Release Notes


This is a patch release from v0.6.2v0.6.3.

Important Notes

  • With the addition of writing ROOT files in uproot v4.1.0 the xmlio extra no longer requires uproot3 and all dependencies on uproot3 and uproot3-methods have been dropped. (PR #1567) uproot4 additionally brings large speedups to writing, which results in an order of magnitude faster conversion time for most workspace conversions from JSON back to XML + ROOT with pyhf json2xml.

  • All backends are now fully compatible and tested with Python 3.9. (PR #1574)

  • The TensorFlow backend now supports compatibility with TensorFlow v2.2.1 and later and TensorFlow Probability v0.10.1 and later. (PR #1001)

  • The with_aux keyword arg has been renamed to include_auxdata to improve API consistency. (PR #1562)


  • The weakref bug with Click v8.0+ was resolved. pyhf is now fully compatible with Click v7 and v8 releases. (PR #1530)


Python API

  • Model parameter names are now propagated to optimizers through addition of the pyhf.pdf._ModelConfig.par_names() API. pyhf.pdf._ModelConfig.par_names() also handles non-scalar modifiers with 1 parameter. (PRs #1536, #1560)

    >>> import pyhf
    >>> model = pyhf.simplemodels.uncorrelated_background(
    ...     signal=[12.0, 11.0], bkg=[50.0, 52.0], bkg_uncertainty=[3.0, 7.0]
    ... )
    >>> model.config.parameters
    ['mu', 'uncorr_bkguncrt']
    >>> model.config.npars
    >>> model.config.par_names()
    ['mu', 'uncorr_bkguncrt[0]', 'uncorr_bkguncrt[1]']
  • The pyhf.pdf._ModelConfig channel_nbins dict is now sorted by keys to match the order of the channels list. (PR #1546)

  • The with_aux keyword arg has been renamed to include_auxdata to improve API consistency. (PR #1562)


This is a patch release from v0.6.1v0.6.2.

Important Notes

  • The pyhf.simplemodels.hepdata_like() API has been deprecated in favor of pyhf.simplemodels.uncorrelated_background(). The pyhf.simplemodels.hepdata_like() API will be removed in pyhf v0.7.0. (PR #1438)

  • There is a small breaking API change for pyhf.contrib.viz.brazil.plot_results(). See the Python API changes section for more information.

  • The pyhf.patchset.PatchSet schema now allows string types for patch values in patchsets. (PR #1488)

  • Only lower bounds on core dependencies are now set. This allows for greater developer freedom and reduces the risk of breaking user’s applications by unnecessarily constraining libraries. This also means that users will be responsible for ensuring that their installed dependencies do not conflict with or break pyhf. c.f. Hynek Schlawack’s blog post Semantic Versioning Will Not Save You for more in-depth coverage on this topic. For most users nothing should change. This mainly affects developers of other libraries in which pyhf is a dependency. (PR #1382)

  • Calling dir() on any pyhf module or trying to tab complete an API will now provide a more helpfully restricted view of the available APIs. This should help provide better exploration of the pyhf API. (PR #1403)

  • Docker images of releases are now published to both Docker Hub and to the GitHub Container Registry. (PR #1444)

  • CUDA enabled Docker images are now available for release v0.6.1 and later on Docker Hub and the GitHub Container Registry. Visit for more information.


  • Allow for precision to be properly set for the tensorlib ones and zeros methods through a dtype argument. This allows for precision to be properly set through the pyhf.set_backend() precision argument. (PR #1369)

  • The default precision for all backends is now 64b. (PR #1400)

  • Add check to ensure that POIs are not fixed during a fit. (PR #1409)

  • Parameter name strings are now normalized to remove trailing spaces. (PR #1436)

  • The logging level is now not automatically set in pyhf.contrib.utils. (PR #1460)


Python API


  • The CLI API now supports a patchset inspect API to list the individual patches in a PatchSet. (PR #1412)

pyhf patchset inspect [OPTIONS] [PATCHSET]


v0.6.2 benefited from contributions from:

  • Alexander Held


This is a patch release from v0.6.0v0.6.1.

Important Notes

  • As a result of changes to the default behavior of torch.distributions in PyTorch v1.8.0, accommodating changes have been made in the underlying implementations for pyhf.tensor.pytorch_backend.pytorch_backend(). These changes require a new lower bound of torch v1.8.0 for use of the PyTorch backend.


  • In the PyTorch backend the validate_args kwarg is used with torch.distributions to ensure a continuous approximation of the Poisson distribution in torch v1.8.0+.


Python API

  • The solver_options kwarg can be passed to the pyhf.optimize.opt_scipy.scipy_optimizer() optimizer for additional configuration of the minimization. See scipy.optimize.show_options() for additional options of optimization solvers.

  • The torch API is now used to provide the implementations of the ravel, tile, and outer tensorlib methods for the PyTorch backend.


This is a minor release from v0.5.4v0.6.0.

Important Notes

  • Please note this release has API breaking changes and carefully read these notes while updating your code to the v0.6.0 API. Perhaps most relevant is the changes to the pyhf.infer.hypotest() API, which now uses a calctype argument to differentiate between using an asymptotic calculator or a toy calculator, and a test_stat kwarg to specify which test statistic the calculator should use, with 'qtilde', corresponding to pyhf.infer.test_statistics.qmu_tilde(), now the default option. It also relies more heavily on using kwargs to pass options through to the optimizer.

  • Following the recommendations of NEP 29 — Recommend Python and NumPy version support as a community policy standard pyhf v0.6.0 drops support for Python 3.6. PEP 494 – Python 3.6 Release Schedule also notes that Python 3.6 will be end of life in December 2021, so pyhf is moving forward with a minimum required runtime of Python 3.7.

  • Support for the discovery test statistic, \(q_{0}\), has now been added through the pyhf.infer.test_statistics.q0() API.

  • Support for pseudoexperiments (toys) has been added through the pyhf.infer.calculators.ToyCalculator() API. Please see the corresponding example notebook for more detailed exploration of the API.

  • The minuit extra, python -m pip install pyhf[minuit], now uses and requires the iminuit v2.X release series and API. Note that iminuit v2.X can result in slight differences in minimization results from iminuit v1.X.

  • The documentation will now be versioned with releases on ReadTheDocs. Please use to access the documentation for the latest stable release of pyhf.

  • pyhf is transtioning away from Stack Overflow to GitHub Discussions for resolving user questions not covered in the documentation. Please check the GitHub Discussions page to search for discussions addressing your questions and to open up a new discussion if your question is not covered.

  • pyhf has published a paper in the Journal of Open Source Software. JOSS DOI Please make sure to include the paper reference in all citations of pyhf, as documented in the Use and Citations section of the documentation.


  • Fix bug where all extras triggered warning for installation of the contrib extra.

  • float-like values are used in division for pyhf.writexml().

  • Model.spec now supports building new models from existing models.

  • \(p\)-values are now reported based on their quantiles, instead of interpolating test statistics and converting to \(p\)-values.

  • Namespace collisions between uproot3 and uproot/uproot4 have been fixed for the xmlio extra.

  • The normsys modifier now uses the pyhf.interpolators.code4 interpolation method by default.

  • The histosys modifier now uses the pyhf.interpolators.code4p interpolation method by default.


Python API

  • The tensorlib API now supports a tensorlib.to_numpy and tensorlib.ravel API.

  • The pyhf.infer.calculators.ToyCalculator() API has been added to support pseudoexperiments (toys).

  • The empirical test statistic distribution API has been added to help support the ToyCalculator API.

  • Add a tolerance kwarg to the optimizer API to set a float value as a tolerance for termination of the fit.

  • The pyhf.optimize.opt_minuit.minuit_optimizer() optimizer now can return correlations of the fitted parameters through use of the return_correlation Boolean kwarg.

  • Add the pyhf.utils.citation API to get a str of the preferred BibTeX entry for citation of the version of pyhf installed. See the example for the CLI API for more information.

  • The pyhf.infer.hypotest() API now uses a calctype argument to differentiate between using an asymptotic calculator or a toy calculator, and a test_stat kwarg to specify which test statistic to use. It also relies more heavily on using kwargs to pass options through to the optimizer.

  • The default test_stat kwarg for pyhf.infer.hypotest() and the calculator APIs is 'qtilde', which corresponds to the alternative test statistic pyhf.infer.test_statistics.qmu_tilde().

  • The return type of \(p\)-value like functions is now a 0-dimensional tensor (with shape ()) instead of a float. This is required to support end-to-end automatic differentiation in future releases.


  • The CLI API now supports a --citation or --cite option to print the preferred BibTeX entry for citation of the version of pyhf installed.

$ pyhf --citation
  author = {Lukas Heinrich and Matthew Feickert and Giordon Stark},
  title = "{pyhf: v0.6.0}",
  version = {0.6.0},
  doi = {10.5281/zenodo.1169739},
  url = {},
  note = {}

  doi = {10.21105/joss.02823},
  url = {},
  year = {2021},
  publisher = {The Open Journal},
  volume = {6},
  number = {58},
  pages = {2823},
  author = {Lukas Heinrich and Matthew Feickert and Giordon Stark and Kyle Cranmer},
  title = {pyhf: pure-Python implementation of HistFactory statistical models},
  journal = {Journal of Open Source Software}


v0.6.0 benefited from contributions from:

  • Alexander Held

  • Marco Gorelli

  • Pradyumna Rahul K

  • Eric Schanet

  • Henry Schreiner


This is a patch release from v0.5.3v0.5.4.


  • Require uproot3 instead of uproot v3.X releases to avoid conflicts when uproot4 is installed in an environment with uproot v3.X installed and namespace conflicts with uproot-methods. Adoption of uproot3 in v0.5.4 will ensure v0.5.4 works far into the future if XML and ROOT I/O through uproot is required.


Without the v0.5.4 patch release there is a regression in using uproot v3.X and uproot4 in the same environment (which was swiftly identified and patched by the fantastic uproot team)

$ python -m pip install "pyhf[xmlio]<0.5.4"
$ python -m pip list | grep "pyhf\|uproot"
pyhf           0.5.3
uproot         3.13.1
uproot-methods 0.8.0
$ python -m pip install uproot4
$ python -m pip list | grep "pyhf\|uproot"
pyhf           0.5.3
uproot         4.0.0
uproot-methods 0.8.0
uproot4        4.0.0

this is resolved in v0.5.4 with the requirement of uproot3

$ python -m pip install "pyhf[xmlio]>=0.5.4"
$ python -m pip list | grep "pyhf\|uproot"
pyhf            0.5.4
uproot3         3.14.1
uproot3-methods 0.10.0
$ python -m pip install uproot4 # or uproot
$ python -m pip list | grep "pyhf\|uproot"
pyhf            0.5.4
uproot          4.0.0
uproot3         3.14.1
uproot3-methods 0.10.0
uproot4         4.0.0


This is a patch release from v0.5.2v0.5.3.


  • Workspaces are now immutable

  • ShapeFactor support added to XML reading and writing

  • An error is raised if a fit initialization parameter is outside of its bounds (preventing hypotest with POI outside of bounds)


Python API

  • Inverting hypothesis tests to get upper limits now has an API with pyhf.infer.intervals.upperlimit

  • Building workspaces from a model and data added with


  • Added CLI API for pyhf fit

  • pyhf combine now allows for merging channels: pyhf combine --merge-channels --join <join option>

  • Added utility to download archived pyhf pallets (workspaces + patchsets) to contrib module: pyhf contrib download


v0.5.3 benefited from contributions from:

  • Karthikeyan Singaravelan