/dist-packages/torch/ in save(obj, f, pickle_module, pickle_protocol, _use_new_zipfile_serialization)
    371             with _open_zipfile_writer(opened_file) as opened_zipfile:
    372                 _save(obj, opened_zipfile, pickle_module, pickle_protocol)
--> 373                 return
    374         _legacy_save(obj, opened_file, pickle_module, pickle_protocol)

/usr/local/lib/python3.7/dist-packages/torch/ in __exit__(self, *args)
    258     def __exit__(self, *args) -> None:
--> 259         self.file_like.write_end_of_file()
    260         self.buffer.flush()

RuntimeError: [enforce fail at] . unexpected pos 2212230208 vs 2212230096```

There are three default arguments that are relevant here, but seeing that you set save_total_limit=1 I am not sure what else could be being saved…

Can you see what’s actually on the disk?

I’ll try setting save_strategy explicitly to epoch? Probably right now its saving at the preset amount of steps and can’t delete the saved steps from the colab/gdrive disk for whatever reason.

As for

Can you see what’s actually on the disk?

There is a file explorer built into google colab and I can also explore the filesystem through ipython magic (i.e. using bash); but I didn’t really find where exactly the virtual disk for the python environment is mounted and therefore where the trainer is seemingly writing to (even though it should be working on the Google Drive mount).

Edit: I rechecked, and it appears that after running the trainer, /root was slowly filling up on the colab disk; I can however not see the contents of that mount point. Curiously save_total_limit=1 does also not seem to limit the checkpoints saved on my google drive partition, as checkpoints are being stored all 500 steps and only sporadically deleted.

I guess I found the reason,
on deleting the previous checkpoint, it goes to the google drive bin and the bin does not delete it then (deletes after 30 days) and this results in occupied space.
@Nagai-san and @BramVanroy if you can verify this, we can
override the _rotate_checkpoint method of the Trainer to also clean the drive bin. That should resolve the issue

I’m not sure Colab gives enough permission for a program to go delete the drive bin. If that’s possible, we can fix the Trainer directly in the Transformers library (it’s easy to check if we’re in a colab noteboook) but if not, I fear the only solution is to not save anything :-/