Update .gitlab-ci.yml

.gitlab-ci.yml

@@ -5,51 +5,86 @@ image: image.sourcefind.cn:5000/dcu/admin/base/vllm:0.9.2-ubuntu22.04-dtk25.04.1
 stages:
   - test
 
-# === 唯一的作业：使用预置文件运行基准测试 ===
+# === 唯一的作业：完整前置操作 + for 循环串行测试 ===
 benchmark_hunyuan-dit:
   stage: test
   tags:
     - demos
 
-  # 【核心修改 1】使用矩阵策略，自动创建多个并行 Job
-  # GitLab 会为 BATCH_SIZE 列表中的每个值创建一个独立的 Job
-  parallel:
-    matrix:
-      - BATCH_SIZE: [1, 2, 4, 8]
-
   script:
     - echo "========================================="
-    - echo "Step 1:Setting up the environment for BATCH_SIZE = $BATCH_SIZE"
+    - echo "Step 1:Setting up the environment from /workspace/packages/hunyuan-dit (One-time setup)"
     - echo "========================================="
     # 定义预置文件的根目录
     - export PACKAGE_DIR="/workspace/packages/hunyuan-dit"
-    # 从预置目录解压库文件
+    # 从预置目录解压库文件 (这些操作只做一次！)
     - tar -xzf "$PACKAGE_DIR/hipblaslt-install0925.tar.gz"
     - tar -xzf "$PACKAGE_DIR/package_0915_ubuntu.tar.gz"
 
     - echo "========================================="
-    - echo "Step 2:Installing Python packages from /workspace/packages/hunyuan-dit"
+    - echo "Step 2:Installing Python packages from /workspace/packages/hunyuan-dit (One-time setup)"
     - echo "========================================="
-    # 从预置目录安装 Python 包
+    # 从预置目录安装 Python 包 (这些操作也只做一次！)
     - pip install "$PACKAGE_DIR/apex-1.5.0+das.opt1.dtk25041-cp310-cp310-linux_x86_64.whl" "$PACKAGE_DIR/deepspeed-0.14.2+das.opt1.dtk25041-cp310-cp310-manylinux_2_28_x86_64.whl" "$PACKAGE_DIR/lightop-0.5.0+das.dtk25041.unknown-cp310-cp310-linux_x86_64.whl" 
 
     - echo "========================================="
-    - echo "Step 3:Running the benchmark with BATCH_SIZE = $BATCH_SIZE"
+    - echo "Step 3:Starting serial benchmark with a for-loop"
+    - echo "========================================="
+
+    # 【核心】定义要测试的批次大小列表
+    - BATCH_SIZES=(1 2 4 8)
+
+    # 创建一个最终的产物目录，用于存放所有批次的测试结果
+    - mkdir -p all_results
+
+    # 【核心】使用 for 循环串行执行
+    - |
+      for bs in "${BATCH_SIZES[@]}"; do
+        echo "============================================================"
+        echo ">>> Running benchmark for BATCH_SIZE = $bs"
+        echo "============================================================"
+        
+        # 设置本次运行的环境变量 (这些变量在每次循环中都需要)
+        export LD_LIBRARY_PATH=/workspace/packages/hunyuan-dit/hipblaslt-install/lib/:$LD_LIBRARY_PATH
+        export LD_LIBRARY_PATH=/workspace/packages/hunyuan-dit/package/miopen/lib/:$LD_LIBRARY_PATH
+        
+        # 运行测试脚本，将 $bs 作为 batch-size 参数
+        python sample_t2i_dcu.py \
+          --model-root /workspace/OEM_ADVTG_TEST/hunyuan/HunyuanDiT-v1.2/ \
+          --batch-size $bs \
+          --infer-mode fa \
+          --prompt "青花瓷风格，一只可爱的哈士奇" \
+          --no-enhance \
+          --load-key module \
+          --image-size 1024 1024 \
+          --infer-steps 20
+
+        # 【关键】整理本次循环的产物
+        # 检查 results 目录是否存在且不为空
+        if [ -d "results" ] && [ "$(ls -A results)" ]; then
+          echo ">>> Saving results for batch size $bs..."
+          # 将本次的 results 目录重命名并移动到 all_results 中
+          mv results "all_results/results_bs_$bs"
+        else
+          echo ">>> No results found for batch size $bs, skipping."
+        fi
+        
+        # 清理环境，为下一次循环做准备（可选，但推荐）
+        rm -rf results # 确保没有残留的空目录
+        echo ">>> Benchmark for batch size $bs finished."
+        echo ""
+      done
+
+    - echo "========================================="
+    - echo "All benchmarks finished."
+    - echo "Final result directory structure:"
+    - ls -R all_results # 打印最终的目录结构，方便检查
     - echo "========================================="
-    # 设置环境变量，模型路径指向预置目录
-    - export LD_LIBRARY_PATH=/workspace/packages/hunyuan-dit/hipblaslt-install/lib/:$LD_LIBRARY_PATH
-    - export LD_LIBRARY_PATH=/workspace/packages/hunyuan-dit/package/miopen/lib/:$LD_LIBRARY_PATH
-    
-    # 【核心修改 2】使用矩阵变量 $BATCH_SIZE
-    # 运行测试脚本，batch-size 从固定的 4 改为动态的 $BATCH_SIZE
-    - python sample_t2i_dcu.py --model-root /workspace/OEM_ADVTG_TEST/hunyuan/HunyuanDiT-v1.2/ --batch-size $BATCH_SIZE --infer-mode fa --prompt "青花瓷风格，一只可爱的哈士奇" --no-enhance --load-key module --image-size 1024 1024 --infer-steps 20
 
-  # 【核心修改 3】优化产物定义
+  # 【优化】定义最终的产物
   artifacts:
-    # 为产物命名，包含批次大小，方便在 GitLab UI 上区分
-    name: "benchmark-hunyuan-dit-bs-$BATCH_SIZE"
-    # 【优化】指定 'results/' 目录，因为你的 Python 脚本把图片保存在这里
-    # 这比 "*.png" 更精确，能确保只收集我们想要的文件
+    name: "benchmark-hunyuan-dit-serial-results"
+    # 只上传包含所有批次结果的最终目录
     paths:
-      - results/
+      - all_results/
     expire_in: 1 week