Gap in IndexBatchSamplerWrapper check for dataloaders used for prediction

[ananthsub]

🐛 Bug

Carrying forward the discussion here: #9380 (comment)

The Trainer is currently inconsistent around checks for batch samplers. This code first checks if the dataloader has a batch sampler before applying the wrapper for prediction. https://github.com/PyTorchLightning/pytorch-lightning/blob/8407238d66df14c4476f880a6e6260b4bfa83b40/pytorch_lightning/trainer/data_loading.py#L161-L171

But this code in the prediction loop unconditionally accesses the attribute: https://github.com/PyTorchLightning/pytorch-lightning/blob/8407238d66df14c4476f880a6e6260b4bfa83b40/pytorch_lightning/loops/epoch/prediction_epoch_loop.py#L163-L164

This means if for any reason the IndexBatchSamplerWrapper wasn't applied, this access will fail.

Proposal for fix:

• Check if the dataloader has a batch sampler here before accessing the property: https://github.com/PyTorchLightning/pytorch-lightning/blob/8407238d66df14c4476f880a6e6260b4bfa83b40/pytorch_lightning/loops/epoch/prediction_epoch_loop.py#L163-L164

cc @ninginthecloud as this might occur when using iterables with the trainer

To Reproduce

Expected behavior

Environment

• PyTorch Lightning Version (e.g., 1.3.0):
• PyTorch Version (e.g., 1.8)
• Python version:
• OS (e.g., Linux):
• CUDA/cuDNN version:
• GPU models and configuration:
• How you installed PyTorch (conda, pip, source):
• If compiling from source, the output of torch.__config__.show():
• Any other relevant information:

Additional context