Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into hackathon9th06

Author: WintersMontagne10335

.github/workflows/H-Coverage.yml

@@ -142,7 +142,9 @@ jobs:
         run: |
           docker exec -t ${{ env.container_name }}  /bin/bash -c '
           flashattn_version=$(git submodule status|grep flashattn|awk "{print \$1}"|sed "s#-##g")
+          echo flashattn_version:$flashattn_version
           url="https://xly-devops.bj.bcebos.com/gpups/flash-attention/cu90/flashattn_libs_${flashattn_version}.tar"
+          echo url:$url
           url_return=`curl -s -o /dev/null -w "%{http_code}" $url`
           if [ "$url_return" != "200" ];then
             echo "flashattn cache not found, please contact umiswing"
@@ -309,6 +311,7 @@ jobs:
           '
 
       - name: Test
+        id: unit_test
         run: |
           docker exec -t ${{ env.container_name }} /bin/bash -c '
           source ${{ github.workspace }}/../../../proxy
@@ -317,6 +320,15 @@ jobs:
           bash $ci_scripts/h-test.sh
           '
 
+      - name: FA Test
+        if: (success() || failure()) && steps.unit_test.conclusion != 'skipped'
+        run: |
+          docker exec -t ${{ env.container_name }} /bin/bash -c '
+          source ${{ github.workspace }}/../../../proxy
+          cd test/test_flashmask_ci
+          bash run.sh
+          '
+
       - name: Terminate and delete the container
         if: always()
         run: |

.github/workflows/rerun.yml

@@ -286,3 +286,13 @@ jobs:
           OWNER: ${{ github.repository_owner }}
           REPO: ${{ github.event.repository.name }}
           JOB_NAME: 'Doc-Preview / Check bypass / Check bypass'
+
+      - name: Rerun Slice
+        if: ${{ contains(github.event.comment.body, 'slice') }}
+        uses: ./.github/actions/rerun-workflow
+        with:
+          PR_ID: ${{ github.event.issue.number }}
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          OWNER: ${{ github.repository_owner }}
+          REPO: ${{ github.event.repository.name }}
+          JOB_NAME: 'Slice / Check bypass / Check bypass'

README.md

@@ -23,16 +23,9 @@ PaddlePaddle originates from industrial practices with dedication and commitment
 Our vision is to enable deep learning for everyone via PaddlePaddle.
 Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddle/releases) to track the latest features of PaddlePaddle.
 
-### Install Latest Stable Release
+### Install Latest Stable Release or Nightly Release
 
-``` sh
-# CPU
-pip install paddlepaddle
-# GPU
-pip install paddlepaddle-gpu
-```
-
-For more information about installation, please view [Quick Install](https://www.paddlepaddle.org.cn/install/quick)
+For detailed information about installation, please view [Quick Install](https://www.paddlepaddle.org.cn/install/quick)
 
 ## **PaddlePaddle New Generation Framework 3.2**
 

README_cn.md

@@ -22,16 +22,9 @@
 
 跟进 PaddlePaddle 最新特性请参考我们的[版本说明](https://github.com/PaddlePaddle/Paddle/releases)
 
-### 安装最新稳定版本
+### 安装最新稳定版本或每日最新构建版本
 
-``` sh
-# CPU
-pip install paddlepaddle
-# GPU
-pip install paddlepaddle-gpu
-```
-
-更多安装信息详见官网 [安装说明](https://www.paddlepaddle.org.cn/install/quick)。
+安装方法详见官网 [安装说明](https://www.paddlepaddle.org.cn/install/quick)。
 
 ## 飞桨新一代框架 3.2
 

README_ja.md

@@ -24,14 +24,7 @@ PaddlePaddle は、工業化に対するコミットメントを持つ工業的
 私たちのビジョンは、PaddlePaddle を通じて、誰もが深層学習を行えるようにすることです。
 PaddlePaddle の最新機能を追跡するために、私たちの[リリースのお知らせ](https://github.com/PaddlePaddle/Paddle/releases)を参照してください。
 
-### 最新の安定版リリースのインストール
-
-``` sh
-# CPU
-pip install paddlepaddle
-# GPU
-pip install paddlepaddle-gpu
-```
+### 最新の安定版またはナイトリービルドをインストールする
 
 インストール方法については、[クイックインストール](https://www.paddlepaddle.org.cn/install/quick)をご覧ください
 

cmake/external/glog.cmake

@@ -47,14 +47,19 @@ if(CMAKE_VERSION VERSION_GREATER_EQUAL "4.0.0")
   )
   set(GLOG_POLICY_ARGS "-DCMAKE_POLICY_VERSION_MINIMUM=3.5")
 endif()
-
+# patch
+file(TO_NATIVE_PATH ${PADDLE_SOURCE_DIR}/patches/glog/indent.patch log_indent)
+set(GLOG_PATCH_COMMAND git checkout -- . && git apply --check ${log_indent} &&
+                       git apply ${log_indent})
 ExternalProject_Add(
   extern_glog
   ${EXTERNAL_PROJECT_LOG_ARGS} ${SHALLOW_CLONE}
   SOURCE_DIR ${SOURCE_DIR}
   DEPENDS gflags
   PREFIX ${GLOG_PREFIX_DIR}
   UPDATE_COMMAND ""
+  PATCH_COMMAND
+  COMMAND ${GLOG_PATCH_COMMAND}
   CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
              -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
              -DCMAKE_CXX_FLAGS=${GLOG_CMAKE_CXX_FLAGS}

paddle/common/flags.cc

@@ -2314,3 +2314,56 @@ PHI_DEFINE_EXPORTED_bool(
 PHI_DEFINE_EXPORTED_bool(use_accuracy_compatible_kernel,
                          false,
                          "Whether use torch compatible version kernel.");
+
+/**
+ * Allocator Compact related FLAG
+ * Name: FLAGS_enable_compact_mem
+ * Since Version: 3.2.2
+ * Value Range: bool, default=false
+ * Example:
+ * Note: whether start compact memory.
+ */
+PHI_DEFINE_EXPORTED_bool(enable_compact_mem,
+                         false,
+                         "whether start compact memory or not.");
+/**
+ * Allocator Compact related FLAG
+ * Name: FLAGS_max_reserved_threshold_in_gb
+ * Since Version: 3.2.2
+ * Value Range: int64, default=70
+ * Example:
+ * Note: Threshold (GB) used in compact memory. Only reserved_mem greater than
+ * threshold may trigger defragmentation.
+ */
+PHI_DEFINE_EXPORTED_int64(
+    max_reserved_threshold_in_gb,
+    70,
+    "Threshold (GB) used in compact memory. Only reserved_mem greater than "
+    "threshold may trigger defragmentation.");
+
+/**
+ * Allocator Compact related FLAG
+ * Name: FLAGS_cur_allocated_threshold_in_gb
+ * Since Version: 3.2.2
+ * Value Range: int64, default=70
+ * Example:
+ * Note: Threshold (GB) used in compact memory. Only reserved_mem greater than
+ * threshold may trigger defragmentation.
+ */
+PHI_DEFINE_EXPORTED_int64(
+    cur_allocated_threshold_in_gb,
+    55,
+    "Threshold (GB) used in compact memory. Only reserved_mem greater than "
+    "threshold may trigger defragmentation.");
+
+/**
+ * Allocator Compact related FLAG
+ * Name: FLAGS_try_allocate
+ * Since Version: 3.2.2
+ * Value Range: bool, default=false
+ * Example:
+ * Note: whether start compact memory.
+ */
+PHI_DEFINE_EXPORTED_bool(try_allocate,
+                         false,
+                         "whether use try allocate in memory compact.");

paddle/fluid/eager/api/utils/global_utils.h

@@ -216,4 +216,52 @@ class EagerBackwardSubGraphNodeRecorder {
   bool need_capture_subgraph_ = false;
 };
 
+/**
+ * @class LogIndent
+ * @brief Singleton class for managing log indentation levels globally
+ *
+ * This class implements the singleton pattern to provide a centralized way
+ * to manage indentation levels for formatted log output. It ensures only
+ * one instance exists throughout the application lifecycle.
+ */
+class LogIndent {
+ public:
+  /**
+   * @brief Get the singleton instance of LogIndent
+   * @return Reference to the singleton instance
+   *
+   * Uses static local variable for thread-safe singleton initialization
+   * (C++11 guarantee). The instance is created on first call and destroyed
+   * automatically at program termination.
+   */
+  static LogIndent& Instance() {
+    static LogIndent instance;
+    return instance;
+  }
+  /**
+   * @brief Increase the current indentation level by 1
+   *
+   * Call this method when entering a nested scope to increase
+   * log indentation for better visual hierarchy.
+   */
+  void IncreaseIndentLevel() { FLAGS_indentlevel = FLAGS_indentlevel + 1; }
+  /**
+   * @brief Decrease the current indentation level by 1
+   *
+   * Reduces the indentation level, but never goes below 0.
+   * Call this when leaving a nested scope.
+   */
+  void DecreaseIndentLevel() {
+    if (FLAGS_indentlevel > 0) {
+      FLAGS_indentlevel = FLAGS_indentlevel - 1;
+    }
+  }
+  LogIndent(const LogIndent&) = delete;
+  LogIndent& operator=(const LogIndent&) = delete;
+
+ private:
+  LogIndent() = default;
+  ~LogIndent() = default;
+};
+
 }  // namespace egr

paddle/fluid/eager/auto_code_generator/generator/eager_gen.py

@@ -311,7 +311,6 @@
     "index_put",
     # others
     "matmul",
-    "expand",
 }
 
 strided_op_need_flags_check_list = {

paddle/fluid/eager/pylayer/py_layer_node.cc

@@ -45,6 +45,7 @@ GradNodePyLayer::operator()(
     egr::CUDAErrorCheck("GradNodePyLayer begin");
   }
   pybind11::gil_scoped_acquire gil;
+  if (VLOG_IS_ON(2)) egr::LogIndent::Instance().IncreaseIndentLevel();
   VLOG(3) << "Running Eager Backward Node: " << name();
   if (FLAGS_call_stack_level == 3) {
     VLOG(3) << "PyLayer forward call stack: " << this->GetForwardTrace();
@@ -256,6 +257,7 @@ GradNodePyLayer::operator()(
   Py_XDECREF(outputs);
   Py_XDECREF(ctx_);
   ctx_ = nullptr;
+  if (VLOG_IS_ON(2)) egr::LogIndent::Instance().DecreaseIndentLevel();
 
   if (FLAGS_check_cuda_error) [[unlikely]] {
     egr::CUDAErrorCheck("GradNodePyLayer finish");