GPT-OSS has high logprob errors on GB200
Open
bug
- Dominant language
- Python
- Stars
- 2k
- Forks
- 561
- Avg merge
- 4d 5h
- Merged PRs (30d)
- 145
Description
On GB200, GPT-OSS 20B's logprob errors are much higher than H100, and reward is lower. This needs to be investigated. In the plots below, green is H100 and pink/purple are GB200
https://wandb.ai/nvidia/nemo-rl?nw=kacgft88ys
Contributor guide
Assessment
This issue has not been assessed yet.