diff --git a/.github/workflows/metax_work.yaml b/.github/workflows/metax_work.yaml
index f73442b6fd5..fd7d04c0843 100644
--- a/.github/workflows/metax_work.yaml
+++ b/.github/workflows/metax_work.yaml
@@ -40,7 +40,7 @@ jobs:
             git fetch origin pull/${{ github.event.pull_request.number }}/head:pull/${{ github.event.pull_request.number }}/head
             git checkout pull/${{ github.event.pull_request.number }}/head
 
-            # git submodule update --init --recursive
+            git submodule update --init --recursive
           fi
 
 
diff --git a/Paddle b/Paddle
index 2588f489910..47699dd459f 160000
--- a/Paddle
+++ b/Paddle
@@ -1 +1 @@
-Subproject commit 2588f4899106cd27bdfcc84ba4c2f5f7aac570ab
+Subproject commit 47699dd459fdc8e525beac030d5c939b42128057
diff --git a/backends/metax_gpu/CMakeLists.txt b/backends/metax_gpu/CMakeLists.txt
index e357a5e5912..3e92996f9a2 100755
--- a/backends/metax_gpu/CMakeLists.txt
+++ b/backends/metax_gpu/CMakeLists.txt
@@ -703,9 +703,9 @@ file(
 set(CUSTOM_DEVICE_SRCS ${CUDA_SRCS} ${CC_SRCS} ${ERNIE_CORE_SRCS})
 
 set_source_files_properties(${CUSTOM_DEVICE_SRCS} PROPERTIES LANGUAGE CUDA)
-
+set(MACA_PATH $ENV{MACA_PATH})
 set(CMAKE_CUCC_COMPILER "cucc")
-set(CMAKE_CUCC_FLAGS "-I /opt/maca/tools/cu-bridge/include/")
+set(CMAKE_CUCC_FLAGS "-I ${MACA_PATH}/tools/cu-bridge/include/")
 
 add_library(${TARGET_NAME} SHARED ${CUSTOM_DEVICE_SRCS})
 
@@ -734,9 +734,9 @@ target_link_libraries(
   ${WARPRNNT_LIBRARIES}
   ${PADDLE_CORE_LIB})
 
-target_link_libraries(${TARGET_NAME} /opt/maca/lib/libmccl.so)
-target_link_libraries(${TARGET_NAME} /opt/maca/lib/libmcFlashAttn.so)
-target_link_libraries(${TARGET_NAME} /opt/maca/lib/libmcpti.so)
+target_link_libraries(${TARGET_NAME} ${MACA_PATH}/lib/libmccl.so)
+target_link_libraries(${TARGET_NAME} ${MACA_PATH}/lib/libmcFlashAttn.so)
+target_link_libraries(${TARGET_NAME} ${MACA_PATH}/lib/libmcpti.so)
 
 include_directories(BEFORE ${PADDLE_SOURCE_DIR})
 
diff --git a/backends/metax_gpu/cmake/dgc.cmake b/backends/metax_gpu/cmake/dgc.cmake
index 4c54e636d5e..4c61f2e6bcb 100644
--- a/backends/metax_gpu/cmake/dgc.cmake
+++ b/backends/metax_gpu/cmake/dgc.cmake
@@ -62,8 +62,8 @@ if(EXISTS ${DGC_DOWNLOAD_DIR}/${DGC_CACHE_FILENAME})
 else()
   download_dgc()
 endif()
-
-set(CU_BRIDGE_PATH "/opt/maca/tools/cu-bridge")
+set(MACA_PATH $ENV{MACA_PATH})
+set(CU_BRIDGE_PATH "${MACA_PATH}/tools/cu-bridge")
 
 add_custom_command(
   OUTPUT "${CU_BRIDGE_PATH}/bin/nvcc"
diff --git a/backends/metax_gpu/kernels/cuda_kernels/activation_grad_kernel_register.cu b/backends/metax_gpu/kernels/cuda_kernels/activation_grad_kernel_register.cu
index 6cdfb2f5242..6c46ef10c0f 100644
--- a/backends/metax_gpu/kernels/cuda_kernels/activation_grad_kernel_register.cu
+++ b/backends/metax_gpu/kernels/cuda_kernels/activation_grad_kernel_register.cu
@@ -119,7 +119,22 @@ void ActivationGradGPUImpl(const Context& dev_ctx,
     ActivationGradGPUImpl<T, Context, funcs::functor_class<T>>( \
         dev_ctx, &x, nullptr, &dout, dx, functor);              \
   }
-
+#define DEFINE_GPU_ACT_GRAD_KERNEL_WITH_TWO_DOUBLE_ATTRS_DEPX(  \
+    name, functor_class, attr1, attr2)                          \
+  template <typename T, typename Context>                       \
+  void name##GradKernel(const Context& dev_ctx,                 \
+                        const DenseTensor& x,                   \
+                        const DenseTensor& dout,                \
+                        double attr1,                           \
+                        double attr2,                           \
+                        DenseTensor* dx) {                      \
+    funcs::functor_class<T> functor;                            \
+    auto attrs = functor.GetAttrs();                            \
+    *(attrs[0].second) = attr1;                                 \
+    *(attrs[1].second) = attr2;                                 \
+    ActivationGradGPUImpl<T, Context, funcs::functor_class<T>>( \
+        dev_ctx, &x, nullptr, &dout, dx, functor);              \
+  }
 #define DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPOUT(name, functor_class) \
   template <typename T, typename Context>                             \
   void name##GradKernel(const Context& dev_ctx,                       \
@@ -239,10 +254,10 @@ DEFINE_GPU_ACT_GRAD_KERNEL_WITH_TWO_ATTRS_DEPX(STanh,
                                                scale_a,
                                                scale_b);
 
-DEFINE_GPU_ACT_GRAD_KERNEL_WITH_TWO_ATTRS_DEPX(Softplus,
-                                               CudaSoftplusGradFunctor,
-                                               beta,
-                                               threshold);
+DEFINE_GPU_ACT_GRAD_KERNEL_WITH_TWO_DOUBLE_ATTRS_DEPX(Softplus,
+                                                      CudaSoftplusGradFunctor,
+                                                      beta,
+                                                      threshold);
 DEFINE_GPU_ACT_GRAD_KERNEL_WITH_TWO_ATTRS_DEPOUT(HardSigmoid,
                                                  CudaHardSigmoidGradFunctor,
                                                  slope,
diff --git a/backends/metax_gpu/kernels/cuda_kernels/activation_kernel_register.cu b/backends/metax_gpu/kernels/cuda_kernels/activation_kernel_register.cu
index f24f3e8abbc..363932cfc28 100644
--- a/backends/metax_gpu/kernels/cuda_kernels/activation_kernel_register.cu
+++ b/backends/metax_gpu/kernels/cuda_kernels/activation_kernel_register.cu
@@ -90,7 +90,21 @@ void ActivationGPUImpl(const Context& dev_ctx,
     ActivationGPUImpl<T, Context, funcs::functor_class<T>>( \
         dev_ctx, x, out, functor);                          \
   }
-
+#define DEFINE_GPU_ACT_KERNEL_WITH_TWO_DOUBLE_ATTRS(        \
+    name, functor_class, attr1, attr2)                      \
+  template <typename T, typename Context>                   \
+  void name##Kernel(const Context& dev_ctx,                 \
+                    const DenseTensor& x,                   \
+                    double attr1,                           \
+                    double attr2,                           \
+                    DenseTensor* out) {                     \
+    funcs::functor_class<T> functor;                        \
+    auto attrs = functor.GetAttrs();                        \
+    *(attrs[0].second) = attr1;                             \
+    *(attrs[1].second) = attr2;                             \
+    ActivationGPUImpl<T, Context, funcs::functor_class<T>>( \
+        dev_ctx, x, out, functor);                          \
+  }
 DEFINE_GPU_ACTIVATION_KERNEL(Cos, CudaCosFunctor)
 DEFINE_GPU_ACTIVATION_KERNEL(Tan, CudaTanFunctor)
 DEFINE_GPU_ACTIVATION_KERNEL(Acos, CudaAcosFunctor)
@@ -139,10 +153,10 @@ DEFINE_GPU_ACT_KERNEL_WITH_TWO_ATTRS(HardTanh,
                                      t_min,
                                      t_max)
 DEFINE_GPU_ACT_KERNEL_WITH_TWO_ATTRS(Stanh, CudaSTanhFunctor, scale_a, scale_b)
-DEFINE_GPU_ACT_KERNEL_WITH_TWO_ATTRS(Softplus,
-                                     CudaSoftplusFunctor,
-                                     beta,
-                                     threshold)
+DEFINE_GPU_ACT_KERNEL_WITH_TWO_DOUBLE_ATTRS(Softplus,
+                                            CudaSoftplusFunctor,
+                                            beta,
+                                            threshold)
 DEFINE_GPU_ACT_KERNEL_WITH_TWO_ATTRS(HardSigmoid,
                                      CudaHardSigmoidFunctor,
                                      slope,
diff --git a/backends/metax_gpu/kernels/funcs/blas/blas.h b/backends/metax_gpu/kernels/funcs/blas/blas.h
index fa4b4643f89..75ea8c921e2 100644
--- a/backends/metax_gpu/kernels/funcs/blas/blas.h
+++ b/backends/metax_gpu/kernels/funcs/blas/blas.h
@@ -282,6 +282,9 @@ class Blas {
   template <typename T>
   T DOT(int n, const T* x, const T* y) const;
 
+  template <typename T>
+  void CUDOT(
+      int n, const T* x, int incx, const T* y, int incy, T* result) const;
   template <typename T>
   void SCAL(int n, const T a, T* x) const;
 
@@ -541,7 +544,10 @@ class BlasT : private Blas<DeviceContext> {
   T DOT(ARGS... args) const {
     return Base()->template DOT<T>(args...);
   }
-
+  template <typename... ARGS>
+  void CUDOT(ARGS... args) const {
+    Base()->template CUDOT<T>(args...);
+  }
   template <typename... ARGS>
   void SCAL(ARGS... args) const {
     Base()->template SCAL<T>(args...);
diff --git a/backends/metax_gpu/patch/paddle.patch b/backends/metax_gpu/patch/paddle.patch
index f2e4f067bb2..7ba32b5b399 100755
--- a/backends/metax_gpu/patch/paddle.patch
+++ b/backends/metax_gpu/patch/paddle.patch
@@ -942,6 +942,19 @@ index f0cca0f701..02ea957240 100644
  
  namespace phi {
  // To determine use cudnn or not.
+diff --git a/paddle/phi/kernels/gpu/dot_kernel.cu b/paddle/phi/kernels/gpu/dot_kernel.cu
+index af27ac89ab..ee0edc6b8e 100644
+--- a/paddle/phi/kernels/gpu/dot_kernel.cu
++++ b/paddle/phi/kernels/gpu/dot_kernel.cu
+@@ -15,7 +15,7 @@
+ #include "paddle/phi/kernels/dot_kernel.h"
+ #include "paddle/phi/backends/gpu/gpu_context.h"
+ #include "paddle/phi/core/kernel_registry.h"
+-#include "paddle/phi/kernels/funcs/blas/blas.h"
++#include "kernels/funcs/blas/blas.h"
+ #include "paddle/phi/kernels/funcs/eigen/common.h"
+ 
+ #include "paddle/phi/kernels/full_kernel.h"
 diff --git a/paddle/phi/kernels/gpu/gelu_funcs.h b/paddle/phi/kernels/gpu/gelu_funcs.h
 index 29fa252e96..4ae72b0935 100644
 --- a/paddle/phi/kernels/gpu/gelu_funcs.h