Fetching the paper…

DHA: Learning Decoupled-Head Attention from Transformer Checkpoints via Adaptive Heads Fusion · Around