A deep stack of transformer blocks would be brittle if each block replaced its input outright. Which associate-level tricks keep that stack trainable?
Select an answer to reveal the explanation.
Short Explanation
A deep stack of blocks would snap if each one threw away the input and started over. Residual skip paths plus layer norm let the signal hop around a block and stay trainable. NCCL rings and tree depth are not that glue, and ripping out the skips makes the stack more brittle.
Full Explanation
Residual paths and layer normalization are the associate-level glue that keeps deep transformer stacks trainable. They are not NCCL topology or tree depth, and deleting skip paths makes the stack more brittle.