🔄 What's New in 0.3.4

📊 lm-evaluation-harness Support - Model evaluation with standard benchmarks (BoolQ, ARC Challenge, etc.) - Documentation
🎯 New RMSNorm Implementation - Precise calculation with ANE hardware ops
🐛 Fixed RoPE Tensor Size Bug - Resolved random overflows (existing pre-0.3.4 models should be re-converted)

####Example ANE vs HF on MPS backend ( Qwen 3 0.6B )

🚀 ANEMLL 0.3.3 Alpha Release

🆕 New Features

• 🎯 Qwen 3 Architecture Support - Initial implementation for Qwen3 (0.6B, 1.7, 4B) models
• 📦 Streamlined Installation - One-command setup with ./create_python39_env.sh +
  ./install_dependencies.sh
• 🧪 Automated Testing Framework - End-to-end validation with python tests/test_qwen_model.py
• 🔧 Enhanced Developer Experience - Improved error handling and better feedback

Sample iOS/macOS inference Chat-Bot App (Alpha)
Updates to Model conversion and upload scripts
Updates to Swift Package and CLI App

• Dependency checks and troubleshooting guide docs/troubleshooting.md added
• Prefill batch size added to the conversion script
• Chat_full interface updated for DeepHermes "think" token "/t" and both chat interfaces added --nw flag to skip warmup step ( docs/chat.md )
• XCode Tools dependency added to the README

Release 0.1.1-alpha

• Single-shot model conversion with see convert_model.md
• Simplified model configuration with meta.yaml
• Automated Hugging Face distribution preparation with prepare_hf.sh
• Enhanced Chat Interfaces with better error handling and configuration support
• Improved LLaMA model with prefill optimization

Initial Support for LLama and Distill DeepSeek for Apple Neural Engine