[data_loader] expand the error message

[stas00]

What does this PR do?

When using deepspeed with a DL that doesn't have a .batch_size (e.g. due to using a batch sampler) it fails with:

[...]
[:0]:    ) = self.accelerator.prepare(
[:0]:  File "/mnt/nvme0/code/huggingface/accelerate-ds-bs-auto/src/accelerate/accelerator.py", line 1209, in prepare
[:0]:    result = self._prepare_deepspeed(*args)
[:0]:  File "/mnt/nvme0/code/huggingface/accelerate-ds-bs-auto/src/accelerate/accelerator.py", line 1419, in _prepare_deepspeed
[:0]:    result = [
[:0]:  File "/mnt/nvme0/code/huggingface/accelerate-ds-bs-auto/src/accelerate/accelerator.py", line 1420, in <listcomp>
[:0]:    self._prepare_one(obj, first_pass=True) if isinstance(obj, torch.utils.data.DataLoader) else obj
[:0]:  File "/mnt/nvme0/code/huggingface/accelerate-ds-bs-auto/src/accelerate/accelerator.py", line 1092, in _prepare_one
[:0]:    return self.prepare_data_loader(obj, device_placement=device_placement)
[:0]:  File "/mnt/nvme0/code/huggingface/accelerate-ds-bs-auto/src/accelerate/accelerator.py", line 1791, in prepare_data_loader
[:0]:    prepared_data_loader = prepare_data_loader(
[:0]:  File "/mnt/nvme0/code/huggingface/accelerate-ds-bs-auto/src/accelerate/data_loader.py", line 863, in prepare_data_loader
[:0]:    new_batch_sampler = BatchSamplerShard(
[:0]:  File "/mnt/nvme0/code/huggingface/accelerate-ds-bs-auto/src/accelerate/data_loader.py", line 155, in __init__
[:0]:    raise ValueError("You need to use `even_batches=False` when the batch sampler has no batch size.")
[:0]:ValueError: You need to use `even_batches=False` when the batch sampler has no batch size.

It's impossible to act on the error message's suggestion, since the deepspeed plugin doesn't expose this parameter to the user. But instead it uses accelerator.even_batches.

So I expanded the error message to suggest to do that.

Of course @pacman100 may consider to do the work for the user and if .batch_size isn't available then to set the correct flag automatically.

Who can review?

• DeepSpeed: @pacman100
• Core parts of the library: @muellerzr @BenjaminBossan

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[BenjaminBossan]

Thanks for improving the error message, LGTM.

[pacman100]

Thank you @stas00 for clarifying the error message!