Release notes

What changed in each RayD release, newest first, from the project changelog.

RayD versions the rayd, rayd-drjit, and rayd-torch distributions together. Releases are tagged on GitHub and published to PyPI.

0.8.0 #

Released 2026-07-30.

True mesh instancing: add_instance shares the source mesh's acceleration structure and gives each instance its own transform, shape id, and scene-global primitive ids. Instance transforms are differentiable and update without a rebuild. See Update geometry in place.
PyTorch: high-level reflection accumulation (Scene.accumulate_reflections) and EPC (Scene.trace_refl_epc, trace_refl_epc_field) APIs.
PyTorch: replicated multi-GPU execution (Scene(devices=[...])) now also covers diffraction paths, coherent diffraction, and reflection accumulation.
PyTorch: batched SDF grids (SdfGridBatch, MixedScene.add_sdf_batch).

AD support is now reported per operation and per input domain, for VJP and JVP separately, in api_manifest()["derivatives"], replacing the backend-wide claim. See Differentiation.
PyTorch trace_refl_epc_field is explicitly forward-only until its Fresnel and polarization derivatives are implemented.
The rayd-torch wheel requires the PyTorch minor version it was built for: torch>=2.10,<2.11.
Broader single- and multi-GPU CI, Stable ABI checks, and release-artifact verification.

Restored the rayd.drjit._C type-stub layout.
PyTorch reflection material defaults now produce a nonzero physical field on a validated reflected path.

Released 2026-07-23.

Automatic pure-CUDA ray-tracing fallback when OptiX is unavailable, with CUDA-fused visibility, reflection, diffraction, and multipath paths. See Backends.
Native RF layer-stack, transmission, scattering, penetration, and diffraction path families for C++ consumers, including source-lane diffraction path layouts (DfrPathLayout.SourceLane).
A validated source bundle for downstream CMake builds and a LibTorch Stable ABI slice tested with Torch 2.10 through 2.13.

Wheels for Python 3.10 through 3.14 on Windows and manylinux_2_28, built with CUDA 12.8.
Prebuilt binaries contain native code for sm_70, sm_75, sm_80, sm_86, sm_87, sm_89, sm_90, sm_100, sm_101, and sm_120, plus compute_120 PTX.
The typed Torch integration API is stable; the legacy integration boundary is retired.

Finite-edge UTD continuation and gating near stationary-point and boundary transitions.
Races in CUDA eager queries and BVH memory, launch, and build robustness.
Consistency of reflection EPC, diffraction, and geometry adjoints between the Dr.Jit and Torch backends.

Released 2026-07-10.

Explicit backend namespaces rayd.drjit and rayd.torch.
Separately installable rayd-drjit and rayd-torch distributions, plus the rayd meta-distribution that pins both to the same version.
A Torch-native C++/CUDA/OptiX backend with autograd, forward AD, torch.compile, geometry, visibility, reflection, diffraction, and EPC operations.
Shared native integration headers for source-linked CMake builds.

rayd is a PEP 420 namespace and no longer picks or re-exports a default backend. Import rayd.drjit or rayd.torch explicitly.
The two backends share geometry and multipath contracts while keeping their own arrays, ownership, streams, and AD graphs.

Finite-edge coherent UTD keeps its endpoint continuation instead of cutting the field to zero when the stationary point lies outside the edge segment.
More robust OptiX pipeline creation, scene cache updates, accumulation, EPC bindings, and AD propagation in both backends.