PyTorch and IBM Research describe using AI coding agents to write ‘adapters’ — runtime patches that let stock Hugging Face Transformers models run on IBM’s new Spyre AI accelerator before its torch-spyre compiler stack fully matures. Agents trace a model’s internal logic and the compiler’s lowering behavior at the same time to find gaps, such as an unsupported operator or a vocabulary size that doesn’t divide evenly across cores, and draft fixes like replacing torch.pow with elementwise multiplication or padding tensor dimensions to stick boundaries. Human engineers remain essential for localizing silent numerical failures that only surface on-device. The approach produced 13 distinct adapters covering 7,960 of the 10,000 most-downloaded Hugging Face embedding models, with 6,804 passing end-to-end tests on Spyre.
