365 lines
12 KiB
Bash
Executable File
365 lines
12 KiB
Bash
Executable File
#!/bin/bash
|
||
# ============================================================
|
||
# EvalScope 完整部署脚本
|
||
# 用法: bash deploy_evalscope.sh [OPTIONS] <BASE_DIR> [TOKEN]
|
||
# 示例:
|
||
# bash deploy_evalscope.sh /data1/sora
|
||
# bash deploy_evalscope.sh --download-swe /data1/sora
|
||
#
|
||
# 断点续传：已加载的镜像/tar 包会记录在 <BASE_DIR>/evalscope/docker/.deploy_checkpoint
|
||
# 重复执行时会自动跳过，避免二次加载。
|
||
# ============================================================
|
||
|
||
set -euo pipefail
|
||
|
||
# --------------------------------------------------
|
||
# 1. 参数解析
|
||
# --------------------------------------------------
|
||
DOWNLOAD_SWE=0
|
||
POSITIONAL_ARGS=()
|
||
|
||
while [[ $# -gt 0 ]]; do
|
||
case "$1" in
|
||
--download-swe)
|
||
DOWNLOAD_SWE=1
|
||
shift
|
||
;;
|
||
--help|-h)
|
||
echo "用法: bash deploy_evalscope.sh [OPTIONS] <BASE_DIR> [TOKEN]"
|
||
echo "选项:"
|
||
echo " --download-swe 同时下载 SWE-Bench 镜像包（默认不下载，体积很大）"
|
||
echo " -h, --help 显示帮助"
|
||
exit 0
|
||
;;
|
||
-*)
|
||
echo "未知选项: $1"
|
||
echo "使用 --help 查看用法"
|
||
exit 1
|
||
;;
|
||
*)
|
||
POSITIONAL_ARGS+=("$1")
|
||
shift
|
||
;;
|
||
esac
|
||
done
|
||
|
||
set -- "${POSITIONAL_ARGS[@]}"
|
||
|
||
BASE_DIR="${1:-/data1/sora}" # 基础目录，默认 /data1/sora
|
||
DEFAULT_TOKEN="ms-3d554a39-6e07-496d-8022-0b0ee64a6389"
|
||
MODELSCOPE_TOKEN="${2:-$DEFAULT_TOKEN}"
|
||
|
||
# 派生路径
|
||
EVALSCOPE_DIR="$BASE_DIR/evalscope"
|
||
DOCKER_DIR="$EVALSCOPE_DIR/docker"
|
||
IMAGES_DIR="$DOCKER_DIR/images"
|
||
CONTAINERD_DIR="$DOCKER_DIR/containerd"
|
||
SWE_IMAGES_DIR="$DOCKER_DIR/swe_images"
|
||
DATASETS_DIR="$EVALSCOPE_DIR/datasets"
|
||
CHECKPOINT_FILE="$DOCKER_DIR/.deploy_checkpoint"
|
||
|
||
# 断点：记录/检查已加载的镜像或 tar 包
|
||
mark_loaded() {
|
||
local key="$1"
|
||
mkdir -p "$DOCKER_DIR"
|
||
touch "$CHECKPOINT_FILE"
|
||
if ! grep -qx "$key" "$CHECKPOINT_FILE" 2>/dev/null; then
|
||
echo "$key" >> "$CHECKPOINT_FILE"
|
||
fi
|
||
}
|
||
|
||
is_loaded() {
|
||
local key="$1"
|
||
if [ -f "$CHECKPOINT_FILE" ] && grep -qx "$key" "$CHECKPOINT_FILE" 2>/dev/null; then
|
||
return 0
|
||
fi
|
||
return 1
|
||
}
|
||
|
||
# --------------------------------------------------
|
||
# 2. 安装依赖
|
||
# --------------------------------------------------
|
||
echo "==> 安装 modelscope..."
|
||
pip install modelscope
|
||
|
||
# --------------------------------------------------
|
||
# 3. 登录 ModelScope
|
||
# --------------------------------------------------
|
||
if [ -n "$MODELSCOPE_TOKEN" ]; then
|
||
echo "==> 登录 ModelScope..."
|
||
modelscope login --token "$MODELSCOPE_TOKEN"
|
||
fi
|
||
|
||
# --------------------------------------------------
|
||
# 4. 下载数据集
|
||
# --------------------------------------------------
|
||
echo "==> 下载 evalscope 数据集..."
|
||
mkdir -p "$DATASETS_DIR"
|
||
python3 -c "
|
||
from modelscope.hub.snapshot_download import snapshot_download
|
||
snapshot_download(
|
||
'SoraAmami/evalscope-datasets',
|
||
repo_type='dataset',
|
||
cache_dir='$EVALSCOPE_DIR',
|
||
local_dir='$DATASETS_DIR'
|
||
)
|
||
"
|
||
|
||
# --------------------------------------------------
|
||
# 5. 下载 Docker 镜像包
|
||
# --------------------------------------------------
|
||
echo "==> 下载 evalscope Docker 镜像..."
|
||
mkdir -p "$DOCKER_DIR"
|
||
python3 -c "
|
||
from modelscope.hub.file_download import model_file_download
|
||
model_file_download(
|
||
model_id='SoraAmami/evalscope-docker',
|
||
file_path='evalscope-complete-py312.tar.gz',
|
||
local_dir='$DOCKER_DIR'
|
||
)
|
||
"
|
||
|
||
# --------------------------------------------------
|
||
# 6. 配置 Docker 和 containerd 数据目录
|
||
# --------------------------------------------------
|
||
echo "==> 配置 Docker 和 containerd..."
|
||
|
||
# 停止服务
|
||
sudo systemctl stop docker.socket 2>/dev/null || true
|
||
sudo systemctl stop docker 2>/dev/null || true
|
||
sudo systemctl stop containerd 2>/dev/null || true
|
||
|
||
# 创建数据目录
|
||
mkdir -p "$IMAGES_DIR"
|
||
mkdir -p "$CONTAINERD_DIR"
|
||
|
||
# 配置 Docker
|
||
echo "==> 写入 Docker 配置..."
|
||
sudo tee /etc/docker/daemon.json <<EOF
|
||
{
|
||
"data-root": "$IMAGES_DIR",
|
||
"features": {
|
||
"containerd-snapshotter": true
|
||
},
|
||
"registry-mirrors": [
|
||
"https://docker.m.daocloud.io",
|
||
"https://docker.1ms.run",
|
||
"https://hub.rat.dev",
|
||
"https://docker.1panel.live",
|
||
"https://dockerproxy.com",
|
||
"https://hub-mirror.c.163.com",
|
||
"https://mirror.baidubce.com",
|
||
"https://docker.mirrors.ustc.edu.cn",
|
||
"https://docker.mirrors.sjtug.sjtu.edu.cn",
|
||
"https://docker.nju.edu.cn",
|
||
"https://docker.mirrors.tuna.tsinghua.edu.cn"
|
||
]
|
||
}
|
||
EOF
|
||
|
||
# 配置 containerd
|
||
echo "==> 写入 containerd 配置..."
|
||
sudo tee /etc/containerd/config.toml <<EOF
|
||
root = "$CONTAINERD_DIR"
|
||
state = "/run/containerd"
|
||
EOF
|
||
|
||
# 重启服务
|
||
sudo systemctl reset-failed docker.service 2>/dev/null || true
|
||
sudo systemctl start containerd
|
||
sudo systemctl start docker
|
||
|
||
# --------------------------------------------------
|
||
# 7. 加载 evalscope 环境镜像
|
||
# --------------------------------------------------
|
||
echo "==> 加载 evalscope Docker 镜像..."
|
||
if docker images --format '{{.Repository}}:{{.Tag}}' | grep -qx "evalscope-complete-py312:latest"; then
|
||
echo "evalscope-complete-py312:latest 已存在，跳过加载"
|
||
elif is_loaded "evalscope-complete-py312.tar.gz"; then
|
||
echo "evalscope-complete-py312.tar.gz 已加载过（checkpoint），跳过加载"
|
||
else
|
||
docker load -i "$DOCKER_DIR/evalscope-complete-py312.tar.gz"
|
||
mark_loaded "evalscope-complete-py312.tar.gz"
|
||
fi
|
||
|
||
# --------------------------------------------------
|
||
# 8. 克隆代码仓库到根目录
|
||
# --------------------------------------------------
|
||
echo "==> 克隆 evalscope 代码..."
|
||
|
||
if [ -f "$EVALSCOPE_DIR/bash/run.py" ]; then
|
||
echo "代码已存在于 $EVALSCOPE_DIR，跳过克隆"
|
||
else
|
||
echo "正在克隆代码到 $EVALSCOPE_DIR ..."
|
||
tmp_backup="$BASE_DIR/.evalscope_deploy_backup_$(date +%s)"
|
||
mkdir -p "$tmp_backup"
|
||
|
||
# 保留已有的 datasets/ 和 docker/，其余清空后克隆
|
||
for item in "$EVALSCOPE_DIR"/* "$EVALSCOPE_DIR"/.[!.]* "$EVALSCOPE_DIR"/..?*; do
|
||
[ -e "$item" ] || continue
|
||
basename_item=$(basename "$item")
|
||
case "$basename_item" in
|
||
.|..|.git) continue ;;
|
||
datasets|docker)
|
||
mv "$item" "$tmp_backup/$basename_item" ;;
|
||
*)
|
||
rm -rf "$item" ;;
|
||
esac
|
||
done
|
||
|
||
git clone "https://git.meta-stone.net/sora/evalstone.git" "$EVALSCOPE_DIR"
|
||
|
||
# 恢复 datasets/ docker/
|
||
for item in "$tmp_backup"/* "$tmp_backup"/.[!.]* "$tmp_backup"/..?*; do
|
||
[ -e "$item" ] || continue
|
||
basename_item=$(basename "$item")
|
||
[ "$basename_item" = "." ] || [ "$basename_item" = ".." ] && continue
|
||
if [ -e "$EVALSCOPE_DIR/$basename_item" ]; then
|
||
continue
|
||
fi
|
||
mv "$item" "$EVALSCOPE_DIR/$basename_item"
|
||
done
|
||
rm -rf "$tmp_backup" 2>/dev/null || true
|
||
fi
|
||
|
||
# --------------------------------------------------
|
||
# 9. 准备 sandbox 镜像
|
||
# --------------------------------------------------
|
||
echo "==> 准备 sandbox 镜像..."
|
||
SANDBOX_IMAGES_DIR="$DOCKER_DIR/sandbox_images"
|
||
mkdir -p "$SANDBOX_IMAGES_DIR"
|
||
|
||
# 从 ModelScope 下载预打包的 sandbox 镜像（不再使用 docker pull，避免 DockerHub 限速）
|
||
download_sandbox_image_from_modelscope() {
|
||
local file="$1"
|
||
if [ -f "$SANDBOX_IMAGES_DIR/$file" ]; then
|
||
echo "sandbox 镜像包 $file 已存在，跳过下载"
|
||
return 0
|
||
fi
|
||
echo "==> 从 ModelScope 下载 $file ..."
|
||
python3 -c "
|
||
from modelscope.hub.file_download import model_file_download
|
||
model_file_download(
|
||
model_id='SoraAmami/evalscope-sandbox-images',
|
||
file_path='$file',
|
||
local_dir='$SANDBOX_IMAGES_DIR'
|
||
)
|
||
" && return 0
|
||
echo "警告：从 ModelScope 下载 $file 失败"
|
||
return 1
|
||
}
|
||
|
||
load_sandbox_image() {
|
||
local file="$1"
|
||
local expected_image="$2"
|
||
if docker images --format '{{.Repository}}:{{.Tag}}' | grep -qx "$expected_image"; then
|
||
echo "镜像 $expected_image 已存在，跳过加载"
|
||
return 0
|
||
fi
|
||
if is_loaded "$file"; then
|
||
echo "镜像包 $file 已加载过（checkpoint），跳过加载"
|
||
return 0
|
||
fi
|
||
if [ -f "$SANDBOX_IMAGES_DIR/$file" ]; then
|
||
echo "==> 加载 $file ..."
|
||
docker load -i "$SANDBOX_IMAGES_DIR/$file"
|
||
mark_loaded "$file"
|
||
return 0
|
||
fi
|
||
return 1
|
||
}
|
||
|
||
# 9.1 bigcodebench-sandbox:latest
|
||
if docker images --format '{{.Repository}}:{{.Tag}}' | grep -qx "bigcodebench-sandbox:latest"; then
|
||
echo "bigcodebench-sandbox:latest 已存在，跳过"
|
||
else
|
||
if ! load_sandbox_image "bigcodebench-sandbox.tar.gz" "bigcodebench-sandbox:latest"; then
|
||
download_sandbox_image_from_modelscope "bigcodebench-sandbox.tar.gz" && \
|
||
load_sandbox_image "bigcodebench-sandbox.tar.gz" "bigcodebench-sandbox:latest"
|
||
fi
|
||
fi
|
||
|
||
# 9.2 python:3.11-slim
|
||
if docker images --format '{{.Repository}}:{{.Tag}}' | grep -qx "python:3.11-slim"; then
|
||
echo "python:3.11-slim 已存在，跳过"
|
||
else
|
||
if ! load_sandbox_image "python-3.11-slim.tar.gz" "python:3.11-slim"; then
|
||
download_sandbox_image_from_modelscope "python-3.11-slim.tar.gz" && \
|
||
load_sandbox_image "python-3.11-slim.tar.gz" "python:3.11-slim"
|
||
fi
|
||
fi
|
||
|
||
# --------------------------------------------------
|
||
# 10. 下载 SWE-Bench 镜像包（默认不下载，需加 --download-swe）
|
||
# --------------------------------------------------
|
||
if [ "$DOWNLOAD_SWE" -eq 1 ]; then
|
||
echo "==> 下载 SWE-Bench 镜像包..."
|
||
mkdir -p "$SWE_IMAGES_DIR"
|
||
modelscope download \
|
||
--repo-type dataset \
|
||
--local_dir "$SWE_IMAGES_DIR" \
|
||
SoraAmami/swe-bench-verified-images || echo "警告：SWE-Bench 镜像下载失败"
|
||
else
|
||
echo "==> 跳过 SWE-Bench 镜像包下载（如需下载请加 --download-swe）"
|
||
fi
|
||
|
||
# --------------------------------------------------
|
||
# 11. 加载 SWE-Bench 镜像（如果存在）
|
||
# --------------------------------------------------
|
||
if [ -d "$SWE_IMAGES_DIR" ] && ls "$SWE_IMAGES_DIR"/swebench_batch_*.tar.gz 1>/dev/null 2>&1; then
|
||
echo "==> 加载 SWE-Bench 镜像..."
|
||
TARS=("$SWE_IMAGES_DIR"/swebench_batch_*.tar.gz)
|
||
TOTAL=${#TARS[@]}
|
||
echo "共找到 $TOTAL 个镜像包"
|
||
|
||
IDX=0
|
||
LOADED_COUNT=0
|
||
SKIPPED_COUNT=0
|
||
for TAR in "${TARS[@]}"; do
|
||
IDX=$((IDX + 1))
|
||
BASENAME=$(basename "$TAR")
|
||
if is_loaded "$BASENAME"; then
|
||
echo "[$IDX/$TOTAL] 跳过（已加载过）: $BASENAME"
|
||
SKIPPED_COUNT=$((SKIPPED_COUNT + 1))
|
||
continue
|
||
fi
|
||
echo "[$IDX/$TOTAL] 加载 $BASENAME ..."
|
||
docker load -i "$TAR"
|
||
mark_loaded "$BASENAME"
|
||
LOADED_COUNT=$((LOADED_COUNT + 1))
|
||
done
|
||
|
||
echo "SWE-Bench 镜像加载完成：本次加载 $LOADED_COUNT 个，跳过 $SKIPPED_COUNT 个，已加载 $(docker images | grep -c '^swebench/' || true) 个"
|
||
else
|
||
echo "跳过 SWE-Bench 镜像加载（未找到镜像包）"
|
||
fi
|
||
|
||
# --------------------------------------------------
|
||
# 12. 验证部署
|
||
# --------------------------------------------------
|
||
echo ""
|
||
echo "========================================"
|
||
echo "部署完成！"
|
||
echo "========================================"
|
||
echo "基础目录: $BASE_DIR"
|
||
echo "EvalScope: $EVALSCOPE_DIR"
|
||
echo "数据集: $DATASETS_DIR"
|
||
echo "Docker 数据: $IMAGES_DIR"
|
||
echo "Containerd: $CONTAINERD_DIR"
|
||
echo ""
|
||
echo "Docker Root Dir:"
|
||
docker info 2>/dev/null | grep "Docker Root Dir" || echo "Docker 未运行"
|
||
echo ""
|
||
echo "已加载镜像:"
|
||
docker images | grep -E "evalscope|swebench|bigcodebench|python" || true
|
||
echo ""
|
||
echo "运行 EvalScope:"
|
||
echo " cd $EVALSCOPE_DIR"
|
||
echo " docker run -it --rm \\"
|
||
echo " --network host \\"
|
||
echo " -v $EVALSCOPE_DIR:/opt/evalscope \\"
|
||
echo " -v /var/run/docker.sock:/var/run/docker.sock \\"
|
||
echo " evalscope-complete-py312:latest \\"
|
||
echo " bash"
|
||
echo "========================================"
|