Implement and evaluate Deep Delta Learning (DDL) residual blocks in our LLM

[vukrosic]

Implement Deep Delta Learning (DDL) residual connections as an alternative to standard additive residuals in our LLM and evaluate their impact on training stability, convergence speed, and downstream performance.

pdf - https://github.com/yifanzhang-pro/deep-delta-learning/blob/master/Deep_Delta_Learning.pdf