scope: perf
time_limit: 3600
key_segments:
  # Modify keys to be renamed (str) or excluded (False) from run identifier. By default, all args under script_args are included.
  data_path: False
  val_check_interval: False
  lr: False
script_args:
  # All arguments referenced in the script string must be specified here.
  # Arguments not referenced in the script string must have the 'arg' field specified.
  # See jet/core/configs.py for the specification of the configuration class
  workspace: /workspace/bionemo2
  data_path: /data/cellxgene_scdl
  model: geneformer
  variant: train
  config_name: geneformer_config
  precision: [bf16-mixed]
  gpus:  8
  max_steps: 1000
  lr: 0.001
  val_check_interval: 500
  acc_grad: 1
  products:
    - nodes: 1
      batch_size: 32
    - nodes: 2
      batch_size: 32

script: |-
   WANDB_API_KEY=$BIONEMO_WANDB_API_KEY ${variant}_${model} \
    --data-dir ${data_path} \
    --experiment-name ${batch_size}bs_${nodes}node_${gpus}gpu_${max_steps}s_${precision}prec \
    --num-gpus ${gpus} \
    --save-last-checkpoint \
    --num-nodes ${nodes} \
    --val-check-interval ${val_check_interval} \
    --num-dataset-workers 8 \
    --num-steps ${max_steps} \
    --seq-length 2048 \
    --limit-val-batches 8 \
    --micro-batch-size ${batch_size} \
    --resume-if-exists \
    --log-every-n-steps 50 \
    --lr ${lr} \
    --create-tensorboard-logger \
    --result-dir=${tensorboard_dir} \
    --wandb-group=${model}_${variant}_${config_name}__${target} \
    --wandb-project ${wandb_project_name} \
    --wandb-job-type=${pipeline_label} \
    --cosine-rampup-frac 0.004331629559040111 \
    --cosine-hold-frac 0.021658147795200554 \
    --accumulate-grad-batches ${acc_grad} \
    --precision ${precision} \
    --disable-checkpointing;