IBM's AI agents bring day-one support for thousands of models on Spyre

The AI model landscape evolves faster than the software stack that runs it. Even mature compilers struggle with novel modules in new architectures, and new hardware widens that gap dramatically: a young stack must support an entire ecosystem of models at once. Traditionally, closing these gaps takes weeks or months of specialist work, delaying deployment. IBM Research proposes a different approach: use AI agents to write thin runtime patches, called adapters, that let stock models run on a given platform today, without waiting for every underlying compiler gap to close.

Adapters work by intercepting operations that don't lower cleanly through the PyTorch compiler and swapping them for mathematically equivalent forms that do. They preserve the model's computation while handing the compiler something it can handle, leaving performance to the stack. IBM demonstrated this on its Spyre AI accelerator, running stock HuggingFace Transformers models end-to-end. A small number of AI-generated adapters enabled thousands of models, covering the harder cases of novel attention variants, custom numeric ranges, and fused operations. The result is day-one enablement across an entire model ecosystem, even as the underlying platform continues to mature.