Changelog¶
All notable changes to MindNLP are documented here.
Version 0.6.x (Current)¶
MindSpore: >=2.7.1 | Python: 3.10-3.11
Highlights¶
- Full HuggingFace Transformers compatibility via patching mechanism
- Full HuggingFace Diffusers compatibility
- Support for latest model architectures (Qwen3, Llama3, etc.)
- Enhanced mindtorch layer for PyTorch API compatibility
- Improved device management and heterogeneous computing support
New Features¶
- Automatic patching of
transformersanddiffuserslibraries - Support for
ms_dtypeparameter in model loading - Enhanced
device_mapsupport for multi-device inference - Improved tensor serialization and checkpoint handling
Version 0.5.x¶
MindSpore: 2.5.0-2.7.0 | Python: 3.10-3.11
Highlights¶
- Major API refactoring for better HuggingFace compatibility
- Introduction of mindtorch compatibility layer
- Support for new model families (Gemma, Phi-3, etc.)
New Features¶
mindnlp.coremodule providing PyTorch-compatible APIs- Enhanced AutoModel classes for various tasks
- Improved tokenizer support
- PEFT/LoRA integration for parameter-efficient fine-tuning
Version 0.4.x¶
MindSpore: 2.2.x-2.5.0 | Python: 3.9-3.11
Highlights¶
- Expanded model support
- Improved training stability
- Enhanced Trainer API
New Features¶
- Support for Qwen2, Mistral, Mixtral models
- Enhanced gradient checkpointing
- Improved distributed training support
- Better memory management for large models
Version 0.3.x¶
MindSpore: 2.1.0-2.3.1 | Python: 3.8-3.9
Highlights¶
- Stable release with comprehensive model coverage
- Improved documentation and examples
New Features¶
- Support for Llama, Llama2 models
- ChatGLM series support (ChatGLM, ChatGLM2, ChatGLM3)
- Enhanced dataset loading utilities
- Improved model serialization
Version 0.2.x¶
MindSpore: >=2.1.0 | Python: 3.8-3.9
Highlights¶
- Major architecture improvements
- Better alignment with HuggingFace APIs
New Features¶
- Refactored model architecture
- Improved tokenizer implementations
- Enhanced training engine
- Better error messages and debugging
Version 0.1.x¶
MindSpore: 1.8.1-2.0.0 | Python: 3.7.5-3.9
Highlights¶
- Initial release of MindNLP
- Core transformer model support
New Features¶
- Basic transformer models (BERT, GPT-2, T5, etc.)
- Tokenizer support
- Dataset loading utilities
- Basic training loop implementation
For detailed release notes, see GitHub Releases.