huggingface config_error ai_generated true

RuntimeError: The number of batches in the dataloader (127) is not divisible by gradient_accumulation_steps (8). This may cause uneven gradient updates.

ID: huggingface/gradient-accumulation-steps-mismatch

Also available as: JSON · Markdown · 中文
85%Fix Rate
84%Confidence
1Evidence
2024-02-10First Seen

Version Compatibility

VersionStatusIntroducedDeprecatedNotes
huggingface/transformers 4.37.0 active — — —
huggingface/accelerate 0.27.0 active — — —
huggingface/transformers 4.38.0 active — — —

Root Cause

The total number of batches per epoch is not a multiple of gradient_accumulation_steps, leading to a partial update at the end of each epoch.

generic

中文

每个 epoch 的总批次数不是 gradient_accumulation_steps 的倍数，导致每个 epoch 结束时出现部分更新。

Official Documentation

https://huggingface.co/docs/transformers/en/main_classes/trainer#gradient-accumulation

Workarounds

  1. 90% success Adjust the batch size or dataset size so that total batches per epoch is a multiple of gradient_accumulation_steps. For example, set `per_device_train_batch_size=8` and `gradient_accumulation_steps=4` if dataset has 256 samples: `TrainingArguments(per_device_train_batch_size=8, gradient_accumulation_steps=4)` 
    Adjust the batch size or dataset size so that total batches per epoch is a multiple of gradient_accumulation_steps. For example, set `per_device_train_batch_size=8` and `gradient_accumulation_steps=4` if dataset has 256 samples: `TrainingArguments(per_device_train_batch_size=8, gradient_accumulation_steps=4)`
  2. 80% success Use `DataLoader` with `drop_last=True` and ensure dataset length is a multiple of batch_size * gradient_accumulation_steps: `train_dataset = train_dataset.select(range(len(train_dataset) - len(train_dataset) % (batch_size * grad_acc_steps)))` 
    Use `DataLoader` with `drop_last=True` and ensure dataset length is a multiple of batch_size * gradient_accumulation_steps: `train_dataset = train_dataset.select(range(len(train_dataset) - len(train_dataset) % (batch_size * grad_acc_steps)))`

中文步骤

  1. Adjust the batch size or dataset size so that total batches per epoch is a multiple of gradient_accumulation_steps. For example, set `per_device_train_batch_size=8` and `gradient_accumulation_steps=4` if dataset has 256 samples: `TrainingArguments(per_device_train_batch_size=8, gradient_accumulation_steps=4)`
  2. Use `DataLoader` with `drop_last=True` and ensure dataset length is a multiple of batch_size * gradient_accumulation_steps: `train_dataset = train_dataset.select(range(len(train_dataset) - len(train_dataset) % (batch_size * grad_acc_steps)))`

Dead Ends

Common approaches that don't work:

  1. Set `dataloader_drop_last=True` in TrainingArguments to drop the last incomplete batch 70% fail

    This only drops the last batch, but the total batch count may still not be divisible by gradient_accumulation_steps; it only works if the dataset size is a multiple of batch_size.

  2. Increase gradient_accumulation_steps to a larger number 80% fail

    Making it larger usually worsens the divisibility problem and increases memory usage; it doesn't address the root cause.