embed tokens is over computed when count flops

[onehaitao]

System Info

when count flops, embed token should not be added to the whole model flops because nn.Embedding is a table lookup operation rather than a computational operation.

For one token, emd_and_lm_head_N should be vocab_size * hidden_size * 1 * 6 rather than vocab_size * hidden_size * 2 * 6.

I think all the calculations involved in this operation need to be adjusted, and the current results are all inflated.

https://github.com/ByteDance-Seed/VeOmni/blob/main/veomni/utils/count_flops.py#L735

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

NA

Expected behavior

correct flops count

[Luosuu]

would you like to make a PR?