CUDA plug-in for Awkward Array, enables GPU-bound arrays and operations.

apache-arrow, cern-root, columnar-format, data-analysis, jagged-array, json, numba, numpy, pandas, python, ragged-array, scikit-hep
pip install awkward-cuda-kernels==1.4.0rc1


PyPI version Conda-Forge Python 3.7‒3.11 BSD-3 Clause License Build Test

Scikit-HEP NSF-1836650 DOI Documentation Gitter

Awkward Array is a library for nested, variable-sized data, including arbitrary-length lists, records, mixed types, and missing data, using NumPy-like idioms.

Arrays are dynamically typed, but operations on them are compiled and fast. Their behavior coincides with NumPy when array dimensions are regular and generalizes when they're not.

Motivating example

Given an array of objects with x, y fields and variable-length nested lists like

array = ak.Array([
    [{"x": 1.1, "y": [1]}, {"x": 2.2, "y": [1, 2]}, {"x": 3.3, "y": [1, 2, 3]}],
    [{"x": 4.4, "y": [1, 2, 3, 4]}, {"x": 5.5, "y": [1, 2, 3, 4, 5]}]

the following slices out the y values, drops the first element from each inner list, and runs NumPy's np.square function on everything that is left:

output = np.square(array["y", ..., 1:])

The result is

    [[], [4], [4, 9]],
    [[4, 9, 16], [4, 9, 16, 25]]

The equivalent using only Python is

output = []
for sublist in array:
    tmp1 = []
    for record in sublist:
        tmp2 = []
        for number in record["y"][1:]:

Not only is the expression using Awkward Arrays more concise, using idioms familiar from NumPy, but it's much faster and uses less memory.

For a similar problem 10 million times larger than the one above (on a single-threaded 2.2 GHz processor),

  • the Awkward Array one-liner takes 4.6 seconds to run and uses 2.1 GB of memory,
  • the equivalent using Python lists and dicts takes 138 seconds to run and uses 22 GB of memory.

Speed and memory factors in the double digits are common because we're replacing Python's dynamically typed, pointer-chasing virtual machine with type-specialized, precompiled routines on contiguous data. (In other words, for the same reasons as NumPy.) Even higher speedups are possible when Awkward Array is paired with Numba.

Our presentation at SciPy 2020 provides a good introduction, showing how to use these arrays in a real analysis.


Awkward Array can be installed from PyPI using pip:

pip install awkward

You will likely get a precompiled binary (wheel), depending on your operating system and Python version. If not, pip attempts to compile from source (which requires a C++ compiler, make, and CMake).

Awkward Array is also available using conda, which always installs a binary:

conda install -c conda-forge awkward

If you have already added conda-forge as a channel, the -c conda-forge is unnecessary. Adding the channel is recommended because it ensures that all of your packages use compatible versions:

conda config --add channels conda-forge
conda update --all

Getting help

How-to tutorials

Python API reference

C++ API reference

Installation for developers

Be sure to clone this repository recursively to get the header-only C++ dependencies.

git clone --recursive

Also be aware that the default branch is named main, not master, which could be important for pull requests from forks.

You can install it on your system with pip, which uses exactly the same procedure as deployment. This is recommended if you do not expect to change the code.

pip install .[test,dev]

Or you can build it locally for incremental development. The following reuses a local directory so that you only recompile what you've changed. This is recommended if you do expect to change the code.

python --pytest tests

The --pytest tests runs the integration tests from the tests directory (drop it to build only).

For more fine-grained testing, we also have tests of the low-level kernels, which can be invoked with

python dev/
python -m pytest -vv -rs tests-spec
python -m pytest -vv -rs tests-cpu-kernels

Release notes and Roadmap

The release notes/history/changelog is generated as part of the documentation.

The roadmap, future version planning, issue prioritization, deprecation schedule, etc. are kept up-to-date on the wiki.

Papers and talks about Awkward Array

Citing Awkward Array in a publication

On the GitHub README, see the "Cite this repository" drop-down menu on the top-right of the page.


Support for this work was provided by NSF cooperative agreement OAC-1836650 (IRIS-HEP), grant OAC-1450377 (DIANA/HEP), PHY-1520942 (US-CMS LHC Ops), and OAC-2103945 (Awkward Array).

We also thank Erez Shinan and the developers of the Lark standalone parser, which is used to parse type strings as type objects.

Thanks especially to the gracious help of Awkward Array contributors (including the original repository).

Jim Pivarski

💻 📖 🚇 🚧

Ianna Osborne


Pratyush Das


Anish Biswas



💻 ⚠️

Henry Schreiner

💻 🚇

Nicholas Smith

💻 ⚠️

Lindsey Gray

💻 ⚠️



Dmitry Kalinkin


Charles Escott


Mason Proffitt


Michael Hedges


Jonas Rembser


Jaydeep Nandi






Doug Davis


Joosep Pata


Martin Durant


Gordon Watts


Nikolai Hartmann


Simon Perkins



💻 ⚠️



Angus Hollands

⚠️ 💻


💻 ⚠️

Bernhard M. Wiedemann


Matthew Feickert


Santam Roy Choudhury


Jeroen Van Goey




Manasvi Goyal


Aryan Roy




💻: code, 📖: documentation, 🚇: infrastructure, 🚧: maintenance, : tests and feedback, 🤔: foundational ideas.