gistfile1.txt

#!/bin/bash
#SBATCH --nodes=2
#SBATCH --ntasks-per-node=1
#SBATCH --gpus-per-node=1
#SBATCH --cpus-per-task=16
#SBATCH --gres=gpu:1
#SBATCH --time=00:20:00
#SBATCH --output=logs/slurm-%j.out
#SBATCH --error=logs/slurm-%j.err
#SBATCH --mail-type=END,FAIL,BEGIN
#SBATCH [email protected]
#SBATCH -p lrz-hgx-h100-94x4


# 必要环境变量
export VLLM_ATTENTION_BACKEND=XFORMERS
export WANDB_API_KEY=xxxxxxxxxx

# 获取所有节点名称
nodes=$(scontrol show hostnames "$SLURM_JOB_NODELIST")
nodes_array=($nodes)

# 获取 Head 节点 IP
head_node=${nodes_array[0]}
head_node_ip=$(srun --nodes=1 --ntasks=1 -w "$head_node" hostname --ip-address)


# if we detect a space character in the head node IP, we'll
# convert it to an ipv4 address. This step is optional.
if [[ "$head_node_ip" == *" "* ]]; then
IFS=' ' read -ra ADDR <<<"$head_node_ip"
if [[ ${#ADDR[0]} -gt 16 ]]; then
  head_node_ip=${ADDR[1]}
else
  head_node_ip=${ADDR[0]}
fi
echo "IPV6 address detected. We split the IPV4 address as $head_node_ip"
fi

port=6379
ip_head=$head_node_ip:$port
export ip_head

source ~/.bashrc
conda activate rllm

echo "Head Node IP: $ip_head"

# 启动 Head 节点上的 Ray
echo "Starting HEAD node: $head_node"
srun --nodes=1 --ntasks=1 -w "$head_node" \
    ray start --head --node-ip-address="$head_node_ip" --port=$port \
    --num-cpus=$SLURM_CPUS_PER_TASK --num-gpus=$SLURM_GPUS_PER_NODE --block &

sleep 10  # 等待 Head 启动

# number of nodes other than the head node
worker_num=$((SLURM_JOB_NUM_NODES - 1))

for ((i = 1; i <= worker_num; i++)); do
    node_i=${nodes_array[$i]}
    echo "Starting WORKER $i at $node_i"
    srun --nodes=1 --ntasks=1 -w "$node_i" \
            ray start --address "$ip_head" --num-cpus "${SLURM_CPUS_PER_TASK}" --num-gpus "${SLURM_GPUS_PER_NODE}" --block &
    sleep 5
done


# Parse command line arguments
while [[ $# -gt 0 ]]; do
    case $1 in
        --model)
            MODEL_PATH="$2"
            shift 2
            ;;
        *)
            break
            ;;
    esac
done

# Set default model path if not provided
if [ -z "${MODEL_PATH:-}" ]; then
    MODEL_PATH="hfmodels/DeepSeek-R1-Distill-Qwen-1.5B"
fi

MODEL_PATH=$(readlink -f "$MODEL_PATH")

# 启动训练（仍然在 Head 节点）
srun --nodes=2 --ntasks-per-node=1 -w "$head_node" \
    python -m verl.trainer.main_ppo_remote \
        algorithm.adv_estimator=grpo \
        "data.train_files=${HOME}/deepscaler/data/train.parquet" \
        "data.val_files=${HOME}/deepscaler/data/aime.parquet" \
        data.train_batch_size=64 \
        data.val_batch_size=512 \
        data.max_prompt_length=1024 \
        data.max_response_length=8192 \
        "actor_rollout_ref.model.path=${MODEL_PATH}" \
        actor_rollout_ref.actor.optim.lr=1e-6 \
        actor_rollout_ref.model.use_remove_padding=True \
        actor_rollout_ref.actor.ppo_mini_batch_size=64 \
        actor_rollout_ref.actor.use_dynamic_bsz=True \
        actor_rollout_ref.actor.ppo_max_token_len_per_gpu=32768 \
        actor_rollout_ref.actor.use_kl_loss=True \
        actor_rollout_ref.actor.kl_loss_coef=0.001 \
        actor_rollout_ref.actor.kl_loss_type=low_var_kl \
        actor_rollout_ref.actor.ulysses_sequence_parallel_size=1 \
        actor_rollout_ref.model.enable_gradient_checkpointing=True \
        actor_rollout_ref.actor.fsdp_config.param_offload=False \
        actor_rollout_ref.actor.fsdp_config.grad_offload=False \
        actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \
        actor_rollout_ref.rollout.tensor_model_parallel_size=1 \
        actor_rollout_ref.rollout.name=vllm \
        actor_rollout_ref.rollout.temperature=0.6 \
        actor_rollout_ref.rollout.val_temperature=0.6 \
        actor_rollout_ref.rollout.gpu_memory_utilization=0.85 \
        actor_rollout_ref.rollout.n=8 \
        actor_rollout_ref.rollout.n_val=16 \
        actor_rollout_ref.ref.fsdp_config.param_offload=True \
        algorithm.kl_ctrl.kl_coef=0.001 \
        trainer.critic_warmup=0 \
        trainer.logger=['console','wandb'] \
        trainer.project_name='deepscaler' \
        trainer.experiment_name='deepscaler-1.5b-h100-8k' \
        +trainer.val_before_train=False \
        trainer.n_gpus_per_node=1 \
        trainer.nnodes=2 \
        trainer.save_freq=5 \
        trainer.test_freq=5 \
        trainer.default_hdfs_dir=null \
        trainer.total_epochs=30 "$@"