Add checkpoint skip for loaded images and --download-swe flag

README.md

@@ -78,7 +78,19 @@ python bash/run.py \
 
 ## 3. Docker 镜像下载与部署
 
-根据是否跑 SWE-bench，选择下面两种方式之一。
+最简方式：直接运行一键部署脚本。
+
+```bash
+# 默认不下载 SWE-Bench 镜像（体积大）
+bash scripts/deploy_evalscope.sh /data1/sora
+
+# 如果需要跑 SWE-Bench，加 --download-swe
+bash scripts/deploy_evalscope.sh --download-swe /data1/sora
+```
+
+脚本会自动完成：安装 modelscope、登录、下载数据集、下载/加载 Docker 镜像、配置 Docker/containerd、加载 sandbox 镜像。重复执行时会通过 checkpoint 自动跳过已加载的镜像。
+
+如果需要手动分步操作，参考下面两种方式。
 
 ### 3.1 方式 A：不测 SWE-bench 的 Docker 配置
 

scripts/deploy_evalscope.sh

@@ -1,8 +1,13 @@
 #!/bin/bash
 # ============================================================
 # EvalScope 完整部署脚本
-# 用法: bash deploy_evalscope.sh <BASE_DIR> [TOKEN]
+# 用法: bash deploy_evalscope.sh [OPTIONS] <BASE_DIR> [TOKEN]
-# 示例: bash deploy_evalscope.sh /data1/sora ms-3d554a39-6e07-496d-8022-0b0ee64a6389
+# 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
@@ -10,6 +15,36 @@ 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:-$DEFAULT_TOKEN}"
@@ -21,6 +56,25 @@ IMAGES_DIR="$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
@@ -119,7 +173,14 @@ sudo systemctl start docker
 # 7. 加载 evalscope 环境镜像
 # --------------------------------------------------
 echo "==> 加载 evalscope Docker 镜像..."
-docker load -i "$DOCKER_DIR/evalscope-complete-py312.tar.gz"
+if docker images --format '{{.Repository}}:{{.Tag}}' | grep -qx "evalscope-complete-py312:latest"; then
+    echo "evalscope-complete-py312:latest 已存在，跳过加载"
+elif is_loaded "evalscope-complete-py312.tar.gz"; then
+    echo "evalscope-complete-py312.tar.gz 已加载过（checkpoint），跳过加载"
+else
+    docker load -i "$DOCKER_DIR/evalscope-complete-py312.tar.gz"
+    mark_loaded "evalscope-complete-py312.tar.gz"
+fi
 
 # --------------------------------------------------
 # 8. 克隆代码仓库
@@ -167,9 +228,14 @@ load_sandbox_image() {
         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
@@ -196,15 +262,17 @@ else
 fi
 
 # --------------------------------------------------
-# 10. 下载 SWE-Bench 镜像包（可选）
+# 10. 下载 SWE-Bench 镜像包（默认不下载，需加 --download-swe）
 # --------------------------------------------------
-if [ -n "$MODELSCOPE_TOKEN" ]; then
+if [ "$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
 
 # --------------------------------------------------
@@ -217,13 +285,23 @@ if [ -d "$SWE_IMAGES_DIR" ] && ls "$SWE_IMAGES_DIR"/swebench_batch_*.tar.gz 1>/d
     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))
-        echo "[$IDX/$TOTAL] 加载 $(basename "$TAR")..."
+        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，已加载 $(docker images | grep '^swebench/' | wc -l) 个"
+    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