HF RL Explorer

Wrong number of TPU hosts reported on newer TPU generations

Wrong number of TPU hosts reported on newer TPU generations: a task in MiMo-V2.6-RL-harbor-code: MiMo-V2.6-RL Code (Harbor) (Harbor dataset). I'm running Ray on a multi-host TPU pod on GCP. When I let Ray autodetect the pod and ask it how many worker hosts there are in the current pod, I get the…

The task

I'm running Ray on a multi-host TPU pod on GCP. When I let Ray autodetect the pod and ask it how many worker hosts there are in the current pod, I get the wrong number on newer TPU generations (e.g. v5e / v5p / v6e). The same code path works correctly for v4.

Part of FineEnvs/MiMo-V2.6-RL-harbor-code.