BeamZ v0.5.2

v0.5.2 - 2026-10-01

Added

• Added multi-GPU CUDA execution and improved streamed CUDA/JAX scaling,
  continuation, and shard-local material preparation for large 3D simulations.
• Added optional constant-radius bent-waveguide eigenmode evaluation through
  solve_grid(..., bend_radius=..., bend_axis=...).
• Added software citation metadata and preparation for Zenodo release archiving.

Changed

• Improved CUDA storage layout selection and CPML execution, and reduced
  full-field communication in sparse JAX source and monitor operations.
• Removed the experimental cuda_hopper backend; use cuda_streamed or JAX.
  Rebuild the optional CUDA extension for ABI 21 / component version 0.21.0.
• Updated dependencies, validation tooling, benchmarks, and documentation.

Fixed

• Fixed package imports on Windows by making Unix process-memory tracing optional.
• Improved material-coefficient equivalence across mesh, GDS, and native Design
  imports, including mesh-region ownership, clipping, adaptive sampling, and
  boundary classification.
• Hardened CUDA graph execution, modal DFT continuation, and distributed CPML
  ownership and halo exchange.

Validation limits

• GPU performance depends on hardware and geometry. Large-capacity benchmark
  completion does not establish optical convergence or full-size backend parity.
• Some CUDA CPML auxiliary-state and refined raw-DFT comparisons remain outside
  tolerance; see the linked validation reports and benchmark evidence in the
  release notes. Bent-mode radiation loss requires mesh/domain/PML convergence.

Evidence and details

• CUDA/JAX scaling and capacity
• Historical scaling evidence archive
• CUDA layout validation and limitations
• Mesh/GDS/Design raster equivalence
• Bent-waveguide mode validation

Full changelog: v0.5.1...v0.5.2