v0.6.6

Bugfix: added alternate URL for libtorch aarch64 download.

v0.6.5

This release brings ir_mul layout support for OpenEquivariance. Pass the parameter layout='ir_mul' to any TPProblem instance to use a transposed layout for the input and output irreps. To transpose input and output irreps, use oeq.transpose_irreps or oeq.jax.transpose_irreps; see our API page for usage details.

v0.6.4

Bugfix: added missing MLIR lowerings for a pair of JAX primitives (thanks @teddykoker!)

v0.6.3

Versioning fix.

v0.6.2

Minor fixes to license and README.md

v0.6.1

Minor updates to address symlinks in README.md and LICENSE.

v0.6.0

OpenEquivariance v0.6.0 brings long-needed improvements to the PyTorch frontend. We strongly encourage all users to upgrade to PyTorch 2.10 and OEQ v0.6.0.

Added:

• OpenEquivariance triggers a build of the CUDA extension module at pip install time and will use this precompiled extension if the user has PyTorch >=2.10 installed. If PyTorch <2.10 is installed, the JIT-compiled extension is used instead.
• PyTorch ABI support for C++ backend, using new features in PyTorch 2.10 to support stable, forward-compatible ahead-of-time extensions.
• Dropped support for TorchBind classes and a new kernel cache in its place, which greatly improves flexibility for automatic mixed precision and AOTI compilation. An inference test in C++ is included.
• openequivariance_extjax has a version number that synchronizes with the main openequivariance package; ensure the two packages stay in sync.

Fixed:

• torch.to() is now called when either TensorProduct or TensorProductConv is a submodule of another PyTorch module.

v0.5.4

v0.5.4

Improvements to JAX frontend.

Added to openequivariance.jax:

• Jacobian Vector Products (JVP) for both TensorProduct and TensorProductConv via custom primitives, in addition to VJP.
• Arbitrary higher-order derivatives. In particular, support for Phonon Fine Tuning in Nequix.
• jax.jit support.
• Initial support for vmap / pmap applied to fused convolution (to-do for unfused tensor products).

Fixed:

• Zero all output buffers in the backwards and double-backwards implementations of convolution before calling kernels.

v0.5.3

Minor bugfix, explicit version for JAX extension.

v0.5.2

Minor symlink fix for JAX package.