Optimum + ONNX Runtime Training accelerates Hugging Face models — migrate to ORTTrainer and ORTTrainingArguments

AI Impact Summary

Optimum is integrated with ONNX Runtime Training to accelerate Hugging Face model training via memory and compute optimizations, delivering up to 35%+ improvements in many cases and up to 130% when combined with DeepSpeed ZeRO-1 on PyTorch baselines. The approach supports both NVIDIA and AMD GPUs and introduces ORTTrainer/ORTTrainingArguments to replace the standard Transformer Trainer flow, enabling faster training loops, mixed precision, and memory efficiency. For engineering teams, this implies a targeted migration path: swap Trainer with ORTTrainer and TrainingArguments with ORTTrainingArguments to realize the acceleration, potentially reducing training time and cloud compute costs for large language, vision, and speech models.