| # ============================================================================= | |
| # serve_judge.sh — Qwen3.5-35B-A3B（採点者）vLLM サーバー起動 | |
| # モデル: Qwen/Qwen3.5-35B-A3B（MoE: 35B全weights/3B active params） | |
| # ポート: 8001 | |
| # VRAM使用量: ~70GB（全weights bf16） | |
| # | |
| # ⚠️ メモリ注意: | |
| # A100 80GB での同時起動: | |
| # - TeenEmo (serve_test.sh, GPU_UTIL=0.10) + Judge (GPU_UTIL=0.88) = ~98% | |
| # - OOM する場合は serve_test.sh を停止してからこちらを起動し、 | |
| # 順次実行モード（setup_eqbench_run.sh の指示に従う）で対応する | |
| # | |
| # Qwen3.5-35B-A3B は VLM のため --language-model-only が必要 | |
| # (pipeline_tag: image-text-to-text, HF: https://huggingface.co/Qwen/Qwen3.5-35B-A3B) | |
| # ============================================================================= | |
| set -euo pipefail | |
| JUDGE_MODEL="${JUDGE_MODEL:-Qwen/Qwen3.5-35B-A3B}" | |
| PORT="${JUDGE_PORT:-8001}" | |
| HOST="${VLLM_HOST:-0.0.0.0}" | |
| GPU_UTIL="${JUDGE_GPU_UTIL:-0.88}" | |
| MAX_MODEL_LEN="${JUDGE_MAX_MODEL_LEN:-8192}" | |
| DTYPE="${VLLM_DTYPE:-auto}" | |
| MAX_NUM_SEQS="${VLLM_MAX_NUM_SEQS:-16}" | |
| echo "=== Qwen3.5-35B-A3B 採点者サーバー起動 ===" | |
| echo " モデル : ${JUDGE_MODEL}" | |
| echo " ポート : ${PORT}" | |
| echo " GPU 使用率 : ${GPU_UTIL}" | |
| echo " ⚠️ TeenEmoと同時起動の場合は serve_test.sh の GPU_UTIL=0.10 を確認してください" | |
| echo "" | |
| python -c "import vllm; print(f' vLLM バージョン : {vllm.__version__}')" 2>/dev/null || true | |
| echo "" | |
| exec vllm serve "${JUDGE_MODEL}" \ | |
| --host "${HOST}" \ | |
| --port "${PORT}" \ | |
| --dtype "${DTYPE}" \ | |
| --gpu-memory-utilization "${GPU_UTIL}" \ | |
| --max-model-len "${MAX_MODEL_LEN}" \ | |
| --tensor-parallel-size 1 \ | |
| --max-num-seqs "${MAX_NUM_SEQS}" \ | |
| --language-model-only \ | |
| --enable-prefix-caching \ | |
| --trust-remote-code | |