New logging keys for LLM training benchmark

[sgpyc]

Hi, we would like to add a few new logging keys for the LLM training benchmark:

No.	key	what is it	expected value
1	opt_adam_weight_decay	ADAM weight decay	0.1
2	opt_learning_rate_decay_schedule	learning rate decay schedule	cosine with linear warmup
3	opt_init_checkpoint_step	the step ckpt is loaded for, in the lr schedule	4000 * 1536 / batch_size
4	opt_adam_beta_1	ADAM beta1	0.9
5	opt_adam_beta_2	ADAM beta2	0.995
6	opt_adam_epsilon	ADAM epsilon	0.1
7	sequence_length	input sequence length	2048
8	trained_samples	number of train samples used to reach the target accuracy

Please advise on how to proceed or whether a full list of keys used by the LLM benchmark is required.
Thanks a lot.

[pgmpablo157321]

@sgpyc Should there be a new pair of .yaml files for the LLM benchmark? In this case how should they be named? closed_LLM.yaml and open_LLM.yaml?
And is this new model supposed to replace bert?

[ShriyaPalsamudram]

@sgpyc's PR adds the yaml files. The model name is gpt3 and it is a new benchmark that does not replace anything.