#!/usr/bin/env bash
#SBATCH --job-name=lfm2vl          # optional
#SBATCH --nodes=1
#SBATCH --gres=gpu:8               # <-- change to your GPU count
#SBATCH --ntasks-per-node=8
#SBATCH --time=08:00:00

export CUDA_DEVICE_MAX_CONNECTIONS=1  # NCCL optimal
export TOKENIZERS_PARALLELISM=false

NUM_GPU=8       # must match --gres above
MASTER_PORT=29500

MLFLOW_TRACKING_URI="./mlruns/finetune_lfm" MLFLOW_OFFLINE_MODE=true MLFLOW_EXPERIMENT_NAME="lfm_further_hf" torchrun \
    --nproc_per_node=$NUM_GPU \
    --master_port=$MASTER_PORT \
    ./finetune_custom_trainer.py