Why it matters
Init determines training stability. Understanding shapes design.
Advertisement
The architecture
Xavier: sqrt(2 / (fan_in + fan_out)).
Kaiming: sqrt(2 / fan_in) for ReLU.
Modern LLM: small std (0.02) for many.
Advertisement
How it works end to end
Preserve variance forward + backward.
LLM: often std = 0.02 empirically.
muP: rescale further.
Special for residual: down-scale output projection.