curated-transformers

Why we included this project

Curated Transformers is a PyTorch library that breaks transformers into small, reusable components, so a bugfix or feature like 4/8-bit quantization lands across every model at once, from Llama and Falcon to BERT and RoBERTa. Because the pieces are independent, you can assemble unusual variants, for instance a BERT encoder with rotary embeddings, without forking the whole codebase. Explosion uses it as the default transformer backend for spaCy 3.7, which gives existing spaCy users a tested route to modern LLMs. The minimal dependency footprint and consistent type annotations make it comfortable for both research experiments and production serving.