SAHI Files

A patch release fixing out-of-memory failures and severe slowdowns when postprocessing large prediction sets, plus model, packaging and documentation fixes gathered since 0.12.1.

🐛 Fixes

• Postprocessing no longer allocates an N x N matrix for large inputs (#1395). v0.11 used a shapely STRtree and only compared nearby boxes; v0.12 replaced it with a dense matrix that is O(N^2) in time and memory regardless of layout. The TorchVision backend had no size guard and built seven N x N tensors on the GPU, so it ran out of memory first. The greedy loops only ever use matrix >= match_threshold, never the metric values, so the thresholded adjacency is now built directly from the pairs an STRtree reports as intersecting and stored as CSR. Fixes #1374.

• RF-DETR local models can be loaded by class name, with corrected Roboflow docs (#1394).

• MMDetection has_mask now handles RepeatDataset (#1387).
• Dependency and minimum-version checks are enforced in import_utils (#1377).
• All OpenCV distributions are kept on one version, avoiding conflicting installs (#1393).

📚 Documentation

• Chinese translations updated and completed (#1371).
• Docs default to YOLO26 and link the Ultralytics YOLO26 page (#1386), and YOLO26 is listed among the CLI models (#1378).

🧹 Maintenance & CI

• Bumped astral-sh/setup-uv 8.2.0 → 8.3.2 (#1389) and updated actions/checkout / actions/cache (#1385).
• Relaxed the build requirement to >=0.10,<1.6 (#1382).

⚡ Performance

33337 boxes, IOS metric, threshold 0.3. Same outputs before and after (6571 predictions for greedy_nmm, 4639 for nmm).

CPU (Intel Core i7-13850HX):

CUDA (RTX 4000 Ada Laptop, 12 GB):

Inputs below 2000 boxes, and any non-positive threshold, stay on the dense path where it is faster.

Full Changelog: https://github.com/obss/sahi/compare/0.12.1...0.12.2