Update installation instructions for cuML with BERTopic.

[csadorf]

What does this PR do?

Updates cuML installation instructions across the documentation to use simplified PyPI.org installation (available as of cuML 25.10) with current CUDA versions.

Changes

Updated Installation Instructions

Updated cuML + BERTopic installation instructions in three documentation files:

• docs/getting_started/clustering/clustering.md
• docs/getting_started/dim_reduction/dim_reduction.md
• docs/getting_started/tips_and_tricks/tips_and_tricks.md

Key Updates

1. Simplified PyPI.org installation: Uses direct pip install from PyPI.org instead of requiring nvidia.com package indexes with --extra-index-url flags
2. Updated CUDA versions: Replaced outdated CUDA 11 packages with current CUDA 12 and CUDA 13 options
3. Streamlined dependencies: Removed auxiliary packages (cudf, dask-cudf, cugraph, cupy) that are not required for basic cuML usage with BERTopic or automatically installed as dependencies
4. Recommended installation order: Install cuML first, then BERTopic to avoid pip resolver dependency conflicts with CUDA runtime packages
5. Added helpful context:
   • Warning about pip resolver limitations
   • Note that cuML is pre-installed on Google Colab
   • Link to RAPIDS installation guide for detailed/advanced installation instructions

Motivation

The updated instructions use pip install from PyPI.org, which is now possible as of cuML release version 25.10. This simplifies the installation process by removing the need for additional package indexes (--extra-index-url) and auxiliary packages (cudf, dask-cudf, cugraph, cupy).

Additionally, the two-step installation approach (install cuML first, then BERTopic) avoids pip resolver limitations when handling conflicting CUDA runtime dependencies between cuML and PyTorch (via sentence-transformers).

Before submitting

• This PR fixes a typo or improves the docs (if yes, ignore all other checks!).
• Did you read the contributor guideline?
• Was this discussed/approved via a Github issue? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes (if applicable)?
• Did you write any new necessary tests?

[csadorf]

See rapidsai/cuml#7374 for details on why we can't install cuml and bertopic with pip in one command. If we get that resolved, I can update the instructions here or we can leave them as-is.

[MaartenGr]

Thank you for the PR and extensive descriptions, that helps me understand the context here a bit more.

I'm also wondering that considering many are moving away from pip to uv, would it make sense to add instructions for uvand potentially fix the dependency conflicts by either pinning indices in pyproject.toml (which seems a lot of maintenance) or potentially add the simple instruction:

uv pip install torch --torch-backend=auto

What do you think?

[csadorf]

Thank you for the PR and extensive descriptions, that helps me understand the context here a bit more.

I'm also wondering that considering many are moving away from pip to uv, would it make sense to add instructions for uvand potentially fix the dependency conflicts by either pinning indices in pyproject.toml (which seems a lot of maintenance) or potentially add the simple instruction:

uv pip install torch --torch-backend=auto

What do you think?

We don't currently support uv installation of cuml. That is not to say it won't work. I'd rather stick to the currently documented approach and then update it if we find that uv installation works reliably or even better, RAPIDS officially supports it.

[csadorf]

@MaartenGr Any chance we could merge this soon or do you think we need to make more changes? Also happy to address those in a follow-up.

[MaartenGr]

@csadorf Apologies for the delay! Yes, sounds good to me.

[csadorf]

@csadorf Apologies for the delay! Yes, sounds good to me.

Thanks a lot!