AttributeError in collator.py with both preprocessed and from-scratch data

[c1ya0]

Hi,

After a long debugging process for the environment setup, I'm now able to start the training script on a 2x RTX 4090 workstation.

However, the training consistently crashes at the beginning of epoch 1 with an AttributeError: 'NoneType' object has no attribute 'max' in fairseq/data/graph/collator.py.

Crucially, this error is reproducible in two scenarios:

1. When using the preprocessed pcqm4mv2-pos.zip downloaded from the link in the README.
2. When downloading the raw OGB data and letting the script preprocess it from scratch.

Since the error is identical in both cases, it strongly points to a bug in the data loading or collation logic that fails on a specific sample in the dataset.

Could you please look into this code-level issue?
Thank you!