some modifications 1. upsample input size from 224*224px to 384*384px 2. interpolate positional embedding from 14*14 to 24*24

Make dataset and collator for supervised fine-tuning.

x: b * c * h * w Assuming x contains num_split**2 sub-squares concatenated along batch dimension, merge the sub-squares back to the o

Collects the state dict and dump to disk.

x: b * c * h * w Deividing x into num_split**2 sub-squares, and concatenate all the sub-squares on the batch dimension

Split a list of indices into `chunks` chunks of roughly equal lengths.

Print the LoRA parameters (r and alpha) for all layers in the model.

Phi3RMSNorm is equivalent to T5LayerNorm