Give example on how to handle gradient accumulation with cross-entropy #5282

Job	Run time
run-tests (latest, test_prod)	1m 57s
run-tests (latest, test_core)	2m 54s
run-tests (latest, test_cli)	2m 32s
run-tests (latest, test_big_modeling)	2m 30s
run-tests (latest, test_deepspeed)	2m 19s
run-tests (latest, test_fsdp)	2m 23s
run-tests (latest, test_example_differences)	2m 15s
run-tests (latest, test_checkpoint_step)	3m 52s
run-tests (latest, test_checkpoint_epoch)	3m 27s
run-tests (latest, test_rest)	9m 41s
run-tests (minimum, test_prod)	2m 41s
run-tests (minimum, test_core)	3m 55s
run-tests (minimum, test_cli)	3m 56s
run-tests (minimum, test_big_modeling)	3m 24s
run-tests (minimum, test_deepspeed)	3m 19s
run-tests (minimum, test_fsdp)	3m 29s
run-tests (minimum, test_example_differences)	3m 28s
run-tests (minimum, test_checkpoint_step)	4m 51s
run-tests (minimum, test_checkpoint_epoch)	4m 45s
run-tests (minimum, test_rest)	11m 54s
	1h 19m 32s

Provide feedback