December 2025: Focused delivery for the HuggingFace Transformers project, with a targeted feature in GraniteMoeHybrid to improve rotary embeddings handling and positional embedding configuration. The work enhances modularity, safety, and configurability, and aligns with transformers v5 compatibility. Key changes include conditional application of rotary embeddings, introduction of a position_embedding_type config, and comprehensive code cleanup and minor fixes to improve maintainability and reliability in production deployments.

June 2025 monthly summary for liguodongiot/transformers. Delivered Granite Speech reliability and performance improvements enabling stable training with the Hugging Face trainer, including updated query handling during training, removal of unused parameters, padding-related crash prevention, and improved mel-spectrogram initialization. Model-level performance enhancements and robust save-state behavior implemented, with adapters to improve efficiency and optimized positional attention. Key commits: be10d4df60bec044ac0c1ab6fd326479874baafc and 22b0a898787f9e34c2b9b4ac1e53d2497c44ff39.