Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Division by zero error when calling find_packing_max_batch_len_and_grad_accum #254

Open
RobotSail opened this issue Oct 7, 2024 · 0 comments

Comments

@RobotSail
Copy link
Member

It seems like there's a division by zero error that can still occur with the right conditions.

[rank0]: Traceback (most recent call last):
[rank0]:   File "/home/oleg/Programming/training/src/instructlab/training/main_ds.py", line 934, in <module>
[rank0]:     main(args)
[rank0]:   File "/home/oleg/Programming/training/src/instructlab/training/main_ds.py", line 546, in main
[rank0]:     packing_max_batch_len, grad_accum = find_packing_max_batch_len_and_grad_accum(
[rank0]:                                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/home/oleg/Programming/training/src/instructlab/training/multipack_sampler.py", line 160, in find_packing_max_batch_len_and_grad_accum
[rank0]:     packing_max_batch_len = find_max_pack_len_with_padding(
[rank0]:                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/home/oleg/Programming/training/src/instructlab/training/multipack_sampler.py", line 86, in find_max_pack_len_with_padding
[rank0]:     avg_bs_per_minibatch = get_effective_samples_per_minibatch(
[rank0]:                            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/home/oleg/Programming/training/src/instructlab/training/multipack_sampler.py", line 79, in get_effective_samples_per_minibatch
[rank0]:     return len(dataset) / len(batches)
[rank0]:            ~~~~~~~~~~~~~^~~~~~~~~~~~~~
[rank0]: ZeroDivisionError: division by zero

To reproduce:

  1. Run a model train with the community dataset
  2. Set your TrainingArgs like so:
train_args = TrainingArgs(
    data_path="path/to/data",
    is_padding_free=False,
    ckpt_output_dir='checkpoints',
    data_output_dir='/dev/shm',
    max_seq_len=70,
    max_batch_len=70,
    effective_batch_size=3840,
    save_samples=20_000,
)
ashna000 added a commit to ashna000/instructlab-training that referenced this issue Oct 15, 2024
Signed-off-by: ashna000 <[email protected]>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

1 participant