textdiffuse

Add Modal training and token cache data pipeline

Improve text diffusion training diagnostics

param count

weighted loss via masked ratios

disable flex for now