feat: Remove support for PyTorch backend (#2611)

* Remove support for PyTorch from the codebase.
   - This is a breaking change.
   - Remove the PyTorch backend and optimize/opt_pytorch.py.
* Remove use of PyTorch from the tests.
   - Remove PyTorch lower bounds from tests/constraints.txt.
   - Remove comparison of tensor library backends in
     tests/test_backend_consistency.py as there is only NumPy and JAX now.
* Remove all docs and examples that mention or use PyTorch.
* Remove mentions of PyTorch in all parts of the website and REAMDE except
  for historical purposes like the release notes, JOSS paper, and notes
  about early pyhf research in FAQs.
   - For abstract use term "automatic differentiation".
* Update NumPy backend docstrings to accommodate NumPy v2.0 returns.

Co-authored-by: Giordon Stark <[email protected]>

Author: matthewfeickert

.zenodo.json

@@ -28,7 +28,6 @@
       "fitting",
       "scipy",
       "numpy",
-      "pytorch",
       "jax",
       "auto-differentiation"
     ],

CITATION.cff

@@ -26,7 +26,6 @@ keywords:
   - fitting
   - scipy
   - numpy
-  - pytorch
   - jax
   - auto-differentiation
 license: "Apache-2.0"
@@ -37,8 +36,8 @@ abstract: |
   of that statistical model for multi-bin histogram-based analysis and its
   interval estimation is based on the asymptotic formulas of "Asymptotic
   formulae for likelihood-based tests of new physics". pyhf supports modern
-  computational graph libraries such as PyTorch and JAX in order
-  to make use of features such as autodifferentiation and GPU acceleration.
+  computational graph libraries such as JAX in order to make use of features
+  such as automatic differentiation and GPU acceleration.
 references:
   - type: article
     authors:

README.rst

@@ -28,9 +28,8 @@ multi-bin histogram-based analysis and its interval estimation is based
 on the asymptotic formulas of “Asymptotic formulae for likelihood-based
 tests of new physics”
 [`arXiv:1007.1727 <https://arxiv.org/abs/1007.1727>`__]. The aim is also
-to support modern computational graph libraries such as PyTorch and
-JAX in order to make use of features such as autodifferentiation
-and GPU acceleration.
+to support modern computational graph libraries such as JAX in order to
+make use of features such as automatic differentiation and GPU acceleration.
 
 ..
   Comment: JupyterLite segment goes here in docs
@@ -144,7 +143,6 @@ Implemented variations:
 
 Computational Backends:
   - ☑ NumPy
-  - ☑ PyTorch
   - ☑ JAX
 
 Optimizers:

docker/gpu/install_backend.sh

@@ -18,10 +18,7 @@ function get_JAXLIB_GPU_WHEEL {
 function install_backend() {
   # 1: the backend option name in setup.py
   local backend="${1}"
-  if [[ "${backend}" == "torch" ]]; then
-    # shellcheck disable=SC2102
-    python3 -m pip install --no-cache-dir .[xmlio,torch]
-  elif [[ "${backend}" == "jax" ]]; then
+  if [[ "${backend}" == "jax" ]]; then
     python3 -m pip install --no-cache-dir .[xmlio]
     python3 -m pip install --no-cache-dir "$(get_JAXLIB_GPU_WHEEL)"
     python3 -m pip install --no-cache-dir jax

docs/api.rst

@@ -65,7 +65,6 @@ The computational backends that :code:`pyhf` provides interfacing for the vector
    :nosignatures:
 
    numpy_backend.numpy_backend
-   pytorch_backend.pytorch_backend
    jax_backend.jax_backend
 
 Optimizers

docs/conf.py

@@ -146,7 +146,6 @@ def setup(app):
 # today_fmt = '%B %d, %Y'
 
 autodoc_mock_imports = [
-    'torch',
     'jax',
     'iminuit',
 ]
@@ -195,13 +194,11 @@ def setup(app):
     'examples/notebooks/StatError.ipynb',
     'examples/notebooks/histogrammar.ipynb',
     'examples/notebooks/histosys.ipynb',
-    'examples/notebooks/histosys-pytorch.ipynb',
     'examples/notebooks/importxml.ipynb',
     'examples/notebooks/multichannel-coupled-normsys.ipynb',
     'examples/notebooks/multichannel-normsys.ipynb',
     'examples/notebooks/normsys.ipynb',
     'examples/notebooks/pullplot.ipynb',
-    'examples/notebooks/pytorch_tests_onoff.ipynb',
 ]
 
 # The reST default role (used for this markup: `text`) to use for all