Ultralytics improves INT8 on macOS
Ultralytics released v8.4.31 with better INT8 export for non‑square images, improved Apple Silicon stability, and multi‑scale auto‑batching — small but useful wins for deploying vision models on Apple hardware. These updates ease model export and execution across varied image sizes on M‑series Macs. (x.com)
PR #24028 corrects INT8 calibration when imgsz is non‑square (examples in the PR notes reference imgsz=640,480), fixing calibration errors that previously produced incorrect quantized export results. (github.com)) A dedicated Apple Silicon stability change (PR #24038) implements more aggressive memory clearing on MPS devices to reduce leak‑related out‑of‑memory (OOM) failures observed during train and validation runs. (github.com)) Auto‑batch logic was updated in PR #24051 so that AutoBatch now includes multi_scale in its effective image‑size calculation, causing automatic batch‑size estimates to reflect larger runtime footprints when multi_scale is enabled. (github.com)) The v8.4.31 commits are dated Mar 26–27, 2026 in the GitHub compare view, and the ultralytics 8.4.31 package is listed on PyPI as the published release. (github.com)) Documentation additions landed alongside these fixes, including a COCO‑to‑YOLO conversion guide (PR #23930) and a YOLO26 training recipe guide (PR #23949) to clarify dataset and training workflows. (github.com)) Because PR #24028 remedies INT8 calibration for non‑square imgsz and PR #24038 targets MPS memory leaks, teams exporting INT8 models to CoreML/ONNX for M‑series Macs can reasonably expect fewer calibration failures and train/validation OOMs during export and tuning (inference based on the cited PR changes). (github.com))
