environment.yml

name: festim-workshop
channels:
  - conda-forge
  - defaults

dependencies:
  - setuptools
  - wheel<0.45
  - jupyter-book<2.0
  # Must be an EXACT pin, not a `<` constraint: repo2docker (Binder / the CI image build)
  # only picks up the Python version when it is pinned exactly, otherwise it builds its base
  # environment on its own default Python and then has to rewrite the whole stack on top.
  # 3.12 because autoemulate is not compatible with Python 3.13 yet.
  - python=3.12
  - pip
  - jupytext # necessary to open MyST files using BinderHub https://jupyterbook.org/en/stable/interactive/launchbuttons.html#launchbuttons-binder
  - sphinx-tags
  - ipywidgets
  - matplotlib
  - meshio[all]
  - ipykernel
  - nbconvert
  - sphinx-codeautolink
  - fenics-dolfinx>=0.11.0  # >=0.11 required for FESTIM 2.2 gas enclosures (real elements)
  - festim==2.2rc2
  - pyvista<0.47.1
  - python-gmsh
  - shapely
  - scipy
  - pyparsing
  - imageio
  - trame
  # No explicit `vtk` entry: it arrives anyway via pyvista, so listing it changed nothing (an
  # image built without it came out byte-identical). Note the `pyvista<0.47.1` pin resolves to
  # pyvista 0.44.1, which requires the full `vtk` metapackage -- qt6-main, pyside6, ffmpeg,
  # OpenVINO, intel-media-driver. pyvista >=0.46 moved to `vtk-base` and >=0.48 is the first
  # that accepts vtk 9.6, so shedding those GBs means relaxing the pyvista pin.
  - trame-vtk<2.10.3
  - trame-vuetify
  - nest-asyncio2
  - pip:
    - h-transport-materials==0.16
    - foam2dolfinx>=1.1
    - openmc2dolfinx>=1.0.2
    # NB: this pulls the CUDA build (nvidia-cublas/cudnn/cusolver/cusparse/cufft/nccl + triton,
    # ~3.3 GB of wheels) even though Binder, Codespaces and RTD are all CPU-only. Swapping it for
    # conda-forge `pytorch-cpu` would shed most of that -- worth revisiting once the workflow's
    # layer-size report shows how close the single conda layer is to GHCR's 10 GB limit.
    - torch<2.12.0
    # Must stay <2.1: autoemulate 2.1 added a dependency on `harmonic`, which cannot be
    # satisfied alongside fenics-dolfinx -- pinning `harmonic>=1.3.1` (the only release with a
    # cp312 manylinux wheel) makes pip report ResolutionImpossible against fenics-dolfinx
    # 0.11.0, and without that pin it backtracks to harmonic 1.3.0, which ships no Linux wheel
    # at all and whose sdist no longer compiles (`model_legacy.pyx:433: Cannot assign type
    # 'double' to 'long'`). pip used to escape this by backtracking all the way past
    # autoemulate 2.1 to 2.0.0 on its own; on a fresh Read the Docs env it stopped doing so and
    # the build failed. Pinning states the version that was being resolved to anyway, and skips
    # a long backtracking search on every env build.
    - autoemulate<2.1