Release 2.6.5
View on GitHub ↗2026-11-16
Summary
- Exact grid symmetries, crops, and pads process vector fields directly in their input dtype. #676 · D4
- In the single-threaded M4 Max test, a batch of 12 finite FP16 CPU gradients of shape
2×1200×1200took 24.89 ms for D4 and collation, down from 341.47 ms. The measurement covers in-memory CPU preparation. Benchmark and workload - CPU Tensor fallbacks use Albucore 0.3.10 so PyTorch can detect overlapping storage when a result shares its input buffer. #674
Breaking changes
Exact field geometry preserves partially non-finite vectors component by component and keeps the original length of normals. If a model requires unit normals or all-NaN sentinels for invalid vectors, prepare the fields at load time. Interpolation excludes a vector from sampling when any component is non-finite. Vector-field contract
Bug fixes
- PyTorch now detects overlapping writes when a CPU Tensor fallback returns a view into the original tensor. Use
result.clone()as the copy source when the view overlaps the destination. #674 - Dense 2D field warps reuse their sampled buffers and validity masks. Dense 3D displacements reuse packed pull maps across inverse iterations. #676