About the NeMo Framework restructuring

[lucacorbucci]

Hi all 👋🏻 ,
We are working on the pre- and post-training of a Speech LLM. We'd like to use Nemo for training, but we noticed you're doing some restructuring.
In the Readme of this repo, we found the "Pivot Notice" saying: "The following collections are no longer available: avlm · diffusion · llm · multimodal · multimodal-autoregressive · nlp · speechlm · vision · vlm".

We are wondering if we should use this repository for pre-training or if there is a more specific one.
For post-training, we discovered https://github.com/NVIDIA-NeMo/RL, which is probably more suitable. Is it correct? I've seen that now it only supports VLMs and LLMs. Is that correct? Will it support Speech LLMs in the future?

Last question: We also found Megatron-Bridge. If we start using NeMo, would it be easy to switch to Megatron-Bridge later if we need to add other data modalities?

Thanks!