Every other model parallelism API I've seen has partitioning annotations in-lined in the model code and uses custom training logic — making it challenging to modify and maintain the partitioning config. Keras 3 dramatically simplifies this.

→ View original post on X — @fchollet