Inference gives different results when using multiple gpus (distributed mode) vs just one gpu (not distributed mode)

[ThuongTNguyen]

Hi, I want to report an issue observed when running inference for a classification task.

Description

When running inference (either do_eval=True or do_predict=True), the results are different whether it's in distributed mode (multiple gpus ) or not (one gpu).

When doing evaluation, data is prepared sequentially in batches using SequentialSampler, BatchSampler, and DistributedBatchSampler https://github.com/microsoft/DeBERTa/blob/4d7fe0bd4fb3c7d4f4005a7cafabde9800372098/DeBERTa/apps/run.py#L172 and then sent to gpus. But once logits are computed, there is a step to gather results across devices - merge_distributed. https://github.com/microsoft/DeBERTa/blob/4d7fe0bd4fb3c7d4f4005a7cafabde9800372098/DeBERTa/apps/run.py#L228 After this step, the order of data instances is no longer similar to the one in the original input file (dev.tsv or test.tsv) in case of distributed modes with multiple gpus, resulting in a different accuracy.

Steps to reproduce

• Use cola.sh to evaluate --init_model deberta-v3-large
• Set other paramaters: --do_eval, --eval_batch_size 4, using 3 gpus CUDA_VISIBLE_DEVICES=7,5,6 or 1 gpu CUDA_VISIBLE_DEVICES=7

• Check oder of instances after logit calculation is done and collected from all gpus. For example, print the first 10 instances in predicts and labels after the above merge_distributed

  • 1 gpu case - as expected according to the input file:

    • labels = [1 1 1 1 0 0 0 1 1 1]
    • predicts = [[-0.04428 -0.167 ] [-0.05667 -0.1713 ] [-0.0438 -0.1727 ] [-0.0396 -0.1794 ] [-0.03604 -0.1823 ] [-0.0433 -0.1809 ] [-0.01921 -0.1947 ] [-0.04788 -0.1741 ] [-0.05774 -0.1755 ] [-0.05173 -0.1692 ]]
    • accuracy = 0.3087248322147651
    • eval_loss = 0.7155781315204284

  • 3 gpu case:

    • labels = [1 1 1 1 1 1 0 0 0 1]
    • predicts =[[-0.04428 -0.167 ] [-0.05667 -0.1713 ] [-0.0438 -0.1727 ] [-0.0396 -0.1794 ] [-0.04428 -0.167 ] [-0.04428 -0.167 ] [-0.03604 -0.1823 ] [-0.0433 -0.1809 ] [-0.01921 -0.1947 ] [-0.04788 -0.1741 ]]
    • accuracy = 0.3231064237775647
    • eval_loss = 0.7120015648589737

Additional information

My system setup is:

• PyTorch 1.10.0+cu113
• 8 GPUs: NVIDIA GeForce GTX 1080 Ti