Fix support with cuDF 2602.

[trivialfis]

Close #12138

This PR drops support for older versions of cuDF, in exchange for cleaner code without the arrow C array parsing. Shared by @mroeschke

• Fix validity mask handling.
• Replace arrow c device array with CUDA array interface.

todos:

• Revert CI tag.

[Copilot]

Pull request overview

Fixes a cuDF 26.02 compatibility break that caused XGBoost to crash when ingesting cuDF categorical columns on GPU by adapting how category columns are converted to pylibcudf / Arrow device arrays.

Changes:

• Update cudf_cat_inf to call to_pylibcudf() without the removed mode argument on newer cuDF versions.
• Add a fallback path for older cuDF versions that still require/accept mode="read".

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated 1 comment.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[mroeschke]

Just curious about the use case for this call.

IIUC at this point, this function wants to return the cuda array interface for categories that are strings but it appears you are going through the arrow c (device) interface to get there?

[trivialfis]

Thank you for looking into this. Yes, that's the intention of the function call. At the time this was suggested to me in an offline conversation, but if there's a better way to achieve this please share.

For context, XGBoost does re-coding in CUDA/C++ and needs to extract the data from cuDF python

[mroeschke]

Sure thing. The results from the pylibcudf.Column.data()/null_mask() functions expose a __cuda_array_interface__ attribute that you can use when composing jnames below. e.g.

In [1]: import pylibcudf as plc, pyarrow as pa

In [2]: plc_col = plc.Column.from_arrow(pa.array(["a", "b", None]))

# the data of the "values"
In [3]: plc_col.data().__cuda_array_interface__
Out[3]: 
{'shape': (2,),
 'strides': None,
 'typestr': '|u1',
 'data': (126954835543040, False),
 'version': 3}

# the null mask of the "values"
In [5]: plc_col.null_mask().__cuda_array_interface__
Out[5]: 
{'shape': (64,),
 'strides': None,
 'typestr': '|u1',
 'data': (126954835542016, False),
 'version': 3}

# the "offsets"
In [6]: plc_col.children()[0].data().__cuda_array_interface__
Out[6]: 
{'shape': (16,),
 'strides': None,
 'typestr': '|u1',
 'data': (126954835542528, False),
 'version': 3}

You would probably need to perform the string dtype check on the cuDF Python object first since the typestr here doesn't indicate "string" e.g.

if not (cats._column.dtype == np.dtype("object") or isinstance(cats._column.dtype, pd.StringDtype)):
   raise TypeError(
            "Unexpected type for category index. It's neither numeric nor string."
        )

[mroeschke]

Tangential question, I also see DfCatAccessor here indicates that you could be passing around an cudf/pd.Series.cat object. Is that intentional?

[trivialfis]

Yes, it's intentional, mostly to get the cat accessor then pass it to these utilities to extract the C data.

[trivialfis]

The results from the pylibcudf.Column.data()/null_mask() functions expose a cuda_array_interface attribute that you can use when composing jnames below

Thank you for sharing, that's really helpful! I will look into it.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated 2 comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

This script is committing a PR-specific CI container tag (IMAGE_TAG=PR-78). If this merges as-is, downstream CI/users will likely fail once that ephemeral tag is removed or never published. Revert this to the standard tag (e.g., main) before merging, and keep any temporary CI tag override confined to the PR/CI configuration.

Suggested change

	IMAGE_TAG=PR-78
	IMAGE_TAG=main