PaddlePaddle
diff --git a/‎cmake/external/xpu.cmake
Copy file name to clipboard
+61-31Lines changed: 61 additions & 31 deletions b/‎cmake/external/xpu.cmake
Copy file name to clipboard
+61-31Lines changed: 61 additions & 31 deletions
diff --git a/‎paddle/phi/backends/xpu/xpu3_op_list.cc
Copy file name to clipboard
-1Lines changed: 0 additions & 1 deletion b/‎paddle/phi/backends/xpu/xpu3_op_list.cc
Copy file name to clipboard
-1Lines changed: 0 additions & 1 deletion
diff --git a/‎paddle/phi/kernels/fusion/xpu/fused_multi_transformer_int8_xpu_kernel.cc
Copy file name to clipboard
+2-2Lines changed: 2 additions & 2 deletions b/‎paddle/phi/kernels/fusion/xpu/fused_multi_transformer_int8_xpu_kernel.cc
Copy file name to clipboard
+2-2Lines changed: 2 additions & 2 deletions
diff --git a/‎paddle/phi/kernels/fusion/xpu/fused_multi_transformer_xpu_kernel.cc
Copy file name to clipboard
+4-4Lines changed: 4 additions & 4 deletions b/‎paddle/phi/kernels/fusion/xpu/fused_multi_transformer_xpu_kernel.cc
Copy file name to clipboard
+4-4Lines changed: 4 additions & 4 deletions
diff --git a/‎paddle/phi/kernels/fusion/xpu/fused_rope_utils.h
Copy file name to clipboard
+15-14Lines changed: 15 additions & 14 deletions b/‎paddle/phi/kernels/fusion/xpu/fused_rope_utils.h
Copy file name to clipboard
+15-14Lines changed: 15 additions & 14 deletions
diff --git a/‎paddle/phi/kernels/xpu/addmm_kernel.cc
Copy file name to clipboard
+4-4Lines changed: 4 additions & 4 deletions b/‎paddle/phi/kernels/xpu/addmm_kernel.cc
Copy file name to clipboard
+4-4Lines changed: 4 additions & 4 deletions
diff --git a/‎paddle/phi/kernels/xpu/c_embedding_kernel.cc
Copy file name to clipboard
+18-18Lines changed: 18 additions & 18 deletions b/‎paddle/phi/kernels/xpu/c_embedding_kernel.cc
Copy file name to clipboard
+18-18Lines changed: 18 additions & 18 deletions
diff --git a/‎paddle/phi/kernels/xpu/distribute_fpn_proposals_kernel.cc
Copy file name to clipboard
+2-2Lines changed: 2 additions & 2 deletions b/‎paddle/phi/kernels/xpu/distribute_fpn_proposals_kernel.cc
Copy file name to clipboard
+2-2Lines changed: 2 additions & 2 deletions
@@ -21,6 +21,7 @@ set(XPU_PROJECT "extern_xpu")
 set(XPU_API_LIB_NAME "libxpuapi.so")
 set(XPU_RT_LIB_NAME "libxpurt.so")
 set(XPU_CUDA_LIB_NAME "libxpucuda.so")
+set(XPU_CUDA_RT_LIB_NAME "libcudart.so")
 set(XPU_XFT_LIB_NAME "libxft.so")
 set(XPU_XPTI_LIB_NAME "libxpti.so")
 set(XPU_XBLAS_LIB_NAME "libxpu_blas.so")
@@ -31,7 +32,7 @@ if(NOT DEFINED XPU_XRE_BASE_VERSION)
   set(XPU_XRE_BASE_VERSION "4.32.0.1")
 endif()
 if(NOT DEFINED XPU_XHPC_BASE_DATE)
-  set(XPU_XHPC_BASE_DATE "eb35/20241104")
+  set(XPU_XHPC_BASE_DATE "dev/20241113")
 endif()
 set(XPU_XCCL_BASE_VERSION "1.2.11e")
 if(NOT DEFINED XPU_XFT_BASE_VERSION)
@@ -139,6 +140,7 @@ set(XPU_API_LIB "${XPU_LIB_DIR}/${XPU_API_LIB_NAME}")
 set(XPU_XBLAS_LIB "${XPU_LIB_DIR}/${XPU_XBLAS_LIB_NAME}")
 set(XPU_RT_LIB "${XPU_LIB_DIR}/${XPU_RT_LIB_NAME}")
 set(XPU_CUDA_LIB "${XPU_LIB_DIR}/${XPU_CUDA_LIB_NAME}")
+set(XPU_CUDA_RT_LIB "${XPU_LIB_DIR}/${XPU_CUDA_RT_LIB_NAME}")
 set(XPU_XFA_LIB "${XPU_LIB_DIR}/${XPU_XFA_LIB_NAME}")
 set(XPU_XPUDNN_LIB "${XPU_LIB_DIR}/${XPU_XPUDNN_LIB_NAME}")
 
@@ -160,29 +162,51 @@ if(WITH_XPU_BKCL)
   include_directories(${XPU_BKCL_INC_DIR})
 endif()
 
-ExternalProject_Add(
-  ${XPU_PROJECT}
-  ${EXTERNAL_PROJECT_LOG_ARGS}
-  PREFIX ${SNAPPY_PREFIX_DIR}
-  DOWNLOAD_DIR ${XPU_DOWNLOAD_DIR}
-  DOWNLOAD_COMMAND
-    bash ${CMAKE_SOURCE_DIR}/tools/xpu/pack_paddle_dependence.sh ${XPU_XRE_URL}
-    ${XPU_XRE_DIR_NAME} ${XPU_XHPC_URL} ${XPU_XHPC_DIR_NAME} ${XPU_XCCL_URL}
-    ${XPU_XCCL_DIR_NAME} ${XPU_XHPC_URL} ${XPU_XHPC_DIR_NAME} && wget
-    ${XPU_XFT_GET_DEPENCE_URL} && bash get_xft_dependence.sh ${XPU_XFT_URL}
-    ${XPU_XFT_DIR_NAME} && bash
-    ${CMAKE_SOURCE_DIR}/tools/xpu/get_xpti_dependence.sh ${XPU_XPTI_URL}
-    ${XPU_XPTI_DIR_NAME}
-  DOWNLOAD_NO_PROGRESS 1
-  UPDATE_COMMAND ""
-  CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${XPU_INSTALL_ROOT}
-  CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${XPU_INSTALL_ROOT}
-  BUILD_BYPRODUCTS ${XPU_API_LIB}
-  BUILD_BYPRODUCTS ${XPU_XBLAS_LIB}
-  BUILD_BYPRODUCTS ${XPU_XPUDNN_LIB}
-  BUILD_BYPRODUCTS ${XPU_XFA_LIB}
-  BUILD_BYPRODUCTS ${XPU_RT_LIB}
-  BUILD_BYPRODUCTS ${XPU_BKCL_LIB})
+if(WITH_XPU_XRE5)
+  ExternalProject_Add(
+    ${XPU_PROJECT}
+    ${EXTERNAL_PROJECT_LOG_ARGS}
+    PREFIX ${SNAPPY_PREFIX_DIR}
+    DOWNLOAD_DIR ${XPU_DOWNLOAD_DIR}
+    DOWNLOAD_COMMAND
+      bash ${CMAKE_SOURCE_DIR}/tools/xpu/pack_paddle_dependence.sh
+      ${XPU_XRE_URL} ${XPU_XRE_DIR_NAME} ${XPU_XHPC_URL} ${XPU_XHPC_DIR_NAME}
+      ${XPU_XCCL_URL} ${XPU_XCCL_DIR_NAME} 1 && wget ${XPU_XFT_GET_DEPENCE_URL}
+      && bash get_xft_dependence.sh ${XPU_XFT_URL} ${XPU_XFT_DIR_NAME} && bash
+      ${CMAKE_SOURCE_DIR}/tools/xpu/get_xpti_dependence.sh ${XPU_XPTI_URL}
+      ${XPU_XPTI_DIR_NAME}
+    DOWNLOAD_NO_PROGRESS 1
+    UPDATE_COMMAND ""
+    CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${XPU_INSTALL_ROOT}
+    CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${XPU_INSTALL_ROOT}
+    BUILD_BYPRODUCTS ${XPU_API_LIB}
+    BUILD_BYPRODUCTS ${XPU_XBLAS_LIB}
+    BUILD_BYPRODUCTS ${XPU_XPUDNN_LIB}
+    BUILD_BYPRODUCTS ${XPU_XFA_LIB}
+    BUILD_BYPRODUCTS ${XPU_RT_LIB}
+    BUILD_BYPRODUCTS ${XPU_CUDA_RT_LIB}
+    BUILD_BYPRODUCTS ${XPU_BKCL_LIB})
+else()
+  ExternalProject_Add(
+    ${XPU_PROJECT}
+    ${EXTERNAL_PROJECT_LOG_ARGS}
+    PREFIX ${SNAPPY_PREFIX_DIR}
+    DOWNLOAD_DIR ${XPU_DOWNLOAD_DIR}
+    DOWNLOAD_COMMAND
+      bash ${CMAKE_SOURCE_DIR}/tools/xpu/pack_paddle_dependence.sh
+      ${XPU_XRE_URL} ${XPU_XRE_DIR_NAME} ${XPU_XHPC_URL} ${XPU_XHPC_DIR_NAME}
+      ${XPU_XCCL_URL} ${XPU_XCCL_DIR_NAME} 0 && wget ${XPU_XFT_GET_DEPENCE_URL}
+      && bash get_xft_dependence.sh ${XPU_XFT_URL} ${XPU_XFT_DIR_NAME} && bash
+      ${CMAKE_SOURCE_DIR}/tools/xpu/get_xpti_dependence.sh ${XPU_XPTI_URL}
+      ${XPU_XPTI_DIR_NAME}
+    DOWNLOAD_NO_PROGRESS 1
+    UPDATE_COMMAND ""
+    CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${XPU_INSTALL_ROOT}
+    CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${XPU_INSTALL_ROOT}
+    BUILD_BYPRODUCTS ${XPU_API_LIB}
+    BUILD_BYPRODUCTS ${XPU_RT_LIB}
+    BUILD_BYPRODUCTS ${XPU_BKCL_LIB})
+endif()
 
 include_directories(${XPU_INC_DIR})
 add_library(shared_xpuapi SHARED IMPORTED GLOBAL)
@@ -192,7 +216,11 @@ set_property(TARGET shared_xpuapi PROPERTY IMPORTED_LOCATION "${XPU_API_LIB}")
 # for cc_library(xxx SRCS xxx.c DEPS xpulib)
 generate_dummy_static_lib(LIB_NAME "xpulib" GENERATOR "xpu.cmake")
 
-target_link_libraries(xpulib ${XPU_API_LIB} ${XPU_RT_LIB})
+if(WITH_XPU_XRE5)
+  target_link_libraries(xpulib ${XPU_API_LIB} ${XPU_RT_LIB} ${XPU_CUDA_RT_LIB})
+else()
+  target_link_libraries(xpulib ${XPU_API_LIB} ${XPU_RT_LIB})
+endif()
 
 if(WITH_XPU_XFT)
   message(STATUS "Compile with XPU XFT!")
@@ -205,15 +233,17 @@ endif()
 
 set(XPU_XHPC_INC_DIR "${XPU_INC_DIR}/xhpc")
 include_directories(${XPU_XHPC_INC_DIR})
-set(XPU_XBLAS_INC_DIR "${XPU_INC_DIR}/xhpc/xblas")
-include_directories(${XPU_XBLAS_INC_DIR})
-set(XPU_XPUDNN_INC_DIR "${XPU_INC_DIR}/xhpc/xpudnn")
-include_directories(${XPU_XPUDNN_INC_DIR})
+set(XPU_XRE_INC_DIR "${XPU_INC_DIR}/xre")
+include_directories(${XPU_XRE_INC_DIR})
 
 if(WITH_XPU_XRE5)
   add_definitions(-DPADDLE_WITH_XPU_XRE5)
+  set(XPU_XBLAS_INC_DIR "${XPU_INC_DIR}/xhpc/xblas")
+  include_directories(${XPU_XBLAS_INC_DIR})
   set(XPU_XFA_INC_DIR "${XPU_INC_DIR}/xhpc/xfa")
   include_directories(${XPU_XFA_INC_DIR})
+  set(XPU_XPUDNN_INC_DIR "${XPU_INC_DIR}/xhpc/xpudnn")
+  include_directories(${XPU_XPUDNN_INC_DIR})
 endif()
 
 if(WITH_XPTI)
@@ -236,14 +266,14 @@ if(WITH_XPU_XRE5)
   target_link_libraries(
     xpulib
     ${XPU_RT_LIB}
+    ${XPU_CUDA_RT_LIB}
     ${XPU_BKCL_LIB}
     ${XPU_XBLAS_LIB}
     ${XPU_API_LIB}
     ${XPU_XFA_LIB}
     ${XPU_XPUDNN_LIB})
 else()
-  target_link_libraries(xpulib ${XPU_RT_LIB} ${XPU_BKCL_LIB} ${XPU_XBLAS_LIB}
-                        ${XPU_API_LIB})
+  target_link_libraries(xpulib ${XPU_RT_LIB} ${XPU_BKCL_LIB} ${XPU_API_LIB})
 endif()
 
 add_dependencies(xpulib ${XPU_PROJECT})
 
@@ -818,7 +818,6 @@ XPUOpMap& get_kl3_ops() {
        XPUKernelSet({
            phi::DataType::FLOAT32,
            phi::DataType::FLOAT16,
-           phi::DataType::BFLOAT16,
        })},
       {"not_equal",
        XPUKernelSet({phi::DataType::INT64,
 
@@ -334,7 +334,7 @@ void FusedMultiTransformerInt8XpuKernel(
           PADDLE_ENFORCE_XDNN_SUCCESS(r, "xpu::gather_part");
         } else {
           if (index_type == DataType::INT32) {
-            r = xpu::gather<XPUTypeT, int32_t>(
+            r = xpu::paddle_gather<XPUTypeT, int32_t>(
                 ctx.x_context(),
                 cache_kv_data,
                 gather_index_t->data<int32_t>(),
@@ -344,7 +344,7 @@ void FusedMultiTransformerInt8XpuKernel(
                                                    : gather_index_t->dims()[0],
                 gather_axis);
           } else {
-            r = xpu::gather<XPUTypeT, int64_t>(
+            r = xpu::paddle_gather<XPUTypeT, int64_t>(
                 ctx.x_context(),
                 cache_kv_data,
                 gather_index_t->data<int64_t>(),
 
@@ -257,7 +257,7 @@ void FusedMultiTransformerXpuKernel(
       const auto& index_type = gather_index_t->dtype();
       if (cache_kv_gather_dims != cache_kv_dims) {
         if (index_type == DataType::INT32) {
-          r = xpu::gather<XPUTypeT, int32_t>(
+          r = xpu::paddle_gather<XPUTypeT, int32_t>(
               ctx.x_context(),
               cache_kv_data,
               gather_index_t->data<int32_t>(),
@@ -267,7 +267,7 @@ void FusedMultiTransformerXpuKernel(
                                                  : gather_index_t->dims()[0],
               gather_axis);
         } else {
-          r = xpu::gather<XPUTypeT, int64_t>(
+          r = xpu::paddle_gather<XPUTypeT, int64_t>(
               ctx.x_context(),
               cache_kv_data,
               gather_index_t->data<int64_t>(),
@@ -287,7 +287,7 @@ void FusedMultiTransformerXpuKernel(
         PADDLE_ENFORCE_XDNN_SUCCESS(r, "xpu::copy");
       } else {  // inplace gather
         if (index_type == DataType::INT32) {
-          r = xpu::gather<XPUTypeT, int32_t>(
+          r = xpu::paddle_gather<XPUTypeT, int32_t>(
               ctx.x_context(),
               cache_kv_data,
               gather_index_t->data<int32_t>(),
@@ -297,7 +297,7 @@ void FusedMultiTransformerXpuKernel(
                                                  : gather_index_t->dims()[0],
               gather_axis);
         } else {
-          r = xpu::gather<XPUTypeT, int64_t>(
+          r = xpu::paddle_gather<XPUTypeT, int64_t>(
               ctx.x_context(),
               cache_kv_data,
               gather_index_t->data<int64_t>(),
 
@@ -75,7 +75,7 @@ void GetSinCosByPassValue(const Context& dev_ctx,
             "The batch_size and seq_len of position_ids must be the same as "
             "those of q."));
 
-    ret = xpu::gather<XPUSCType, int64_t>(
+    ret = xpu::paddle_gather<XPUSCType, int64_t>(
         dev_ctx.x_context(),
         reinterpret_cast<const XPUSCType*>(sin->data()),
         position_ids->data<int64_t>(),
@@ -84,7 +84,7 @@ void GetSinCosByPassValue(const Context& dev_ctx,
         batch_size * seq_len,
         0);
     PADDLE_ENFORCE_XDNN_SUCCESS(ret, "gather");
-    ret = xpu::gather<XPUSCType, int64_t>(
+    ret = xpu::paddle_gather<XPUSCType, int64_t>(
         dev_ctx.x_context(),
         reinterpret_cast<const XPUSCType*>(cos->data()),
         position_ids->data<int64_t>(),
@@ -361,13 +361,14 @@ void XPUFusedRotaryHalf(const Context& dev_ctx,
       true,
       common::errors::Unimplemented("The xpu rotary half do not support "
                                     "sin/cos with different dtype as input."));
-  auto single_func = &xpu::rotary_no_freqs_embedding_v2<XPUType>;
-  auto fusion_func = &xpu::rotary_no_freqs_qk_embedding_v2<XPUType>;
+  auto single_func = &xpu::rotary_no_freqs_embedding_v2<XPUType, XPUSCType>;
+  auto fusion_func = &xpu::rotary_no_freqs_qk_embedding_v2<XPUType, XPUSCType>;
   const char* single_func_name = "rotary_no_freqs_embedding_v2";
   const char* fusion_func_name = "xpu::rotary_no_freqs_qk_embedding_v2";
   if (is_bwd) {
-    single_func = &xpu::rotary_no_freqs_embedding_v2_grad<XPUType>;
-    fusion_func = &xpu::rotary_no_freqs_qk_embedding_v2_grad<XPUType>;
+    single_func = &xpu::rotary_no_freqs_embedding_v2_grad<XPUType, XPUSCType>;
+    fusion_func =
+        &xpu::rotary_no_freqs_qk_embedding_v2_grad<XPUType, XPUSCType>;
   }
 
   if (head_dim * sizeof(XPUType) <= 1024 && head_dim % 64 == 0 && in_k) {
@@ -376,8 +377,8 @@ void XPUFusedRotaryHalf(const Context& dev_ctx,
         dev_ctx.x_context(),
         reinterpret_cast<const XPUType*>(in_q.data()),
         reinterpret_cast<const XPUType*>(in_k->data()),
-        reinterpret_cast<const XPUType*>(sin_data),
-        reinterpret_cast<const XPUType*>(cos_data),
+        reinterpret_cast<const XPUSCType*>(sin_data),
+        reinterpret_cast<const XPUSCType*>(cos_data),
         reinterpret_cast<XPUType*>(out_q->data()),
         reinterpret_cast<XPUType*>(out_k->data()),
         {batch_size, seq_len, num_heads, head_dim},
@@ -390,8 +391,8 @@ void XPUFusedRotaryHalf(const Context& dev_ctx,
     int ret = single_func(
         dev_ctx.x_context(),
         reinterpret_cast<const XPUType*>(in_q.data()),
-        reinterpret_cast<const XPUType*>(sin_data),
-        reinterpret_cast<const XPUType*>(cos_data),
+        reinterpret_cast<const XPUSCType*>(sin_data),
+        reinterpret_cast<const XPUSCType*>(cos_data),
         reinterpret_cast<XPUType*>(out_q->data()),
         {batch_size, seq_len, num_heads, head_dim},
         {batch_size, seq_len, 1, head_dim},
@@ -402,8 +403,8 @@ void XPUFusedRotaryHalf(const Context& dev_ctx,
       int64_t num_heads_k = in_k->dims()[2];
       int ret = single_func(dev_ctx.x_context(),
                             reinterpret_cast<const XPUType*>(in_k->data()),
-                            reinterpret_cast<const XPUType*>(sin_data),
-                            reinterpret_cast<const XPUType*>(cos_data),
+                            reinterpret_cast<const XPUSCType*>(sin_data),
+                            reinterpret_cast<const XPUSCType*>(cos_data),
                             reinterpret_cast<XPUType*>(out_k->data()),
                             {batch_size, seq_len, num_heads_k, head_dim},
                             {batch_size, seq_len, 1, head_dim},
@@ -421,8 +422,8 @@ void XPUFusedRotaryHalf(const Context& dev_ctx,
     int ret = single_func(
         dev_ctx.x_context(),
         reinterpret_cast<const XPUType*>(in_v->data()),
-        reinterpret_cast<const XPUType*>(sin_data),
-        reinterpret_cast<const XPUType*>(cos_data),
+        reinterpret_cast<const XPUSCType*>(sin_data),
+        reinterpret_cast<const XPUSCType*>(cos_data),
         reinterpret_cast<XPUType*>(out_v->data()),
         {batch_size, seq_len, num_heads_v, head_dim},
         {batch_size, seq_len, 1, head_dim},
 
@@ -16,14 +16,14 @@
 #include "paddle/phi/backends/xpu/enforce_xpu.h"
 #include "paddle/phi/backends/xpu/xpu_context.h"
 #include "paddle/phi/core/kernel_registry.h"
-#include "xblas/cublasLt.h"
 
-#ifndef PADDLE_WITH_XPU_XRE5
+#ifdef PADDLE_WITH_XPU_XRE5
+#include "xblas/cublasLt.h"
+namespace xblas = baidu::xpu::xblas;
+#else
 #include "paddle/phi/kernels/xpu/xpu_api_wrapper.h"
 #endif
 
-namespace xblas = baidu::xpu::xblas;
-
 namespace phi {
 
 template <typename T, typename Context>
 
@@ -41,26 +41,26 @@ void CEmbeddingKernel(const Context& dev_ctx,
   // ym: number of elements of input ids.
   const auto& index_type = ids.dtype();
   if (index_type == phi::DataType::INT32) {
-    int r = xpu::embedding(dev_ctx.x_context(),
-                           reinterpret_cast<const XPUType*>(table_data),
-                           ids.data<int32_t>(),
-                           reinterpret_cast<XPUType*>(output_data),
-                           height,
-                           width,
-                           ids.numel(),
-                           -1,
-                           static_cast<int32_t>(start_index));
+    int r = xpu::paddle_embedding(dev_ctx.x_context(),
+                                  reinterpret_cast<const XPUType*>(table_data),
+                                  ids.data<int32_t>(),
+                                  reinterpret_cast<XPUType*>(output_data),
+                                  height,
+                                  width,
+                                  ids.numel(),
+                                  -1,
+                                  static_cast<int32_t>(start_index));
     PADDLE_ENFORCE_XDNN_SUCCESS(r, "embedding");
   } else if (index_type == phi::DataType::INT64) {
-    int r = xpu::embedding(dev_ctx.x_context(),
-                           reinterpret_cast<const XPUType*>(table_data),
-                           ids.data<int64_t>(),
-                           reinterpret_cast<XPUType*>(output_data),
-                           height,
-                           width,
-                           ids.numel(),
-                           -1,
-                           static_cast<int64_t>(start_index));
+    int r = xpu::paddle_embedding(dev_ctx.x_context(),
+                                  reinterpret_cast<const XPUType*>(table_data),
+                                  ids.data<int64_t>(),
+                                  reinterpret_cast<XPUType*>(output_data),
+                                  height,
+                                  width,
+                                  ids.numel(),
+                                  -1,
+                                  static_cast<int64_t>(start_index));
     PADDLE_ENFORCE_XDNN_SUCCESS(r, "embedding");
   } else {
     PADDLE_THROW(common::errors::Unavailable(
 
@@ -146,11 +146,11 @@ void DistributeFpnProposalsKernel(
       multi_fpn_rois[i]->Resize({sub_rois_num, funcs::kBoxDim});
       dev_ctx.template Alloc<T>(multi_fpn_rois[i]);
 
-      std::vector<int> fpn_rois_shape(fpn_rois.dims().size());
+      std::vector<int64_t> fpn_rois_shape(fpn_rois.dims().size());
       for (int i = 0; i < fpn_rois.dims().size(); ++i) {
         fpn_rois_shape[i] = fpn_rois.dims()[i];
       }
-      int r1 = xpu::gather<XPUType, int>(
+      int r1 = xpu::paddle_gather<XPUType, int>(
           dev_ctx.x_context(),
           reinterpret_cast<const XPUType*>(fpn_rois.data<T>()),
           sub_idx.data<int>(),