New Release: Ultralytics v8.4.197

Discussion
, ,
1

:rocket: Ultralytics v8.4.197: Offline AMP Checks and Leaner Dependencies

Ultralytics YOLO v8.4.197 is here! This release removes weight downloads from Automatic Mixed Precision (AMP) checks and replaces PyYAML with a built-in YAML utility—making offline training smoother and standard installations leaner.

:high_voltage: More Independent AMP Checks

AMP validation now builds an untrained YOLO26n from its YAML configuration, rather than downloading a checkpoint.

  • No pretrained weights or internet access required: Avoid unexpected downloads and skipped checks when training offline.
  • Layer-level precision checks: Compare FP32 and AMP results across convolution and attention layers to help identify precision issues before training.
  • Less dependence on training setup: Checks use the model’s device without resetting callbacks.

See PR #26967 by @glenn-jocher for the implementation details.

:package: Built-In YAML Support

Ultralytics now includes a minimal YAML reader and writer, removing PyYAML from the base dependencies.

  • Supports YAML structures commonly used in Ultralytics configurations, datasets, and export metadata.
  • Provides consistent loading and saving through ultralytics.utils.YAML.
  • Updates documentation examples to use the built-in utility when saving YAML files.

Compatibility note: The built-in parser covers common Ultralytics YAML needs; it is not intended to support every YAML feature.

Explore PR #26983 by @glenn-jocher for details on the replacement.

:hammer_and_wrench: Try It Out

Upgrade to the new release:

pip install --upgrade ultralytics==8.4.197

You can review the v8.4.197 release notes for the overview or browse the full changelog for all changes.

:speech_balloon: Give it a try and share your feedback below! We’d especially love to hear how offline AMP checks and the built-in YAML utility work with your training workflows.