#!/bin/bash
#SBATCH --job-name=stemz
#SBATCH --nodes=8
#SBATCH --gres=gpu:8
#SBATCH --ntasks-per-node=8  # match this to n_gpu if python
#SBATCH --cpus-per-task=4  # n_cpu*n_task has to be <= cpus per node (~64)


# other useful slurm commands
# --nodelist=h100-ord01-03-[]
# --exclude=h100-ord01-03-[721,138,520,120,839,526,565,741,375,696,684,431]

# this is fairly arbity
export MASTER_ADDR=$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)
export MASTER_PORT=12885

# cluster specific defaults
export OMP_NUM_THREADS=1
export NCCL_CROSS_NIC=2

# explicit cache dirs to not get user conflicts
export TRITON_CACHE_DIR=/mnt/localdisk/.triton_cache_victor

# using bare python rather than torchrun prevents hanging on nightly
# -K1 means slurm job will crash if run crashes rather than just hang
# -u means python will stream stoud as it comes rather than buffer
echo "Starting train script..."
srun -K1 "/home/christian/miniconda3/envs/suno_clone_diff/bin/python" -u train.py \
    --master_addr=$MASTER_ADDR \
    --master_port=$MASTER_PORT \
    --config config/25hz_v45_stemz_main.json