v0.6.0

What’s Changed

• Add support for stopping criteria: MaxLength, MaxTime, and Interruptable for more flexible generation control.
• Add PSR-3 logging support
• Extend PretrainedConfig to reduce code repetition across model files and improve maintainability.
• Restructure AutoModel resolution to prioritize generic model selection when no task-specific class is found.
• Add support for new model families: Gemma, Gemma2, Gemma3, Qwen3, and Phi3, with their respective causal language models.
• Add support for eos and last_token pooling strategies in the FeatureExtractionPipeline.
• Add support for new merge format in BPEModel with better compatibility via JSON-encoded token pair mapping.
• Introduce PretrainedModel::$sessions array to streamline model session handling across subclasses.
• Simplify streamer implementation to improve clarity and flexibility.
• Enhance image processing methods and VIPS integration.
• Convert to a platform package with native shared library support for Linux, macOS, and Windows (x86_64 and ARM64).
• Introduce dynamic shared library loading with platform-specific path resolution logic.
• Update example configurations, docs, and usage instructions to reflect new architecture.
• Add tests for image utilities and inference session logic.

Bug Fixes

• Fix Tensor::slice() error during generation.
• Fix logic in RepetitionPenaltyLogitsProcessor to properly apply penalties based on generated tokens.
• Fix compatibility issues with PHP 8.1 and adjust dependency versions accordingly.
• Fix bug in HubTest for Windows paths by using DIRECTORY_SEPARATOR.
• Fix edge case in BPEModel when null value encountered in merge map.
• Fix various doc inconsistencies and default constructor bugs in model subclasses.

Improvements

• Refactor library architecture for improved platform compatibility and modularity.
• Refactor Samplerate, Sndfile, and other FFI wrappers to use instance-based invocation.
• Improve generation configuration merging logic and model resolution flow.
• Unify code style across components and improve clarity of core components.
• Update shared native dependencies:
  • onnxruntime → 1.21.0
  • rindowmatlib → 1.1.1
• Clean up feature extractors and consolidate model session logic.
• Improve error handling in FFI calls and utility classes.
• Rename internal config and model mapping classes for better clarity (PretrainedMixin → AutoModelBase, etc.)

New Contributors

• @panariga made their first contribution in #73
• @Deltachaos made their first contribution in #77

Full Changelog: 0.5.3...0.6.0