Questions on Nemotron Nano Omni - including async GRPO, multi-node RDMA, R3, and fully async training
- Dominant language
- Python
- Stars
- 2k
- Forks
- 561
- Avg merge
- 4d 5h
- Merged PRs (30d)
- 145
Description
Hi, I have a few questions regarding Nemotron Nano Omni and would like to raise them as an issue.
1. Are there any plans to support **async GRPO for Nano-Omni** (including non-colocated setups and in-flight weight updates)?
2. Related to (1): when I tried async GRPO with Nemo-RL on my own in a non-colocated setting, **cross-node communication of pixel_values was extremely heavy.** Since Ray does not yet seem to support IB RDMA, I ran into significant difficulties scaling to multi-node. Have you encountered the same issue, and are there any plans to address it?
3. My understanding is that **R3 (Routing Replay; https://arxiv.org/abs/2510.11370)**, which is supported in Verl and slime, has not yet been implemented in Nemo-RL. Are there plans to support it?
4. Are there plans to support **fully async training**, including partial rollouts?
Thanks in advance!
Contributor guide
Assessment
This issue has not been assessed yet.