> the pro run restarted at step 17 due to a GPU OOM issue caused by expert load imbalance. we have adjusted the training parallelism strategy.

no idea why I laughed so hard at this, feel so familiar