RoPE scaling config confusing

[tattrongvu]

Hi Yarn team,
thank you guys for the awesome work. Currently I'm trying to evaluate several rope scaling methods and fortunately there are all available in this git. I have some question related to the Config of rope scaling.
I see that in the requirements.txt you already include transformers >= 4.34.0, so it mean I could use the "linear" and "dynamic-ntk" out of the box with transformers, just by add the rope scaling in AutoConfig.from_pretrained() like that:
config.rope_scaling = { "type": "linear", "factor": args.linear }
or
config.rope_scaling = { "type": "dynamic", "factor": args.dynamic_ntk }
I tried that and remove the patch for linear & dynamic-ntk and the result look identical when using your implemented patch.
Moreover, it also support Falcon architecture. (https://github.com/huggingface/transformers/blob/main/src/transformers/models/falcon/modeling_falcon.py#L162)
So my question is that, are there any different between this two implementation? or your implementation for linear & dynamic-ntk patch is for keeping the reproduction eval consistent?