Optimum 1.2 enables ONNX Runtime-backed inference for Hugging Face Transformers pipelines

AI Impact Summary

Optimum now enables inference for Hugging Face Transformers pipelines via ONNX Runtime, allowing accelerated, hardware-optimized transformer inference across NLP, CV, and speech workloads. The integration introduces API-compatible ORTModelForXxx variants and tools like ORTOptimizer and ORTQuantizer to achieve graph-level optimizations and dynamic quantization, delivering lower latency and smaller model footprints. The example pathways reference RoBERTa-base-squad2 and seamless Hub integration for optimized checkpoints, with support for accelerators such as Graphcore IPU and Habana Gaudi. This creates a production-ready path to scale transformer workloads without rewriting pipelines, though teams should verify installation of optimum[onnxruntime] and performance against their specific workloads.