[BUG] Error running ETL with NVTabular tutorial notebook -- version compatibility issue?

Bug description

When running 02-ETL-with-NVTabular.ipynb on REES46 data, the following error was reported:

Failed to transform operator <nvtabular.ops.groupby.Groupby object at 0x7f62fed519a0>
Traceback (most recent call last):
  File "/home/hwan/miniconda3/envs/trans4rec/lib/python3.9/site-packages/nvtabular/workflow/workflow.py", line 519, in _transform_partition
    output_df = node.op.transform(selection, input_df)
  File "/home/hwan/miniconda3/envs/trans4rec/lib/python3.9/contextlib.py", line 79, in inner
    return func(*args, **kwds)
  File "/home/hwan/miniconda3/envs/trans4rec/lib/python3.9/site-packages/nvtabular/ops/groupby.py", line 132, in transform
    new_df = _apply_aggs(
  File "/home/hwan/miniconda3/envs/trans4rec/lib/python3.9/site-packages/nvtabular/ops/groupby.py", line 244, in _apply_aggs
    df[f"{col}{name_sep}{_agg}"] = _first_or_last(
  File "/home/hwan/miniconda3/envs/trans4rec/lib/python3.9/site-packages/nvtabular/ops/groupby.py", line 288, in _first_or_last
    return _first(x)
  File "/home/hwan/miniconda3/envs/trans4rec/lib/python3.9/site-packages/nvtabular/ops/groupby.py", line 302, in _first
    return elements[offsets[:-1]]
TypeError: 'NumericalColumn' object is not subscriptable
....
File ~/miniconda3/envs/trans4rec/lib/python3.9/site-packages/nvtabular/ops/groupby.py:302, in _first(x)
    300     offsets = x.list._column.offsets
    301     elements = x.list._column.elements
--> 302     return elements[offsets[:-1]]
    303 else:
    304     # cpu/pandas
    305     return x.apply(lambda y: y[0])

TypeError: 'NumericalColumn' object is not subscriptable

Steps/Code to reproduce bug

1. Loading REES Oct-2019.parquet into examples/tutorial/02-ETL-with-NVTabular.ipynb
2. Run the cells until hitting the cell right beneath Initialize the NVTabular dataset object and workflow graph.

Expected behavior

The workflow is expected to fit and transform the data without error.

Environment details

• Transformers4Rec version: The current codebase from Github (commit 0fd858dcee5ce5b2508f7e91733ae68710affd76)
• Platform: Ubuntu 18.04
• Python version: 3.9
• Huggingface Transformers version: 4.18
• PyTorch version (GPU?): 1.12 (GPU)
• Tensorflow version (GPU?):
• NVTabular version: 1.3.3
• CUDF version: 22.08 (cuda version, from rapidsai)

Additional context

I am suspecting it might be compatibility issues related to NVTabular and CUDF versions. Much appreciate your help! Thanks!

[benfred]

I believe this issue is fixed with https://github.com/NVIDIA-Merlin/NVTabular/pull/1654 - this will be in the next release of nvtabular

[hui-wan]

Update: The issue went away with the installation of the latest NVTabular version 1.4.0 released on 09/06/2022.

Thank you for your help @benfred @rnyak!