v2.1.0

[2.1.0] - 2026-09-15

Added

• AnalogConv3d layer. torch.nn.Conv3d modules are now converted by
  convert_to_analog and covered by the correctness tests against AIHWKIT
  (#39).
• Runtime dependencies (torch, typing_extensions, tqdm) and the optional
  extras dev, huggingface and triton are declared in pyproject.toml,
  together with a PEP 517 build backend (#28).
• PyPI classifiers in the project metadata (#66).
• uv project files, lockfiles and updated READMEs for the
  basic_huggingface, deepspeed_and_huggingface and deepspeed_cifar10
  examples.
• examples/aihwkit_evaluation: example that converts a Lightning model to
  AIHWKIT and evaluates it there, with a model summary and an evaluation
  report. Referenced from the user guide (#25).

Changed

• AnalogOptimizer now takes a callable that returns the analog layers
  (e.g. model.analog_layers) instead of an already-created generator, so the
  post-step hook can iterate the layers on every optimizer step. Update calls
  from AnalogOptimizer(SGD, model.analog_layers(), ...) to
  AnalogOptimizer(SGD, model.analog_layers, ...).
• Minimum supported Python version raised from 3.8 to 3.10.
• torch pinned to 2.13.0. PyTorch 2.14 introduced C++ header changes that
  break the build (#68).
• transformers requirement raised to >=5.10.0. The HuggingFace examples
  were adapted to the v5 API, where warmup_ratio became warmup_steps
  (#77).
• Maintainers updated to Manuel Le Gallo-Bourdeau and Pablo Carmona Gonzalez
  (#80).
• Source file headers now reference the MIT license, matching LICENSE.txt
  and the package metadata. They previously still cited Apache 2.0 (#79).
• Read the Docs build moved to an Ubuntu LTS image and a Python version that
  matches the minimum requirement (#81).
• CI workflow runs on Python 3.14 (#21), pins the Triton commit and
  setuptools version, pins torch, torchvision and torchaudio for GCC
  compatibility (#68), and uses actions/checkout v7 (#64) and
  actions/setup-python v7 (#65).
• Type hints in the RNN module were tightened (AnalogRNNCell alias and
  explicit casts) to satisfy the newer mypy (#53).
• Development tooling bumps via Renovate: black 26.5.1, mypy 1.20.2,
  pycodestyle 2.14.0, pytest 9.x, Sphinx 9, sphinx-rtd-theme 3,
  myst-parser 5 and recommonmark 0.7.1.

Fixed

• WeightModifierType is re-exported from aihwkit_lightning.simulator.configs
  again. It was dropped from that module in 2.0.1, which broke imports in
  code written for 1.x.
• AnalogOptimizer post-step hook silently stopped clipping weights after the
  first step because the generator of analog layers was exhausted.
• Type hints in the convolution modules.

Security

• Refreshed uv.lock for the package and the examples to resolve Dependabot
  and Renovate vulnerability alerts (transformers, black, pytest,
  gitpython, urllib3, pillow, setuptools, msgpack, accelerate)
  (#42, #43, #54, #55, #67, #72, #73, #74, #75, #76, #77, #82, #83).