Databricks (who also participated in OLMo, it's probably the same codebase) trained on AMD before, see 2023 post https://www.databricks.com/blog/amd-mi250. It was probably seamless, as any issues were fixed by Databricks in 2023.