v2.7.2
Internal change
Use sparse_categorical_crossentropy for test as the loss object default does not work on tpustrategy + the single task trainer already handles the reduction.
Fix a bug that `loop_fns.create_tf_while_loop_fn` doesn't handle nested structure states.
Remove internal link in Orbit README.
Don't early return if the global step is not updated by `num_steps` as expected. This still allows summary and metric reporting.
#KerasNLP Update TransformerEncoderBlock to support Q, KV as two input streams.