Skip to content

IBM Analog Hardware Accelerator Kit 1.2.0

Latest

Choose a tag to compare

@PabloCarmona PabloCarmona released this 21 Oct 10:58
· 10 commits to master since this release

[1.2.0] - 2026/10/21

Added

  • Define data attribution for AnalogContext with read-only protection (#765)
  • Non-zero gamma support in ChoppedTransferCompound (#764)
  • Add a section to the README linking to aihwkit-lightning (#804)
  • Add cibuildwheel configuration to pyproject.toml for building wheels for Python 3.12–3.14 on Linux x86_64 and macOS arm64 (#820)
  • Add examples/run_examples.py to run the numbered examples and write a report, and examples/compare_runs.py to compare the outputs of two runs (#782)
  • Add a section to the README for running on AMD GPUs with ROCm/HIP (#789)

Changed

  • Updated and corrected analog_device_nonidealities_tutorial notebook (#758)
  • Remove the reset bias clamp in rpu_pulsed_device.h so that a negative reset bias is kept (#755) (#768)
  • Update the install instructions in the notebooks and docs to use the GPU-enabled 1.1.0 wheel (#759) (#760)
  • Update CUDA.Dockerfile to CUDA 13.3.0 and change the default CUDA_ARCH from 86 to 75
  • Replace the per-Python-version wheel workflows with a single build-wheels workflow driven by the cibuildwheel configuration
  • Bump version to 1.2.0, make the package version dynamic and update the license year in the license file and source headers (#818)

Fixed

  • Build against torch >= 2.14, whose headers require C++20: the C++ standard is now detected from the installed torch headers (C++17 for torch <= 2.13, C++20 for torch >= 2.14) and can be overridden with the new RPU_CXX_STANDARD cmake option (#792)
  • Preserve device runtime state (update counters, chopper states, transfer slice indices, pulse counters) when calling set_hidden_parameters on CUDA tiles (#794)
  • Support CUDA 13.x / CCCL 3.x toolchains: add compatibility aliases for the CUB iterators removed in CCCL 3.0, forward the MKL/OpenBLAS include path to nvcc, silence the GCC 15 -Wtemplate-body error in PyTorch headers, and fix an int/float mismatch in a CUB reduction call (#783)
  • Build the CUDA native extension for Blackwell GPUs by adding 120 to the default RPU_CUDA_ARCHITECTURES (#805)
  • Wrap partial in staticmethod in the test decorators to fix tile parameters on Python 3.14 (#817)
  • Return the plot figure instead of w_trace in utils/visualization.py, which had broken example 21 since #702 (#819)
  • Add POST_BUILD explicitly to custom commands in CMake to remove the warning for policy CMP0175 (#822)
  • Use SPDX license expression and license-files in pyproject.toml (requires setuptools >= 77.0.0) (#823)
  • Keep the input clamp gradient at the bound for torch >= 2.14, fixing the input range gradient in the analog MVM (#827)
  • Use the restored ChoppedTransfer choppers in the CUDA weight reduction (#821)
  • Improve Python executable detection and error handling in CMake (#757)
  • Retune PWU kernel when m_batch grows after initial m=1 update (#763)
  • Tests for dynamic tolerance for cuDNN TF32 precision on Ampere+ GPUs (#771)
  • Correct self-move assignments in RPU device move operators, removing a -Wself-move warning on GCC 13+ builds (#793)
  • Support PyTorch *_like factory operations in AnalogContext placeholder mode without reading backing weight values, with CPU and CUDA regression coverage (#785)
  • Fix double bracket indexing in HermesNoiseModel and export it from aihwkit.inference (#786)
  • Pass the y-limits of plot_device_symmetry as separate arguments to plt.ylim (#773)
  • Remove has_names from the AnalogContext placeholder metadata, deprecated in torch >= 2.13
  • Add cmake and mypy as build dependencies in pyproject.toml to fix pip install ., and fix the Read the Docs build

Security

  • Bump jupyter-server from 2.18.2 to 2.20.0 (#777)
  • Fix code scanning alerts for multiplication results converted to a larger type in the C++ simulator and tile periphery (#797) (#798) (#802)
  • Add explicit permissions to GitHub workflows to fix code scanning alerts (#796) (#799) (#800) (#801)
  • Bump urllib3 from 2.7.0 to 2.8.0 (#808)
  • Bump pillow from 12.2.0 to 12.3.0 (#784)
  • Update uv.lock to solve dependency vulnerabilities, including accelerate (#791)
  • Bump several dependencies to solve security vulnerabilities (#816)