I have a training job running as a pathways workload on the GKE pathways cluster (through xpk). To monitor the metrics of my training job, I use wandb for scalar metrics logging. WandB by default saves the metrics log files to local file system. In a pathways workload case, it's the docker image.
However, the issue with the flow I described is that if my job fails or is preempted in the middle, the metrics file of wandb will be lost.
Resuming of my training job from checkpoint won't get back the metrics history.
We need xpk workload/clusters to be able to mount the gcs bucket to the head node so that we can specify the mounted directory to wandb logging.
I have a training job running as a pathways workload on the GKE pathways cluster (through xpk). To monitor the metrics of my training job, I use wandb for scalar metrics logging. WandB by default saves the metrics log files to local file system. In a pathways workload case, it's the docker image.
However, the issue with the flow I described is that if my job fails or is preempted in the middle, the metrics file of wandb will be lost.
Resuming of my training job from checkpoint won't get back the metrics history.
We need xpk workload/clusters to be able to mount the gcs bucket to the head node so that we can specify the mounted directory to wandb logging.