FAILED_CONTROLLER after a preemption, no error in logs about skypilot

[Michaelvll]

Thanks for reporting this @Hubert-Bonisseur! This is quite weird. Possibly the controller process is somehow killed.

Could you share how many spot jobs you were running concurrently and if you have seen any issue with the other spot jobs?

It would be nice to share the job task yaml you were running as well : )

[Hubert-Bonisseur]

I was running only one job at that time. I since launched 2 concurrent spot jobs and they are working fine so far, but there hasn't been a preemption yet. I will update then.

Here is the task.yml

name: small-yt

resources:
  cloud: gcp
  region: europe-west4
  cpus: 12+
  accelerators: A100
  memory: 6+

  disk_size: 500
  disk_tier: 'medium'

file_mounts:
  ~/secret/service_account.json: /Users/datalab/épellations/STT/finetune/secrets/finetuning-414911-4d293f61509f.json

envs:
  COMMIT: b11a10fb86feef059d4798ff883ea719e4169218
  MODEL_ID: small-yt-V2
  NUM_WORKERS: 12

setup: |
  echo "Begin setup."
  sudo apt-get update
  sudo apt-get -y install ffmpeg
  cd ~/sky_workdir
  git clone [email protected]:data-science/speech/finetune.git
  cd finetune_whisper
  git checkout $COMMIT
  pip install -r requirements.txt
  pip install "git+https://github.com/skypilot-org/skypilot.git#egg=sky-callback&subdirectory=sky/callbacks/"
  mkdir ~/checkpoints/
  if gsutil ls gs://finetuning-checkpoints/$MODEL_ID; then
    gsutil -m cp -r gs://finetuning-checkpoints/$MODEL_ID/* ~/checkpoints/
  else
    echo "Remote folder does not exist. Starting a new training run"
  fi  
  echo "Setup complete."

run: |
  echo "Beginning task."
  cd finetune
  export GOOGLE_APPLICATION_CREDENTIALS=$(realpath ~/secret/service_account.json)
  export PYTHONPATH=$PWD
  python finetune run configs/training_config_mosaicML.yml