OpenAI OSS

[photomz]

Today OpenAI released 2 OSS models that are SOTA: 20B, 120B MoE.

https://github.com/openai/gpt-oss

vLLM already supports inference. This would be a game changer for RL training. What's Verl's roadmap to supporting this?

[plmsmile]

+1

[qiaoyu-zheng]

+1

[phoenixbai]

+1

[aongwachi1]

+1

[Tonyzhou98]

+1

[zhuokaizhao]

+1

[vermouth1992]

Flash-attention support is still lacking. Once it's ready, we can start to integrate

[bturtel]

+1

[paulwilczewski]

+1

[tedspare]

Flash-attention support is still lacking. Once it's ready, we can start to integrate

Dao-AILab/flash-attention#1797

[SeungyounShin]

+1

[sosofun]

+1

[cbartmann]

+1