You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
In the training scripts, we often ignore the validation values and instead do checkpoint averaging + periodic evaluation that's scheduled externally and runs in trtllm. If that is the case, the validation that nemo does is just wasting compute, but it's not possible to disable it currently. The best way is to fix it on nemo/nemo-aligner side so that we can just disable it. But if not, let's at least set the minimum possible validation batch size by default, which is something like mbs * dp size (might be a bit different if more involved parallelism is used)
The text was updated successfully, but these errors were encountered:
In the training scripts, we often ignore the validation values and instead do checkpoint averaging + periodic evaluation that's scheduled externally and runs in trtllm. If that is the case, the validation that nemo does is just wasting compute, but it's not possible to disable it currently. The best way is to fix it on nemo/nemo-aligner side so that we can just disable it. But if not, let's at least set the minimum possible validation batch size by default, which is something like mbs * dp size (might be a bit different if more involved parallelism is used)
The text was updated successfully, but these errors were encountered: