transformers

Enable Gradient Accumulation fix across all models + trainer fully in forward()

#34283

Merged