All notable changes to fairseq2 are documented in this file.
The format is based on Keep a Changelog.
- Introduced
kv_dim
option toStandardMultiheadAttention
(for different sizes of encoder and decoder) - Added the
CheckpointManager.get_model_checkpoint_path
method - Added the
target_twoway mode
in the NLLB tokenizer for formatting target sequences in NLLB or SONAR-like models
- Introduced LLaMA and LLaMA 2
- Introduced Mistral 7B
- Introduced LoRA fine-tuning
- Revised the sequence generator API
- Introduced support for lazy padding and attention masks
- Many smaller improvements to existing APIs
- Improvements to the build system and CI pipelines
- Improvements to the installation instructions and contribution guidelines
- Initial release