paddle_musa v2.6.0 release initialization (#64265)

* Revert "fix rpc_sync and rpc_async doc;test=develop (#64107)"

This reverts commit 1319992.

* Revert "[Dy2St][2.6] Disable `test_sentiment` on release/2.6 (#63197)"

This reverts commit 9013831.

* Revert "Revert "fix security (#62626) (#62683)" (#62890)"

This reverts commit 89a60d7.

* Revert "Enhance several unit tests (#62477) (#62776)"

This reverts commit 0348f3f.

* Revert "[Fix_ci] set PLUGIN_TAG release/2.6 (#62731)"

This reverts commit 97ffa07.

* Revert "fix security (#62626) (#62683)"

This reverts commit 6a73547.

* Revert "add more capi to support stride (#62716)"

This reverts commit 683a141.

* Revert "[XPU] default no autotune (#62636)"

This reverts commit fde63d1.

* Revert "[DCU] fix dcu compile failure (#62573)"

This reverts commit d527fb5.

* Revert "[AutoParallel] Adjust time restriction for test_semi_auto_parallel_hybrid_strategy.py (#62278)"

This reverts commit fbf852d.

* Revert "disable llm_int8 ut (#62282)"

This reverts commit e816529.

* Revert "fix openssl-cpu compile bug (#62079) (#62224)"

This reverts commit 59c61db.

* Revert "[CINN] Add IntrinsicOps into ir_codes_collector (#60556) (#62245)"

This reverts commit 773ea41.

* Revert "rm graph_reindex_test (#62057)"

This reverts commit 521dc70.

* Revert "fix (#61923) (#62186)"

This reverts commit d077553.

* Revert "fix cpups training bug:executor trainer use_ps_gpu value;test=develop (#62111)"

This reverts commit d804975.

* Revert "[cherry-pick 2.6] Fix bug of put_along_axis/take_along_axis (#62065)"

This reverts commit 3a083c3.

* Revert "[Cherry-pick] Fix indexing shape bug and Optimize (#62117)"

This reverts commit 609f55e.

* Revert "cherry pick: reduce log for type promotion. (#62116)"

This reverts commit f4d9adf.

* Revert "fix test_communicator_half_async random core;test=develop (#62092)"

This reverts commit dba9992.

* Revert "fix the unqiue op that generate the wrong the inreverse result (#62104)"

This reverts commit b89066a.

* Revert "[Cherry-pick] Fix Paddle-TRT UT fails (#61605)"

This reverts commit 867ab0d.

* Revert "fix se (#61640) (#61702)"

This reverts commit c0f4a49.

* Revert "fix dataloaer for toolkit (#61867) (#61994)"

This reverts commit b50e906.

* Revert "[Cherry-Pick] Fix CacheKV Quant Bug (#61966)"

This reverts commit 04ac1c0.

* Revert "[Paddle-TRT] fix solve (#61806)"

This reverts commit df0155f.

* Revert "fix launch when elastic run (#61847) (#61878)"

This reverts commit f09d9d8.

* Revert "Support Fake GroupWise Quant (#61900)"

This reverts commit 2175de0.

* Revert "repeat_interleave support bf16 dtype (#61854) (#61899)"

This reverts commit 96c2aaf.

* Revert "[security] refine _get_program_cache_key (#61827) (#61896)"

This reverts commit b6a38d0.

* Revert "merge (#61866)"

This reverts commit 39010bf.

* Revert "fix doc style (#61688)"

This reverts commit 12e5c97.

* Revert "fix layer_norm decompose dtyte bugs, polish codes (#61631)"

This reverts commit e5a85b6.

* Revert "remove _wget (#61356) (#61569)"

This reverts commit 9250f66.

* Revert "cinn(py-dsl): skip eval string in python-dsl (#61380) (#61586)"

This reverts commit a37f6fb.

* Revert "Fix unique (#60840) (#61044)"

This reverts commit 3452e61.

* Revert "[CherryPick] Fix issue 60092 (#61427)"

This reverts commit f025385.

* Revert "[cherry-pick] adapt c_embedding to phi namespace for custom devices (#60774) (#61045)"

This reverts commit 0ccb9cb.

* Revert "check eval for security (#61389)"

This reverts commit 60325a1.

* Revert "[Security] fix download security problem (#61162) (#61388)"

This reverts commit 5f3bbeb.

* Revert "[Security] fix security problem for run_cmd (#61285) (#61398)"

This reverts commit 9cd0c91.

* Revert "[Security] fix security problem for prune_by_memory_estimation (#61382)"

This reverts commit af9b8c5.

* Revert "Fix CVE-2024-0521 (#61032) (#61287)"

This reverts commit f99d4f2.

* Revert "fix _decompress security problem (#61294) (#61337)"

This reverts commit 0227a0d.

* Revert "[Security] fix draw security problem (#61161) (#61338)"

This reverts commit aeaa0ca.

* Revert "fix qat tests (#61211) (#61284)"

This reverts commit ff119d0.

* Revert "fix core dump when fallback gather_nd_grad and MemoryAllocateHost (#61067)"

This reverts commit ac1702b.

* Revert "[cherry-pick] This PR enable offset of generator for custom device. (#60616) (#60772)"

This reverts commit 0f732a5.

* Revert "[Cherry-pick] fix set_value with scalar grad (#60930)"

This reverts commit 1aa5f4b.

* Revert "[Dy2St][2.6] Increase `test_transformer` and `test_mobile_net` ut time (#60829) (#60875)"

This reverts commit d788e9b.

* Revert "[Dy2St][2.6] Disable `test_transformer` on `release/2.6` and update README (#60786)"

This reverts commit e738f49.

* Revert "fix bug of ci (#59926) (#60785)"

This reverts commit 7b0d2e9.

* Revert "[Dy2St][2.6] Disable `test_grad` on release/2.6 (#60662)"

This reverts commit e50f43e.

* Revert "[cherry-pick]update pdsa-2023-019 (#60649)"

This reverts commit ccdf528.

* Revert "[cherry-pick]fix fleetutil get_online_pass_interval bug3 (#60620)"

This reverts commit bbc13eb.

* Revert "fix fused_rope diff (#60217) (#60593)"

This reverts commit 97b65c7.

* Revert "fix fleetutil get_online_pass_interval bug2; test=develop (#60545)"

This reverts commit ae2e588.

* Revert "update 2023 security advisory, test=document_fix (#60532)"

This reverts commit 83ce809.

* Revert "add chunk allocator posix_memalign return value check (#60208) (#60495)"

This reverts commit b065877.

* Revert "tile (#60261)"

This reverts commit 203754e.

* Revert "[Cherry-pick] fix weight quant kernel bug when n div 64 != 0 (#60184)"

This reverts commit 20d3558.

* Revert "[Dy2St] Disable `test_bert` on CPU (#60173) (#60324)"

This reverts commit a4cd847.

* Revert "fix windows bug for common lib (#60308)"

This reverts commit 1b696a1.

* update to v2.6.0

* enable WITH_DISTRIBUTED in CMakeLists.txt and port related source file from cuda to musa

* fix some bugs when WITH_DISTRIBUTED is enabled

* delete useless cout in ../paddle/phi/backends/gpu/musa/musa_info.cc and set compute capacity to 9.9 for UT

.gitmodules

@@ -110,11 +110,3 @@
 	path = third_party/cccl
 	url = https://github.com/NVIDIA/cccl.git
 	ignore = dirty
-[submodule "third_party/cryptopp"]
-	path = third_party/cryptopp
-	url = https://github.com/weidai11/cryptopp.git
-	ignore = dirty
-[submodule "third_party/cryptopp-cmake"]
-	path = third_party/cryptopp-cmake
-	url = https://github.com/noloader/cryptopp-cmake.git
-	ignore = dirty

CMakeLists.txt

@@ -41,13 +41,14 @@ if(NOT CMAKE_BUILD_TYPE)
 endif()
 
 project(paddle CXX C)
-
+# set(CMAKE_VERBOSE_MAKEFILE ON)
 # enable language CUDA
 # TODO(Shibo Tao): remove find_package(CUDA) completely.
 find_package(CUDA QUIET)
 find_package(MKL CONFIG QUIET)
 option(WITH_ONEMKL "Compile PaddlePaddle with oneMKL" OFF)
-option(WITH_GPU "Compile PaddlePaddle with NVIDIA GPU" ${CUDA_FOUND})
+option(WITH_GPU "Compile PaddlePaddle with NVIDIA GPU" OFF)
+option(WITH_MUSA "Compile PaddlePaddle with MUSA" ON)
 option(WITH_MPI "Compile PaddlePaddle with MPI" OFF)
 option(WITH_TENSORRT "Compile PaddlePaddle with NVIDIA TensorRT" OFF)
 option(WITH_XPU "Compile PaddlePaddle with BAIDU KUNLUN XPU" OFF)
@@ -89,6 +90,9 @@ endif()
 if(WITH_GPU AND WITH_ROCM)
   message(FATAL_ERROR "Error when compile CUDA and ROCM at the same time")
 endif()
+if(WITH_GPU AND WITH_MUSA)
+  message(FATAL_ERROR "Error when compile CUDA and MUSA at the same time")
+endif()
 
 if(WITH_GPU AND NOT APPLE)
   enable_language(CUDA)
@@ -252,7 +256,7 @@ option(WITH_TESTING "Compile PaddlePaddle with unit testing" OFF)
 option(WITH_MULTINODE_TESTING "Test multinode apis and ops" OFF)
 option(WITH_MKL "Compile PaddlePaddle with MKL support." ${AVX_FOUND})
 option(WITH_SYSTEM_BLAS "Use system blas library" OFF)
-option(WITH_DISTRIBUTE "Compile with distributed support" OFF)
+option(WITH_DISTRIBUTE "Compile with distributed support" ON)
 option(WITH_BRPC_RDMA "Use brpc rdma as the rpc protocal" OFF)
 option(ON_INFER "Turn on inference optimization and inference-lib generation"
        ON)
@@ -285,6 +289,7 @@ option(WITH_LITE "Compile Paddle Fluid with Lite Engine" OFF)
 option(WITH_CINN "Compile PaddlePaddle with CINN" OFF)
 option(WITH_NCCL "Compile PaddlePaddle with NCCL support" ON)
 option(WITH_RCCL "Compile PaddlePaddle with RCCL support" ON)
+option(WITH_MCCL "Compile PaddlePaddle with MCCL support" ON)
 option(WITH_XPU_BKCL "Compile PaddlePaddle with BAIDU KUNLUN XPU BKCL" OFF)
 option(WITH_CRYPTO "Compile PaddlePaddle with crypto support" ON)
 option(WITH_ARM "Compile PaddlePaddle with arm support" OFF)
@@ -352,6 +357,7 @@ endif()
 if(LINUX
    AND NOT WITH_CUSTOM_DEVICE
    AND NOT WITH_GPU
+   AND NOT WITH_MUSA
    AND NOT WITH_ROCM
    AND NOT WITH_XPU
    AND NOT WITH_XPU_KP
@@ -404,6 +410,14 @@ if(NOT WITH_GPU AND WITH_NCCL)
       CACHE STRING "Disable NCCL when compiling without GPU" FORCE)
 endif()
 
+if(NOT WITH_MUSA AND WITH_MCCL)
+  message(
+    WARNING "Disable MCCL when compiling without MUSA. Force WITH_MCCL=OFF.")
+  set(WITH_MCCL
+      OFF
+      CACHE STRING "Disable MCCL when compiling without MUSA" FORCE)
+endif()
+
 if(NOT WITH_GPU AND WITH_CUDNN_DSO)
   message(
     WARNING
@@ -461,6 +475,19 @@ else()
   endif()
 endif()
 
+if(WITH_MCCL)
+  add_definitions("-DPADDLE_WITH_MCCL")
+  include(mccl)
+else()
+  if(WITH_MUSA)
+    message(
+      WARNING
+        "If the environment is multi-card, the WITH_MCCL option needs to be turned on, otherwise only a single card can be used."
+    )
+  endif()
+endif()
+
+
 if(WITH_BRPC_RDMA)
   message(STATUS "Use brpc with rdma.")
   if(NOT WITH_DISTRIBUTE)
@@ -486,6 +513,11 @@ if(WITH_ROCM)
   include(cupti)
 endif()
 
+if(WITH_MUSA)
+  include(musa)
+  include(mudnn)
+endif()
+
 if(WITH_XPU_KP)
   include(xpu_kp)
 endif()
@@ -498,6 +530,14 @@ if(NOT WITH_ROCM AND WITH_RCCL)
       CACHE STRING "Disable RCCL when compiling without ROCM" FORCE)
 endif()
 
+if(NOT WITH_MUSA AND WITH_MCCL)
+  message(
+    WARNING "Disable MCCL when compiling without MUSA. Force WITH_MCCL=OFF.")
+  set(WITH_MCCL
+      OFF
+      CACHE STRING "Disable MCCL when compiling without MUSA" FORCE)
+endif()
+
 if(WITH_RCCL)
   add_definitions("-DPADDLE_WITH_RCCL")
   include(rccl)
@@ -510,6 +550,18 @@ else()
   endif()
 endif()
 
+if(WITH_MCCL)
+  add_definitions("-DPADDLE_WITH_MCCL")
+  include(mccl)
+else()
+  if(WITH_MUSA)
+    message(
+      WARNING
+        "If the environment is multi-card, the WITH_MCCL option needs to be turned on, otherwise only a single card can be used."
+    )
+  endif()
+endif()
+
 if(WITH_HETERPS AND WITH_PSLIB)
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -D_GLIBCXX_USE_CXX11_ABI=0")
 endif()
@@ -560,6 +612,13 @@ if(WITH_RPC)
         OFF
         CACHE BOOL "Disable WITH_RPC when compiling with ROCM" FORCE)
   endif()
+  if(WITH_MUSA AND WITH_RPC)
+  message(
+    WARNING "Disable WITH_RPC when compiling with MUSA. Force WITH_RPC=OFF.")
+  set(WITH_RPC
+      OFF
+      CACHE BOOL "Disable WITH_RPC when compiling with MUSA" FORCE)
+  endif()
   if(WITH_XPU AND WITH_RPC)
     message(
       WARNING "Disable WITH_RPC when compiling with XPU. Force WITH_RPC=OFF.")
@@ -631,6 +690,12 @@ include(configure) # add paddle env configuration
 
 include_directories("${PADDLE_SOURCE_DIR}")
 
+# distribute need openssl
+# openssl install tutorial: https://www.howtoforge.com/tutorial/how-to-install-openssl-from-source-on-linux/
+include_directories("/usr/local/ssl/include")
+link_directories("/usr/local/ssl/lib64")
+
+
 if(WITH_NV_JETSON)
   set(WITH_ARM
       ON

README.md

@@ -20,7 +20,7 @@ PaddlePaddle is originated from industrial practices with dedication and commitm
 
 ## Installation
 
-### Latest PaddlePaddle Release: [v2.6](https://github.com/PaddlePaddle/Paddle/tree/release/2.6)
+### Latest PaddlePaddle Release: [v2.5](https://github.com/PaddlePaddle/Paddle/tree/release/2.5)
 
 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.

README_cn.md

@@ -18,9 +18,9 @@
 
 ## 安装
 
-### PaddlePaddle 最新版本: [v2.6](https://github.com/PaddlePaddle/Paddle/tree/release/2.6)
+### PaddlePaddle最新版本: [v2.5](https://github.com/PaddlePaddle/Paddle/tree/release/2.5)
 
-跟进 PaddlePaddle 最新特性请参考我们的[版本说明](https://github.com/PaddlePaddle/Paddle/releases)
+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:
 ```

README_ja.md

@@ -20,7 +20,7 @@ PaddlePaddle は、工業化に対するコミットメントを持つ工業的
 
 ## インストール
 
-### PaddlePaddle の最新リリース: [v2.6](https://github.com/PaddlePaddle/Paddle/tree/release/2.6)
+### PaddlePaddle の最新リリース: [v2.5](https://github.com/PaddlePaddle/Paddle/tree/release/2.5)
 
 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。

cmake/configure.cmake

@@ -175,6 +175,19 @@ elseif(WITH_ROCM)
   if(${MIOPEN_VERSION} VERSION_LESS 2090)
     message(FATAL_ERROR "Paddle needs MIOPEN >= 2.9 to compile")
   endif()
+elseif(WITH_MUSA)
+  add_definitions(-DPADDLE_WITH_MUSA)
+  add_definitions(-DEIGEN_USE_GPU)
+  add_definitions(-DEIGEN_USE_MUSA)
+  if(MUPTI_FOUND)
+    include_directories(${CUPTI_INCLUDE_DIR})
+    add_definitions(-DPADDLE_WITH_MUPTI)
+  else()
+    message(STATUS "Cannot find MUPTI, GPU Profiling is incorrect.")
+  endif()
+  if(NOT MUDNN_FOUND)
+    message(FATAL_ERROR "Paddle needs mudnn to compile")
+  endif()
 else()
   add_definitions(-DHPPL_STUB_FUNC)
   list(APPEND CMAKE_CXX_SOURCE_FILE_EXTENSIONS cu)

cmake/cupti.cmake

@@ -1,11 +1,15 @@
-if(NOT WITH_GPU AND NOT WITH_ROCM)
+if(NOT WITH_GPU AND NOT WITH_ROCM AND NOT WITH_MUSA)
   return()
 endif()
 
 if(WITH_ROCM)
   set(CUPTI_ROOT
       "${ROCM_PATH}/cuda/extras/CUPTI"
       CACHE PATH "CUPTI ROOT")
+elseif(WITH_MUSA)
+  set(CUPTI_ROOT
+      "/usr/local/musa"
+      CACHE PATH "CUPTI ROOT")      
 else()
   set(CUPTI_ROOT
       "/usr"

cmake/external/cryptopp.cmake

@@ -14,13 +14,12 @@
 
 include(ExternalProject)
 
-set(CRYPTOPP_SOURCE_DIR ${PADDLE_SOURCE_DIR}/third_party/cryptopp)
-set(CRYPTOPP_CMAKE_SOURCE_DIR ${PADDLE_SOURCE_DIR}/third_party/cryptopp-cmake)
 set(CRYPTOPP_PREFIX_DIR ${THIRD_PARTY_PATH}/cryptopp)
 set(CRYPTOPP_INSTALL_DIR ${THIRD_PARTY_PATH}/install/cryptopp)
 set(CRYPTOPP_INCLUDE_DIR
     "${CRYPTOPP_INSTALL_DIR}/include"
     CACHE PATH "cryptopp include directory." FORCE)
+set(CRYPTOPP_REPOSITORY ${GIT_URL}/weidai11/cryptopp.git)
 set(CRYPTOPP_TAG CRYPTOPP_8_2_0)
 
 if(WIN32)
@@ -64,16 +63,17 @@ include_directories(${CRYPTOPP_INCLUDE_DIR})
 ExternalProject_Add(
   extern_cryptopp
   ${EXTERNAL_PROJECT_LOG_ARGS} ${SHALLOW_CLONE}
+  GIT_REPOSITORY ${CRYPTOPP_REPOSITORY}
+  GIT_TAG ${CRYPTOPP_TAG}
   PREFIX ${CRYPTOPP_PREFIX_DIR}
-  SOURCE_DIR ${CRYPTOPP_SOURCE_DIR}
   UPDATE_COMMAND ""
   PATCH_COMMAND
-  COMMAND ${CMAKE_COMMAND} -E copy "${CRYPTOPP_CMAKE_SOURCE_DIR}/CMakeLists.txt"
-          "<SOURCE_DIR>/CMakeLists.txt"
-  COMMAND
-    ${CMAKE_COMMAND} -E copy
-    "${CRYPTOPP_CMAKE_SOURCE_DIR}/cryptopp-config.cmake"
-    "<SOURCE_DIR>/cryptopp-config.cmake"
+  COMMAND ${CMAKE_COMMAND} -E remove_directory "<SOURCE_DIR>/cmake/"
+  COMMAND git clone ${GIT_URL}/noloader/cryptopp-cmake "<SOURCE_DIR>/cmake"
+  COMMAND cd "<SOURCE_DIR>/cmake" && git checkout tags/${CRYPTOPP_TAG} -b
+          ${CRYPTOPP_TAG}
+  COMMAND ${CMAKE_COMMAND} -E copy_directory "<SOURCE_DIR>/cmake/"
+          "<SOURCE_DIR>/"
   COMMAND ${CRYPTOPP_PATCH_COMMAND}
   INSTALL_DIR ${CRYPTOPP_INSTALL_DIR}
   CMAKE_ARGS ${CRYPTOPP_CMAKE_ARGS}

cmake/external/eigen.cmake

@@ -60,6 +60,76 @@ if(CMAKE_COMPILER_IS_GNUCC)
         ${EIGEN_PATCH_COMMAND} && patch -Nd
         ${SOURCE_DIR}/Eigen/src/Core/arch/SSE/ < ${complex_header})
   endif()
+  if(WITH_MUSA)
+    file(
+      TO_NATIVE_PATH
+      ${PADDLE_SOURCE_DIR}/patches/eigen/Eigen_src_Core_util_ConfigureVectorization.h.patch
+      configure_vectorization_header)
+    set(EIGEN_PATCH_COMMAND
+        ${EIGEN_PATCH_COMMAND} && patch -Nd ${SOURCE_DIR}/Eigen/src/Core/util/
+        < ${configure_vectorization_header})
+    file(TO_NATIVE_PATH
+         ${PADDLE_SOURCE_DIR}/patches/eigen/Eigen_src_Core_util_Macros.h.patch
+         util_macros_header)
+    set(EIGEN_PATCH_COMMAND
+        ${EIGEN_PATCH_COMMAND} && patch -Nd ${SOURCE_DIR}/Eigen/src/Core/util/
+        < ${util_macros_header})
+    file(TO_NATIVE_PATH
+         ${PADDLE_SOURCE_DIR}/patches/eigen/Eigen_src_Core_util_Meta.h.patch
+         meta_header)
+    set(EIGEN_PATCH_COMMAND ${EIGEN_PATCH_COMMAND} && patch -Nd
+                            ${SOURCE_DIR}/Eigen/src/Core/util/ < ${meta_header})
+    file(TO_NATIVE_PATH
+         ${PADDLE_SOURCE_DIR}/patches/eigen/unsupported_Eigen_CXX11_Tensor.patch
+         cxx11_tensor)
+    set(EIGEN_PATCH_COMMAND
+        ${EIGEN_PATCH_COMMAND} && patch -Nd
+        ${SOURCE_DIR}/unsupported/Eigen/CXX11/ < ${cxx11_tensor})
+    file(
+      TO_NATIVE_PATH
+      ${PADDLE_SOURCE_DIR}/patches/eigen/unsupported_Eigen_CXX11_src_Tensor_TensorContractionGpu.h.patch
+      tensor_contraction_gpu_header)
+    set(EIGEN_PATCH_COMMAND
+        ${EIGEN_PATCH_COMMAND} && patch -Nd
+        ${SOURCE_DIR}/unsupported/Eigen/CXX11/src/Tensor/ <
+        ${tensor_contraction_gpu_header})
+    file(
+      TO_NATIVE_PATH
+      ${PADDLE_SOURCE_DIR}/patches/eigen/unsupported_Eigen_CXX11_src_Tensor_TensorDeviceDefault.h.patch
+      tensor_device_default_header)
+    set(EIGEN_PATCH_COMMAND
+        ${EIGEN_PATCH_COMMAND} && patch -Nd
+        ${SOURCE_DIR}/unsupported/Eigen/CXX11/src/Tensor/ <
+        ${tensor_device_default_header})
+    file(
+      TO_NATIVE_PATH
+      ${PADDLE_SOURCE_DIR}/patches/eigen/unsupported_Eigen_CXX11_src_Tensor_TensorGpuHipCudaDefines.h.patch
+      tensor_gpu_hip_cuda_defines_header)
+    set(EIGEN_PATCH_COMMAND
+        ${EIGEN_PATCH_COMMAND} && patch -Nd
+        ${SOURCE_DIR}/unsupported/Eigen/CXX11/src/Tensor/ <
+        ${tensor_gpu_hip_cuda_defines_header})
+    file(
+      TO_NATIVE_PATH
+      ${PADDLE_SOURCE_DIR}/patches/eigen/unsupported_Eigen_CXX11_src_Tensor_TensorReduction.h.patch
+      tensor_reduction_header)
+    set(EIGEN_PATCH_COMMAND
+        ${EIGEN_PATCH_COMMAND} && patch -Nd
+        ${SOURCE_DIR}/unsupported/Eigen/CXX11/src/Tensor/ <
+        ${tensor_reduction_header})
+    file(TO_NATIVE_PATH ${PADDLE_SOURCE_DIR}/patches/eigen/Eigen_CORE.patch
+         eigen_core)
+    set(EIGEN_PATCH_COMMAND ${EIGEN_PATCH_COMMAND} && patch -Nd
+                            ${SOURCE_DIR}/Eigen/ < ${eigen_core})
+    file(
+      TO_NATIVE_PATH
+      ${PADDLE_SOURCE_DIR}/patches/eigen/unsupported_Eigen_CXX11_src_Tensor_TensorDeviceGpu.h.patch
+      tensor_device_gpu_header)
+    set(EIGEN_PATCH_COMMAND
+        ${EIGEN_PATCH_COMMAND} && patch -Nd
+        ${SOURCE_DIR}/unsupported/Eigen/CXX11/src/Tensor/ <
+        ${tensor_device_gpu_header})
+  endif()
 endif()
 
 set(EIGEN_INCLUDE_DIR ${SOURCE_DIR})

cmake/flags.cmake

@@ -246,6 +246,11 @@ if(WITH_GPU)
   set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} ${SAFE_GPU_COMMON_FLAGS}")
 endif()
 
+if(WITH_MUSA)
+  set(CMAKE_MUSA_FLAGS "${CMAKE_MUSA_FLAGS} ${SAFE_GPU_COMMON_FLAGS}")
+endif()
+
+
 if(WITH_ROCM)
   set(HIP_HIPCC_FLAGS "${HIP_HIPCC_FLAGS} ${SAFE_GPU_COMMON_FLAGS}")
 endif()