Software stacks¶

Analysis software at Purdue AF can come from several sources, listed here from most to least recommended:

1. Pixi environments — the recommended way to manage analysis software.
2. Conda environments — supported; use Pixi for new projects.
3. LCG stacks distributed via CVMFS — useful for special cases such as CUDA-enabled ROOT.
4. Apptainer/Singularity images via CVMFS — a last resort for software that requires a different operating system.

Pixi environments¶

Pixi is a modern package manager and a successor of Conda/Mamba. It is significantly faster than Conda, resolves Conda and PyPI dependencies together, and enforces reproducibility via lock files.

Unlike Conda environments, Pixi environments are meant to be project-specific: the environment definition (pixi.toml) and the environment itself live in the project directory, next to your analysis code. A detailed guide on how to start using Pixi is available here: Pixi guide.

The global Pixi environment¶

In addition to project-specific environments, we provide a global Pixi environment at /work/pixi/global/, which contains all common HEP analysis packages and ML libraries. It is a good starting point for new projects and for code that is not part of any Pixi project.

[workspace]
authors = ["Dmitry Kondratyev"]
channels = ["conda-forge"]
name = "pixi-global"
platforms = ["linux-64"]
version = "0.1.0"

[dependencies]
awkward = "*"
awkward-pandas = "*"
bokeh = "3.10.0.*"
# boost = "*"
cabinetry = "*"
ca-certificates = "*"
certifi = "*"
click = "*"
cms-combine = "==11.0.0"
coffea = "2026.9.0.*"
correctionlib="*"
dask = "*"
dask-gateway = "*"
dask-histogram = "*"
dask-jobqueue = "*"
dask-memusage = "*"
dask-ml = "*"
dask-xgboost = "*"
distributed = "*"
eigen = "*"
fsspec-xrootd = "*"
gcc_linux-64 = "*"
graphviz = "*"
gpustat = "*"
gsl = "*"
gxx_linux-64 = "*"
h5py = "<3.16.1"
hepdata-lib = "*"
hist = "*"
ipykernel = "*"
ipympl = "*"
ipywidgets = "*"
jupyter = "*"
jupyter-sshd-proxy = "0.3.0.*"
law = "*"
lhapdf = "*"
libgcc-devel_linux-64 = "*"
libgcc-ng = "*"
libstdcxx-devel_linux-64 = "*"
libstdcxx-ng = "*"
lmfit = "*"
# madminer = "*"
mamba = "*"
matplotlib = "*"
mimesis = "*"
# floor: an unconstrained solve can pick mlflow 1.x, whose protos fail under protobuf >= 4
mlflow = ">=3"
mplhep = ">=1.1.2"
numba = "*"
numpy = "*"
omegaconf = "*"
openssl = "*"
openssh = "*"
optuna = "*"
pandas = "*"
particle = "*"
pcre2 = "*"
pip = "*"
plotly = "*"
py-spy = "*"
pyarrow = "*"
pycurl = "*"
pydot = "*"
pyhf = "*"
pyjuliacall = "*"
pytest = "*"
python = "3.12.*"
python-graphviz = "*"
pytorch = "2.12.*"
rucio-clients = "*"
root = ">=6.40.2,<6.41"
scikit-learn = "*"
scikit-optimize = "*"
scipy = "*"
seaborn = "*"
setuptools = "*"
sysroot_linux-64 = "*"
tbb = "*"
tensorflow = {version = "==2.19.1", build = "cuda*"}
tensorflow-probability = "*"
tf-keras = "*"
torch-geometric = "*"
tqdm = "*"
uncertainties = "*"
uproot = "*"
vdt = "*"
vector = "*"
wrapt = "<2.4.2"
xgboost = "*"
xrootd = "*"
yaml = "*"
zuko = "*"

[system-requirements]
cuda = "12.4"

[pypi-dependencies]
cmsstyle = "*"
dbs3-client = "*"
phyprakit = "*"
rabbit-fit = { version = "*", extras = ["plotting"] }
scalpl = "*"
servicex = "==3.3.1"
tensor2tensor = "*"
termplotlib = "*"
xgbfir = "*"
roastcoffea = "*"
pyroscope-io = "*"

If a package that you consider common is missing from the global environment, let us know — we update it regularly.

Jupyter kernels¶

We provide multiple types of Jupyter kernels to execute analysis code in notebooks.

Pixi kernels¶

There is no one-to-one mapping between Pixi environments and Jupyter kernels. Instead, we provide two special Pixi kernels:

• Python (pixi global) — always uses the global environment at /work/pixi/global/.
• Python (pixi project-aware) — automatically discovers the environment local to the directory where the notebook is located. If no local environment is found, the kernel falls back to the global environment.

Conda kernels¶

Conda environments are discovered automatically and appear as kernels if they have the ipykernel package installed and are stored in a publicly readable directory — see Creating Conda environments and Jupyter kernels.

ROOT C++ kernel¶

This kernel provides an interactive interface to the ROOT command line, allowing you to execute ROOT macros and produce plots inside Jupyter notebooks.

LCG kernels¶

These kernels are based on LCG "views" loaded via CVMFS:

• LCG_106b — the standard LCG software stack;
• LCG_106b_cuda — contains the CUDA-enabled ROOT build and is suitable for running RooFit on GPUs.

Combine¶

Combine is included in the global Pixi environment (package cms-combine) — see Using Combine at Purdue AF.

CMSSW¶

CMSSW releases are available via CVMFS in the usual way:

source /cvmfs/cms.cern.ch/cmsset_default.sh
cmsrel CMSSW_13_0_13
cd CMSSW_13_0_13/src
cmsenv

Note that Purdue AF is based on an EL8 system (RHEL8-compatible, el8/slc8 architectures), so CMSSW releases built for other architectures (e.g. slc7) must be run inside an Apptainer container such as cmssw-el7 (see below, and the MC generation guide for a worked example).

Apptainer / Singularity images¶

In rare cases when you need to run code that requires a specific operating system, you can load Apptainer/Singularity images via CVMFS.

Example of loading an Apptainer image based on EL7:

$ /cvmfs/cms.cern.ch/common/cmssw-el7
Singularity>