[CI] RAPIDS XGBoost packages block 26.12 conda environments #8583

Summary

Python conda jobs fail during environment creation because strict channel priority selects RAPIDS’ xgboost 3.3.0 packages, whose librmm constraints are incompatible with libnvforest 26.12.

Failing test/component: Conda test environment creation

Failure observed in:

• https://github.com/NVIDIA/cuml/actions/runs/34335597461
• https://github.com/NVIDIA/cuml/actions/runs/34343304452

Environment

• OS: Rocky Linux 8; Ubuntu 22.04, 24.04, and 26.04
• Python version: 3.12, 3.13, and 3.14
• CUDA version: 12.2.2, 12.9.2, 13.0.3, and 13.3.0
• GPU: L4, H100, and RTX PRO 6000
• Dependencies: latest-deps; nightly 26.12 conda packages; strict channel priority

Test Details

• Test file: N/A (environment creation fails before tests run)
• Test name: conda-python-* and conda-notebook-tests
• Error message:

  error    libmamba Could not solve for environment specs
  
  libcuml 26.12 requires libnvforest 26.12
  libnvforest 26.12 requires librmm 26.12.*
  
  xgboost 3.3.0 requires libxgboost 3.3.0
  libxgboost 3.3.0 requires librmm >=26.8,<26.9
  or librmm >=26.10,<26.11
  

Root Cause Analysis

cuML requests xgboost>=3.3.0. With rapidsai-nightly and rapidsai ahead of conda-forge under strict channel priority, conda selects RAPIDS’ XGBoost packages.

The available RAPIDS libxgboost 3.3.0 builds require an older RMM release:

• stable builds require librmm >=26.8,<26.9
• nightly builds require librmm >=26.10,<26.11

The same environment requires libnvforest 26.12, which requires librmm 26.12.*. No librmm version satisfies both constraints.

The published CUDA constraints for the 26.12 libnvforest and librmm packages are correct. The CUDA conflicts shown by libmamba are consequences of exploring incompatible package variants.

Removing or relabeling RAPIDS’ XGBoost packages would allow strict channel priority to resolve standard XGBoost from conda-forge. That migration is tracked in:

rapidsai/build-planning#312