82 lines
2.9 KiB
Bash
Executable File
82 lines
2.9 KiB
Bash
Executable File
#!/bin/bash
|
||
# 在宿主机本地构建 SciCode 沙箱镜像（不从 Docker Hub 拉 scicode-benchmark）。
|
||
# 用法与 Terminal Bench 预加载相同：先在宿主机准备好镜像，再 docker run 评测容器并挂 docker.sock。
|
||
#
|
||
# bash bash/images_load/preload_k3_sandbox_images.sh
|
||
#
|
||
# 注意：
|
||
# - 评测仍在 evalscope 容器里跑，沙箱是宿主机 Docker 再起的子容器
|
||
# - SciCode 首次运行还会按内容哈希打 tag，有本地层缓存就不会再拉 Hub
|
||
# - 构建时不访问 PyPI：先在宿主机用国内镜像下载 cp311 wheel，再离线 pip install
|
||
set -euo pipefail
|
||
|
||
PROJECT_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
|
||
SCICODE_DIR="$PROJECT_ROOT/evalscope/evalscope/benchmarks/scicode/docker"
|
||
WHEEL_DIR="$SCICODE_DIR/wheels"
|
||
H5_URL="https://modelscope.cn/datasets/evalscope/SciCode/resolve/master/test_data.h5"
|
||
PIP_INDEXES=(
|
||
"https://mirrors.aliyun.com/pypi/simple/"
|
||
"https://pypi.tuna.tsinghua.edu.cn/simple"
|
||
)
|
||
|
||
download_wheels() {
|
||
mkdir -p "$WHEEL_DIR"
|
||
local index
|
||
for index in "${PIP_INDEXES[@]}"; do
|
||
echo " pip download via $index"
|
||
if python3 -m pip download \
|
||
-d "$WHEEL_DIR" \
|
||
-r "$SCICODE_DIR/docker_requirements.txt" \
|
||
--python-version 3.11 \
|
||
--implementation cp \
|
||
--abi cp311 \
|
||
--platform manylinux2014_x86_64 \
|
||
--only-binary=:all: \
|
||
-i "$index" \
|
||
--default-timeout=120; then
|
||
return 0
|
||
fi
|
||
echo " 该源失败，尝试下一个..."
|
||
done
|
||
echo "ERROR: 无法下载 SciCode 构建所需的 wheel"
|
||
return 1
|
||
}
|
||
|
||
echo "==> 检查基座镜像（已有则跳过 pull）"
|
||
if docker image inspect python:3.11-slim >/dev/null 2>&1; then
|
||
echo " ok python:3.11-slim"
|
||
else
|
||
echo " docker pull python:3.11-slim"
|
||
docker pull python:3.11-slim
|
||
fi
|
||
|
||
echo "==> SciCode: 确保 test_data.h5"
|
||
if [ ! -f "$SCICODE_DIR/test_data.h5" ]; then
|
||
echo " downloading test_data.h5 ..."
|
||
curl -L --fail -o "$SCICODE_DIR/test_data.h5" "$H5_URL"
|
||
fi
|
||
ls -lh "$SCICODE_DIR/test_data.h5"
|
||
|
||
echo "==> SciCode: 确保离线 wheel（避免构建时访问 files.pythonhosted.org 超时）"
|
||
need_wheels=0
|
||
if [ ! -d "$WHEEL_DIR" ] || [ -z "$(ls -A "$WHEEL_DIR"/*.whl 2>/dev/null || true)" ]; then
|
||
need_wheels=1
|
||
else
|
||
echo " 已有 wheel，跳过下载"
|
||
ls -lh "$WHEEL_DIR"
|
||
fi
|
||
if [ "$need_wheels" -eq 1 ]; then
|
||
download_wheels
|
||
ls -lh "$WHEEL_DIR"
|
||
fi
|
||
|
||
echo "==> docker build scicode-benchmark:latest（离线安装 wheel）"
|
||
docker build --network=host -t scicode-benchmark:latest "$SCICODE_DIR"
|
||
|
||
echo ""
|
||
echo "完成。本地镜像："
|
||
docker images --format 'table {{.Repository}}\t{{.Tag}}\t{{.Size}}' | grep -E 'REPOSITORY|scicode-benchmark|python'
|
||
echo ""
|
||
echo "下一步：docker run 评测镜像，挂 -v /var/run/docker.sock:/var/run/docker.sock"
|
||
echo "SciCode 需要 run.py 打开 sandbox（已加入 SANDBOX_DATASETS）。"
|