# Accelerate inside a notebook cell just ends abruptly without doing anything

**URL:** <https://discuss.huggingface.co/t/accelerate-inside-a-notebook-cell-just-ends-abruptly-without-doing-anything/48864>\
**Category:** 🤗Accelerate\
**Created:** [July 31, 2023, 5:57pm UTC](https://discuss.huggingface.co/t/accelerate-inside-a-notebook-cell-just-ends-abruptly-without-doing-anything/48864 "2023-07-31T17:57:23Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![taksqth](https://avatars.discourse-cdn.com/v4/letter/t/b77776/32.png) [@taksqth](https://discuss.huggingface.co/u/taksqth)\
**Post date:** [July 31, 2023, 5:57pm UTC](https://discuss.huggingface.co/t/accelerate-inside-a-notebook-cell-just-ends-abruptly-without-doing-anything/48864/1 "2023-07-31T17:57:23Z")

</div>

`notebook_launcher(train_and_evaluate, num_processes=2)`

I’m calling the above from within a Kaggle T4x2 notebook and it does start train the model for 2 epochs on both GPUs, but after that training just abruptly ends without logging anything. Is there some way to get more info on what might be going wrong? I’m using a custom training loop similar to the following (the miniai framework has a lot of details hidden, tell me if something is missing that might help):

```auto
def train_and_evaluate():
  from sklearn.model_selection import KFold
  
  kfold = KFold(n_splits=n_splits, shuffle=True)
  
  eval_metrics = []
  for fold, (train_ids, val_ids) in enumerate(kfold.split(df_train)):
      dls = DataLoaders(
          #valid,
          #train,
      )
  
      model = #automodelfrompretrained
  
      lr = #lr
      epochs = #epochs
  
      tmax = epochs * len(dls.train)
      sched = partial(optim.lr_scheduler.OneCycleLR, anneal_strategy='cos', pct_start=0.01, max_lr=lr, total_steps=tmax)
      cbs = [
        HFTrainCB(), #adapt interface for huggingface models and use accelerate
        # Inherits from: https://github.com/johnowhitaker/miniminiai/blob/e84407d11ec2d9d244f7d32b4052b988e887ae0c/miniminiai/miniminiai.py#L384C27-L384C27
        ProgressCB(plot=True), #printprogress
        HFMetricsCB(), #getmetrics
        BatchSchedCB(sched), #step scheduler
      ]
  
      opt_func = partial(optim.AdamW)
      learn = Learner(model, dls, lr=lr, cbs=cbs, opt_func=opt_func)
      learn.fit(epochs)

```