File size: 4,328 Bytes
abdaf13 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 | #!/bin/bash
# =============================================================================
# EvalPlus 代码评测 自包含脚本 (HumanEval+ / MBPP+, vLLM 后端, greedy pass@1)
# 由 code_eval/run_evalplus.sh 抽取, 路径改为本目录相对, 可独立搬运。
# 官方项目: https://github.com/evalplus/evalplus (已装在集群 vllm conda 环境)
#
# 数据: ./data/ (HumanEvalPlus-v0.1.10.jsonl, MbppPlus-v0.2.0.jsonl);
# 运行时若 ~/.cache/evalplus 缺数据, 自动从 ./data 离线补齐 (不联网也能跑)
#
# 用法:
# sbatch --export=ALL,MODEL=/data/icto/mc56499/model/Qwen2.5-Coder-3B \
# /data/icto/mc56499/eval/code/run_evalplus.sh # pro6000 2卡
# sbatch --export=ALL,MODEL=... -p gpu-l40s run_evalplus.sh # 换分区
# bash run_evalplus.sh [/模型路径] [TP] # srun/交互调试
#
# 产物: <本目录>/evalplus_results/
# samples: evalplus_results/<dataset>/<model>_vllm_temp_0.0.jsonl
# 报告: 同目录 *_eval_results.json (含 pass@1 base/plus)
# =============================================================================
#SBATCH --job-name code_eval
#SBATCH --partition gpu-pro6000
#SBATCH --nodes 1
#SBATCH --tasks-per-node 1
#SBATCH --time 06:00:00
#SBATCH --mem 128G
#SBATCH --cpus-per-task 16
#SBATCH --gres gpu:2
#SBATCH --output /data/icto/mc56499/eval/logs/code_eval-%j.out
#SBATCH --error /data/icto/mc56499/eval/logs/code_eval-%j.err
set -e
SCRIPT_DIR=$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)
cd "$SCRIPT_DIR"
mkdir -p /data/icto/mc56499/eval/logs
# 环境防护: IDE/TRAE 终端提交时继承的 LD_LIBRARY_PATH 污染 (旧 libstdc++)
unset LD_LIBRARY_PATH LD_PRELOAD
module load nvidia
module load cuda/13.0.2
source /data/icto/mc56499/miniconda3/etc/profile.d/conda.sh
conda activate vllm # evalplus.* 入口脚本安装在此环境
export PYTHONUNBUFFERED=1
export TOKENIZERS_PARALLELISM=false
export HF_HOME=/data/icto/mc56499/hf_cache
export HF_HUB_CACHE="$HF_HOME"
# ====== 离线数据种子: ~/.cache/evalplus 缺数据时从 ./data 补齐 ======
mkdir -p "$HOME/.cache/evalplus"
for f in data/*.jsonl; do
[ -f "$f" ] || continue
dst="$HOME/.cache/evalplus/$(basename "$f")"
if [ ! -f "$dst" ]; then
cp "$f" "$dst"
echo "已补齐 evalplus 数据: $dst"
fi
done
# ====== 模型与并行度 ======
MODEL_PATH=${MODEL:-${1:-}}
if [ -z "$MODEL_PATH" ]; then
echo "用法: sbatch --export=ALL,MODEL=/path/to/model run_evalplus.sh"
echo " 或: bash run_evalplus.sh /path/to/model [TP]"
exit 1
fi
TP=${2:-auto}
if [ "$TP" = "auto" ] || [ -z "$TP" ]; then
TP=$(nvidia-smi -L 2>/dev/null | wc -l)
fi
[ "$TP" -lt 1 ] && TP=1
ROOT="$SCRIPT_DIR/evalplus_results"
PARALLEL=16 # 评测进程数 (代码沙箱执行)
echo "========================================================"
echo "=== EvalPlus 代码评测: $(basename "$MODEL_PATH") ==="
echo "=== 数据集: HumanEval+ / MBPP+ GPU: $TP 张 ==="
echo "=== 输出: $ROOT ==="
echo "=== 开始: $(date) ==="
echo "========================================================"
mkdir -p "$ROOT"
for DATASET in humaneval mbpp; do
echo ""
echo "===== [EvalPlus] dataset=$DATASET model=$MODEL_PATH tp=$TP ====="
evalplus.evaluate \
--model "$MODEL_PATH" \
--dataset "$DATASET" \
--backend vllm \
--tp "$TP" \
--greedy \
--force_base_prompt \
--root "$ROOT" \
--parallel "$PARALLEL"
done
# ====== 结果汇总 ======
echo ""
echo "===== EvalPlus pass@1 汇总 ====="
for f in $(find "$ROOT" -name "*_eval_results.json" | sort); do
echo "--- $f"
python - "$f" <<'EOF'
import json, sys
with open(sys.argv[1]) as f:
r = json.load(f)
tasks = r["eval"].values()
base = sum(1 for samples in tasks if any(s["base_status"] == "pass" for s in samples))
plus = sum(1 for samples in r["eval"].values() if any(
s["base_status"] == "pass" and s["plus_status"] == "pass" for s in samples))
n = len(r["eval"])
print(f" pass@1 (base tests): {base / n:.4f} ({base}/{n})")
print(f" pass@1 (base+extra): {plus / n:.4f} ({plus}/{n})")
EOF
done
echo ""
echo "===== DONE: $(date '+%F %T') ====="
|