New Release: Ultralytics v8.3.213

Discussion
, ,
1

Ultralytics v8.3.213 — Stability first: automatic NaN recovery, faster Objects365, unified dataset URLs :ambulance::repeat_button::high_voltage:

A stability-focused release that automatically recovers from NaN/Inf training issues, cleans up resume logic, speeds up Objects365 setup, and unifies dataset download URLs. Upgrade with confidence and keep training moving.

  • Version: v8.3.213
  • Focus: Training robustness, dataset reliability, and setup speed
  • Release: See the full notes on the v8.3.213 release page

TL;DR

  • Auto-recovery from NaN/Inf losses and metric collapse during training (DDP-aware, capped retries) :shield:
  • Safer resumes with centralized checkpoint loading and scheduler reset
  • Objects365 preparation runs significantly faster with multithreading :rocket:
  • Unified ASSETS_URL across dataset YAMLs for more reliable downloads :link:

Highlights

Training robustness and resume improvements

  • Automatically detects NaN/Inf losses and fitness collapse, then safely restores from the latest checkpoint with capped retries (up to 3).
  • DDP-aware broadcasting keeps distributed training in sync during recovery.
  • Validates checkpoint weights to avoid reloading corrupted EMA states.
  • Centralizes checkpoint loading via _load_checkpoint_state() and uses it in resume_training() for consistency and less drift.
  • Resets the scheduler state after recovery to maintain the intended LR schedule.
  • Includes a new test test_nan_recovery that injects a NaN to verify the recovery path. :white_check_mark:
    Details are in the NaN epoch recovery PR (#22352) by Glenn Jocher.

Dataset YAMLs: unified asset URLs

Objects365 setup speedups

CI and tests

  • Temporarily disabled Jetson JetPack 5 Docker build while the new recovery path stabilizes.
  • Skips training tests on Jetson/Raspberry Pi to reflect that edge devices are not training targets.

Improvements

  • More resilient training at scale with automatic recovery from NaN/Inf losses or sudden metric collapse.
  • Safer, more consistent resume logic for optimizer, scaler, EMA, and best-fitness states.
  • Faster dataset preparation, especially for Objects365, thanks to multithreading.
  • More reliable dataset downloads via centralized hosting and unified URLs.

Bug Fixes

  • Prevents reloading corrupted EMA states by validating checkpoint weights during recovery.
  • Avoids scheduler desynchronization by resetting LR scheduler after recovery.

Quick start

  • Upgrade:
    pip install -U ultralytics
  • Train (recovery is automatic; no extra flags needed):
    yolo detect train model=yolo11n.pt data=coco128.yaml epochs=100 imgsz=640
  • Resume anytime:
    yolo detect train resume=True

Learn more about training options in the Train mode documentation.

Model guidance

  • YOLO11 is our latest stable and recommended default for all tasks; explore the YOLO11 model docs.
  • Community models YOLO12 and YOLO13 are not recommended; see notes in the YOLO12 model docs.
  • Ultralytics R&D for YOLO26 is underway; follow the YOLO26 R&D preview.

What’s changed (PRs and authors)

Review the complete changelog for v8.3.212 → v8.3.213.

Try it and share feedback

Please upgrade, put the new recovery path through its paces, and let us know how it performs in your workflows. You can start a conversation in Ultralytics GitHub Discussions or open issues with reproducible examples. Your feedback helps the YOLO community and Ultralytics team keep improving.