#!/bin/bash
#SBATCH --job-name=gk_gpt45
#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)
#SBATCH --output=/app/suno/slurm/logs/run_%j.txt
#SBATCH --error=/app/suno/slurm/logs/run_%j_err.txt

# other useful slurm commands
# --exclude=h100-ord01-03-[]
# --nodelist=h100-ord01-03-[]

# 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_$USER

# 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/georg/anaconda3/envs/gpt_n/bin/python -u train.py \
    --master_addr=$MASTER_ADDR \
    --master_port=$MASTER_PORT \
    \
    --out_dir="/app/suno/checkpoints" \
    --data_dir="/app/suno/data/chirp_v5/v0" \
    --weights_multiplier="pond5:0.25" \
    \
    --n_layer=28 \
    --n_head=28 \
    --d_head=112 \
    \
    --learning_rate=5e-4 \
    --max_iters=50_000 \
    --warmup_iters=1_000 \
    --eval_interval=1_000 \
    \
    --fsdp=True \
    --grad_checkpointing=True \
    --compile=False \
    \
    --wandb_log=True \
    --wandb_dir="/app/suno/wandb" \
    --wandb_project="gpt-45" \
    --wandb_run_name="base"