Implement depthwise truncated model init #20
Description
Need to add a proper llama style init, which uses the number of layers to properly init weights (with and without meta_init).
This will ensure weights are properly setup for multi-billion param training.