v0.27.1#

This patch release contains fixes to CI/CD pipelines, documentation builds, notebook execution, and several runtime bugs. Key areas include resolving nightly CI pipeline issues, fixing ~150 Sphinx build warnings, standardizing notebook setup cells across all CI-tracked notebooks, correcting GPU device detection and seed handling in perturbers, and adding slant range validation for TurbulenceAperturePerturber.

Fixes#

  • Fixed README images and links not rendering on PyPI by replacing relative paths with absolute URLs pointing to the public GitHub repository.

  • Fixed ~150 Sphinx documentation build warnings including malformed RST tables, missing toctree entries, broken cross-references, non-consecutive header levels, incorrect notebook cross-reference links, and docstring formatting issues.

  • Fixed scripts/update_release_notes.sh to produce a single commit instead of two, and to correctly update docs/release_notes/index.rst (previously wrote to a nonexistent docs/release_notes.rst file).

  • Fixed scripts/combine_release_notes.sh to use the v prefix in release note titles for consistency, and to concatenate full fragment content instead of only extracting bullet-point lines.

  • Fixed nightly scheduled CI pipeline wiping the coverage badge by removing the schedule exclusion from test and coverage jobs.

  • Fixed nightly scheduled CI builds to run the full MR pipeline (tests, quality, docs, security, containers). Mirror and publish stages remain unaffected.

  • Fixed compliance trigger job running on tag-initiated publish pipelines by scoping its rules to MR events and default branch pushes only.

  • Fixed notebook rendering being out of sync with releases by redesigning the tox papermill environment to build a local wheel and install it with --no-index, so notebook %pip install commands use the local build instead of PyPI. Added nbstripout to strip personal metadata from executed notebooks. Changed base_python to py313.

  • Fixed print_extras_status() showing unknown for package versions by updating _try_import() to use importlib.metadata.version() instead of the module __version__ attribute. Changed the readthedocs URL from a version-pinned path to /en/stable/.

  • Fixed inconsistent notebook setup cells across all 17 CI-tracked notebooks by standardizing %pip install format, correcting per-notebook extras, removing unnecessary numpy<2.0 pins, and consolidating scattered install cells into a single setup cell per notebook.

  • Fixed diffusion notebook hardcoding device="cuda" which caused failures on CPU-only runners. The notebook now uses DiffusionPerturber’s built-in auto-detection (CUDA when available, CPU otherwise).

  • Fixed jatic-perturbations-saliency notebook by replacing JitterPerturber with AverageBlurPerturber.

  • Fixed PDF docs CI job failing by switching latex_engine to lualatex which handles Unicode/emoji characters natively, installing texlive-luatex and fonts-freefont-ttf in CI, using -pdflua latexmk flag, and overriding Sphinx’s default fontpkg to use Latin Modern fonts bundled with texlive.

  • Fixed notebook CI jobs using a shared runner tag, which prevented GPU-dependent notebooks from being tested on CUDA. Each notebook now specifies its own runner tag so that GPU notebooks run on GPU runners.

  • Fixed tox blanket CUDA_VISIBLE_DEVICES override hiding GPUs from the papermill environment. The papermill environment now passes through the host’s CUDA_VISIBLE_DEVICES so that GPU-dependent notebooks can access CUDA when run on a GPU runner.

  • Fixed a bug with TorchRandomPerturbImage._set_seed not selecting the correct device before setting seed.

  • Fixed diffusion tests to match changes from the _set_seed bug fix.

  • Fixed _get_device CUDA availability check to support specific GPU device strings (e.g., "cuda:0", "cuda:1"). Previously only the bare "cuda" string was matched, so requests for a specific GPU would bypass the fallback-to-CPU logic when CUDA was unavailable.

  • Fixed TurbulenceAperturePerturber._create_simulator to validate that slant_range is greater than or equal to altitude.

  • Fixed TestTurbulenceAperturePerturber.test_configuration_bounds to test for slant range less than altitude.

On this page