·
3 commits
to main
since this release
Immutable
release. Only release title and notes can be modified.
Notable Changes
- 🎯 Major fine-tuning improvements. Train with more dataset formats and sources, combine multiple datasets, tune preferences with DPO, focus training on assistant responses, and export portable PEFT LoRA adapters. Fine-tuning is also more reproducible and cache-efficient, with safer GPU access and immutable model snapshots. Existing Alpaca, full-sequence loss, and GGUF behavior remain the defaults. #814 #815 #827 #830 #831 #832 #833 #834 #835 #837 #838
- 🟠 AMD ROCm and Strix Halo support. Build hardware-accelerated inference images for supported AMD GPUs. Thanks @keithmattix for this contribution. #771
- ⚡ Faster builds and smaller images. Project benchmarks show large BuildKit graphs improving by up to 82% and 1,000-file packaging operations by 76–89%. Measured llama.cpp runner images are also 29–38% smaller. #805 #819
- 🚀 Preload models at startup. The new opt-in
loadToMemorysetting can warm selected LocalAI models when a container starts. #836
What's changed?
Features
- Add AMD ROCm and Strix Halo inference support. #771
- Expand fine-tuning dataset formats, loaders, and multi-dataset support. #827 #830 #831 #832 #833 #834
- Add DPO preference tuning. #835
- Load configured inference models at startup. #836
- Export portable PEFT LoRA adapter bundles. #838
Bug Fixes
- Restore compatibility with the current Unsloth and TRL stack. #814
- Regenerate cached model configurations when switching inference backends. #818
- Harden remote dataset loading, validation, timeouts, and cache handling. #837
Performance Improvements
- Optimize BuildKit graphs, local contexts, and model packaging. #805
- Make Unsloth builds reproducible and cache-aware. #815
- Reduce published inference runner image sizes. #819
Chores
New Contributors
- @keithmattix made their first contribution in #771
Full Changelog: v0.21.0...v0.22.1
Contributors
keithmattix