山本さんのバージョン4を、HLE全問推論用に更新したものです

cat > ~/run_deepseek.sbatch <<'SBATCH'
#!/bin/bash
#SBATCH --job-name=deepseek_r1_hle
#SBATCH --partition=P10
#SBATCH --nodes=2
#SBATCH --ntasks-per-node=1
#SBATCH --gpus-per-node=8
#SBATCH --cpus-per-task=240
#SBATCH --mem=0
#SBATCH --time=12:00:00
#SBATCH --output=%x-%j.out
#SBATCH --error=%x-%j.err
#SBATCH --export=ALL

ulimit -n 65536
ulimit -u 131072
ulimit -s unlimited

set -eo pipefail

#=============== ユーザ設定（必要なら書き換え） ===============
LOG_DIR="${LOG_DIR:-$HOME/logs}"
MODEL_PATH="${MODEL_PATH:-/home/Competition2025/P10/shareP10/llm_project/models/deepseek/deepseek-r1-0528-merged/20250818_medmcqa_n10/fp8}"
DEEPSEEK_ENV="${DEEPSEEK_ENV:-$HOME/conda_env_llm}"
VLLM_API_KEY="${VLLM_API_KEY:-p10-deepseek-key}"

export LOG_DIR
export MODEL_PATH
export DEEPSEEK_ENV
export VLLM_API_KEY

#=============== モード切替（--skip-env対応） ===============
MODE_ARG="${1:-}"
case "$MODE_ARG" in
  --skip-env) SKIP_ENV_SETUP=1 ;;
  --setup-env) SKIP_ENV_SETUP=0 ;;
  *) : ;;
esac
: "${SKIP_ENV_SETUP:=0}"

mkdir -p "$LOG_DIR"

#=============== 前処理 / 片付け ===============
cleanup () {
  [[ -n "${pid_vllm:-}"        ]] && kill "${pid_vllm}"        2>/dev/null || true
  [[ -n "${pid_vllm_pretty:-}" ]] && kill "${pid_vllm_pretty}" 2>/dev/null || true
  [[ -n "${pid_nvsmi:-}"       ]] && kill "${pid_nvsmi}"       2>/dev/null || true
  srun --overlap -N ${SLURM_NNODES:-1} -n ${SLURM_NNODES:-1} /bin/bash -lc '
    source /home/appli/miniconda3/24.7.1-py311/etc/profile.d/conda.sh
    conda activate "$DEEPSEEK_ENV" || true
    timeout 10s ray stop --force || true
    pkill -9 ray || true
    pkill -9 gcs_server || true
    find /tmp/ray -maxdepth 1 -type d -name "session_*" -user "$USER" -print0 2>/dev/null | xargs -0r rm -rf
    find /tmp -maxdepth 1 -type d -name "ray_tmp_*" -user "$USER" -print0 2>/dev/null | xargs -0r rm -rf
    rm -rf /dev/shm/ray* 2>/dev/null || true
  ' >/dev/null 2>&1 || true
}
export PYTHONNOUSERSITE=1
unset PYTHONPATH
export PIP_USER=no
export TORCH_COMPILE_DISABLE=1
export TORCHDYNAMO_DISABLE=1
trap cleanup EXIT ERR SIGINT SIGTERM SIGQUIT

#=============== Hugging Face 認証とキャッシュ ===============
export HF_TOKEN="${HF_TOKEN:-<your_hf_token>}"
export HF_HOME=${SLURM_TMPDIR:-$HOME}/.hf_cache
export TRANSFORMERS_CACHE=$HF_HOME
export HUGGINGFACE_HUB_TOKEN=$HF_TOKEN
mkdir -p "$HF_HOME"

#=============== GPU 監視 ======================
nvidia-smi -i 0,1,2,3,4,5,6,7 -l 3 > "$LOG_DIR/nvidia-smi.$SLURM_JOB_ID.log" &
pid_nvsmi=$!

#=============== モジュールと Conda 環境（条件付き） ===============
module reset
module load nccl/2.24.3
module load hpcx/2.18.1-gcc-cuda12/hpcx-mt
module load cudnn/9.4.0
module load miniconda/24.7.1-py311
source /home/appli/miniconda3/24.7.1-py311/etc/profile.d/conda.sh

if [[ "$SKIP_ENV_SETUP" -eq 1 ]]; then
  conda activate "$DEEPSEEK_ENV" || true
  export PATH="$DEEPSEEK_ENV/bin:$PATH"
else
  [ -d "$DEEPSEEK_ENV" ] && conda env remove --prefix "$DEEPSEEK_ENV" -y && rm -rf "$DEEPSEEK_ENV"
  conda create --prefix "$DEEPSEEK_ENV" python=3.11 -y
  conda activate "$DEEPSEEK_ENV"
  export PATH="$DEEPSEEK_ENV/bin:$PATH"
  conda install cuda-toolkit=12.6 -c conda-forge -y
  conda install gcc_linux-64 gxx_linux-64 -y
  pip install --upgrade pip wheel cmake ninja
  pip install huggingface-hub hydra-core openai
  pip cache purge && rm -rf ~/.cache/pip/*
  pip install vllm==0.9.2
  pip install --index-url https://download.pytorch.org/whl/cu126 torch==2.7.0 torchvision==0.22.0+cu126 torchaudio==2.7.0
  pip install "ray[cgraph,data,train,tune,serve]==2.43.0"
  pip install transformers==4.53.2 accelerate==1.2.1 sentencepiece==0.2.0 \
             protobuf==4.25.5 pydantic==2.10.3 aiohttp==3.11.10 \
             numpy==2.2.6 scipy==1.14.1 datasets requests tqdm psutil gputil
  pip install https://github.com/Dao-AILab/flash-attention/releases/download/v2.8.2/flash_attn-2.8.2+cu12torch2.7cxx11abiTRUE-cp311-cp311-linux_x86_64.whl
  pip install flashinfer-python
fi

#=============== Ray 前処理（自分の所有物のみ掃除） ===============
srun --overlap -N ${SLURM_NNODES:-1} -n ${SLURM_NNODES:-1} /bin/bash -lc '
  source /home/appli/miniconda3/24.7.1-py311/etc/profile.d/conda.sh
  conda activate "$DEEPSEEK_ENV" || true
  export PYTHONNOUSERSITE=1; unset PYTHONPATH; export PATH="$DEEPSEEK_ENV/bin:$PATH"
  timeout 10s ray stop --force || true
  pkill -9 ray || true
  pkill -9 gcs_server || true
  find /tmp/ray -maxdepth 1 -type d -name "session_*" -user "$USER" -print0 2>/dev/null | xargs -0r rm -rf
  find /tmp -maxdepth 1 -type d -name "ray_tmp_*" -user "$USER" -print0 2>/dev/null | xargs -0r rm -rf
  rm -rf /dev/shm/ray* 2>/dev/null || true
'

#=============== ノード情報・Ray起動 ===============
HOSTS=($(scontrol show hostnames "$SLURM_JOB_NODELIST"))
HEAD_NODE=${HOSTS[0]}
HEAD_IF=enp25s0np0
HEAD_IP=$(ip -o -4 addr show dev "$HEAD_IF" | awk '{print $4}' | cut -d/ -f1)
[[ -z "$HEAD_IP" ]] && HEAD_IP=$(hostname -i | awk '{print $1}')

export RAY_TMPDIR="${SLURM_TMPDIR:-/tmp}/ray-${SLURM_JOB_ID:-manual}"
mkdir -p "$RAY_TMPDIR"

export VLLM_HOST_IP=$HEAD_IP          # Rayのnodeラベル用に実IPを使う
export NCCL_SOCKET_IFNAME=enp25s0np0
export NCCL_IB_DISABLE=0
"$DEEPSEEK_ENV/bin/ray" start --head --node-ip-address="$HEAD_IP" --port=6379 \
  --dashboard-host=0.0.0.0 --dashboard-port=8265 \
  --temp-dir "$RAY_TMPDIR"
export RAY_ADDRESS=$HEAD_IP:6379

if [[ ${#HOSTS[@]} -gt 1 ]]; then
  for node in "${HOSTS[@]:1}"; do
    srun -N1 -n1 -w "$node" --overlap /bin/bash -lc '
      source /home/appli/miniconda3/24.7.1-py311/etc/profile.d/conda.sh
      conda activate "$DEEPSEEK_ENV" || true
      ulimit -n 65536
      ulimit -u 131072
      ulimit -s unlimited
      ulimit -l unlimited
      export PYTHONNOUSERSITE=1; unset PYTHONPATH; export PATH="$DEEPSEEK_ENV/bin:$PATH"
      export NCCL_SOCKET_IFNAME=enp25s0np0; export NCCL_IB_DISABLE=0
      export RAY_TMPDIR="${SLURM_TMPDIR:-/tmp}/ray-${SLURM_JOB_ID:-manual}"
      mkdir -p "$RAY_TMPDIR"
      IF=enp25s0np0
      WORKER_IP=$(ip -o -4 addr show dev "$IF" | awk "{print \$4}" | cut -d/ -f1)
      [[ -z "$WORKER_IP" ]] && WORKER_IP=$(hostname -i | awk "{print \$1}")
      export VLLM_HOST_IP=$WORKER_IP
      "$DEEPSEEK_ENV/bin/ray" start --address="$RAY_ADDRESS" --node-ip-address="$WORKER_IP" \
        --temp-dir "$RAY_TMPDIR" --block
    ' &
  done
  # === srun を使わず、ヘッド内でローカルにポーリング ===
  for i in {1..60}; do
    python - <<'PY' 2>/dev/null | tail -n1 > .ray_nodes_count.tmp || true
import os, ray
ray.init(address=os.environ.get("RAY_ADDRESS","auto"), log_to_driver=False)
print(len(ray.nodes()))
PY
    ok="$(cat .ray_nodes_count.tmp 2>/dev/null || echo 0)"
    [[ "$ok" == "${SLURM_NNODES}" ]] && echo "all workers joined" && break
    sleep 2
  done
fi

#=============== vLLM サーバ起動（バックグラウンド） ===============
cat > deepseek_server.sh << 'EOF'
#!/bin/bash
export VLLM_USE_RAY_SPMD_WORKER=0
export MODEL_PATH="$MODEL_PATH"
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
export NCCL_SOCKET_IFNAME=enp25s0np0
export NCCL_IB_DISABLE=0
export NCCL_DEBUG=WARN
export RAY_ADDRESS=${RAY_ADDRESS:-"auto"}
export PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:128
export CUDA_LAUNCH_BLOCKING=1

"$DEEPSEEK_ENV"/bin/python -m vllm.entrypoints.openai.api_server --model "$MODEL_PATH" \
  --served-model-name deepseek-r1 \
  --tensor-parallel-size 16 \
  --enable-expert-parallel \
  --distributed-executor-backend ray \
  --trust-remote-code \
  --dtype auto \
  --enforce-eager \
  --gpu-memory-utilization 0.90 \
  --max-model-len 8192 \
  --max-num-seqs 64 \
  --max-num-batched-tokens 30000 \
  --host 0.0.0.0 --port 8000 \
  --api-key "$VLLM_API_KEY" \
  --disable-log-requests \
  --swap-space 16
EOF
chmod +x deepseek_server.sh

./deepseek_server.sh > "$LOG_DIR/vllm_server.$SLURM_JOB_ID.log" 2>&1 & pid_vllm=$!
stdbuf -oL -eL tail -n +1 -F "$LOG_DIR/vllm_server.$SLURM_JOB_ID.log" \
  | awk '{ print strftime("[%Y-%m-%d %H:%M:%S]"), "[vllm]", $0; fflush(); }' \
  > "$LOG_DIR/vllm_server.$SLURM_JOB_ID.pretty.log" & pid_vllm_pretty=$!

#=============== レディネス確認 ===============
OPENAI_BASE_URL="http://127.0.0.1:8000/v1"
for i in $(seq 1 180); do
  resp=$(curl -fsS --max-time 2 -H "Authorization: Bearer ${VLLM_API_KEY}" \
         "${OPENAI_BASE_URL}/models" 2>/dev/null || true)
  echo "$resp" | grep -q '"id"[[:space:]]*:[[:space:]]*"deepseek-r1"' && echo "vLLM HTTP API is ready." && break
  sleep 2
done

#=============== 推論・評価 ===============
cd ~/llm_bridge_prod/eval_hle/ || exit 0
mkdir -p predictions judged
"$DEEPSEEK_ENV/bin/python" predict.py > predict.log 2>&1
"$DEEPSEEK_ENV/bin/python" judge.py   > judge.log   2>&1

#=============== 後片付け ===============
kill "${pid_vllm}" 2>/dev/null || true
kill "${pid_vllm_pretty}" 2>/dev/null || true
kill "${pid_nvsmi}" 2>/dev/null || true
srun --overlap -N ${SLURM_NNODES:-1} -n ${SLURM_NNODES:-1} /bin/bash -lc '
  source /home/appli/miniconda3/24.7.1-py311/etc/profile.d/conda.sh
  conda activate "$DEEPSEEK_ENV" || true
  timeout 10s ray stop --force || true
  pkill -9 ray || true
  pkill -9 gcs_server || true
  find /tmp/ray -maxdepth 1 -type d -name "session_*" -user "$USER" -print0 2>/dev/null | xargs -0r rm -rf
  find /tmp -maxdepth 1 -type d -name "ray_tmp_*" -user "$USER" -print0 2>/dev/null | xargs -0r rm -rf
  rm -rf /dev/shm/ray* 2>/dev/null || true
' || true
echo "Done. Logs: $LOG_DIR"
SBATCH

<aside> ⛔

run_deepseek.sbatch実行で、環境変数が空あるいは未伝達というエラーが出た場合

</aside>

<aside> ⛔

環境変数が空あるいは未伝達の場合は、下記で実行

環境構築あり

JOBID=$(sbatch --parsable -p P10 \
  --export=ALL,OPENAI_API_KEY="your_api_key",HF_TOKEN="your_hf_path",MODEL_PATH="your_model_path",LOG_DIR="$HOME/logs" \
  ~/run_deepseek.sbatch)
echo "Submitted $JOBID"

環境構築なし（推論のみ）

JOBID=$(sbatch --parsable -p P10 \
  --export=ALL,OPENAI_API_KEY="your_api_key",HF_TOKEN="your_hf_path",MODEL_PATH="your_model_path",LOG_DIR="$HOME/logs" \
  ~/run_deepseek.sbatch --skip-env)
echo "Submitted $JOBID"

</aside>