[XPU] Fix code to support xpu with aarch64 (#73071)

HydrogenSulfate · dynamicheart · web-flow · commit 5c1b862f5087 · 2025-06-04T16:07:02.000+08:00
* [XPU] support aarch64

* fix

* fix

* fix

* fix

* fix

* fix

* fix

* fix for support xpu + aarch64 env

---------

Co-authored-by: Jianbang Yang &lt;yangjianbang112@gmail.com&gt;
diff --git a/cmake/external/xpu.cmake b/cmake/external/xpu.cmake
@@ -35,7 +35,7 @@ add_compile_definitions(XPUAPI_NOT_INCLUDE_DEPRECATED)
 if(NOT DEFINED XPU_XHPC_BASE_DATE)
   set(XPU_XHPC_BASE_DATE "dev/20250417")
 endif()
-set(XPU_XCCL_BASE_VERSION "3.0.2.5") # For XRE5
+set(XPU_XCCL_BASE_VERSION "3.0.2.7") # For XRE5
 if(NOT DEFINED XPU_XFT_BASE_VERSION)
   set(XPU_XFT_BASE_VERSION "20250507/xpu3")
 endif()
@@ -95,10 +95,18 @@ if(WITH_XPU_FFT)
   set(XPU_FFT_DIR_NAME "xpufft_ubuntu2004-x86_64")
 endif()
 
-if(WITH_AARCH64)
-  set(XPU_XRE_DIR_NAME "xre-kylin_aarch64")
-  set(XPU_XCCL_DIR_NAME "") # TODO: xccl has no kylin output now.
-  set(XPU_XFT_DIR_NAME "") # TODO: xft has no kylin output at now.
+if(WITH_ARM)
+  if(WITH_XPU_XRE5)
+    set(XPU_XRE_DIR_NAME "xre-kylin_v10_server-aarch64-${XPU_XRE_BASE_VERSION}")
+    # TODO: xccl has no kylin output now. set default value here.
+    set(XPU_XCCL_DIR_NAME "xccl_Linux_x86_64")
+    set(XPU_XHPC_DIR_NAME "xhpc-kylinv4_aarch64")
+    set(XPU_XFT_DIR_NAME "") # TODO: xft has no kylin output at now.
+  else()
+    set(XPU_XRE_DIR_NAME "")
+    set(XPU_XCCL_DIR_NAME "") # TODO: xccl has no kylin output now.
+    set(XPU_XFT_DIR_NAME "") # TODO: xft has no kylin output at now.
+  endif()
 elseif(WITH_SUNWAY)
   set(XPU_XRE_DIR_NAME "xre-deepin_sw6_64")
   set(XPU_XCCL_DIR_NAME "") # TODO: xccl has no deepin output at now.
@@ -349,17 +357,14 @@ if(WITH_XPU_XRE5)
     ${XPU_XBLAS_LIB}
     ${XPU_API_LIB}
     ${XPU_XFA_LIB}
-    ${XPU_XPUDNN_LIB})
+    ${XPU_XPUDNN_LIB}
+    ${XPU_ML_LIB})
 else()
   target_link_libraries(xpulib ${XPU_RT_LIB} ${XPU_API_LIB})
 endif()
 
 if(WITH_XPU_BKCL)
-  if(WITH_XPU_XRE5)
-    target_link_libraries(xpulib ${XPU_ML_LIB} ${XPU_BKCL_LIB})
-  else()
-    target_link_libraries(xpulib ${XPU_BKCL_LIB})
-  endif()
+  target_link_libraries(xpulib ${XPU_BKCL_LIB})
 endif()
 
 add_dependencies(xpulib ${XPU_PROJECT})
diff --git a/cmake/flags.cmake b/cmake/flags.cmake
@@ -152,7 +152,6 @@ if(NOT WIN32)
       -fdata-sections
       -Wl
       -gc-sections
-      -Werror
       -Wall
       -Wextra
       -Wno-unused-parameter
diff --git a/paddle/phi/core/memory/allocation/mmap_allocator.h b/paddle/phi/core/memory/allocation/mmap_allocator.h
@@ -63,7 +63,7 @@ class MemoryMapAllocation : public Allocation {
         map_size_(size) {}
 
   inline const std::string &ipc_name() const { return ipc_name_; }
-  inline const int shared_fd() const { return fd_; }
+  inline int shared_fd() const { return fd_; }
 
   virtual void close();
 
@@ -121,7 +121,7 @@ class MemoryMapWriterAllocation : public Allocation {
         ipc_name_(std::move(ipc_name)) {}
 
   inline const std::string &ipc_name() const { return ipc_name_; }
-  inline const int shared_fd() const { return fd_; }
+  inline int shared_fd() const { return fd_; }
 
   ~MemoryMapWriterAllocation() override;
 
@@ -139,7 +139,7 @@ class MemoryMapReaderAllocation : public Allocation {
         ipc_name_(std::move(ipc_name)) {}
 
   inline const std::string &ipc_name() const { return ipc_name_; }
-  inline const int shared_fd() const { return fd_; }
+  inline int shared_fd() const { return fd_; }
 
   ~MemoryMapReaderAllocation() override;
 
diff --git a/paddle/phi/kernels/xpu/c_embedding_kernel.cc b/paddle/phi/kernels/xpu/c_embedding_kernel.cc
@@ -26,6 +26,7 @@ void CEmbeddingKernel(const Context& dev_ctx,
                       int64_t start_index,
                       int64_t vocab_size,
                       DenseTensor* out) {
+#if defined(PADDLE_WITH_XPU_BKCL)
   const T* table_data = w.data<T>();
   T* output_data = dev_ctx.template Alloc<T>(out);
   using XPUType = typename XPUTypeTrait<T>::Type;
@@ -60,6 +61,11 @@ void CEmbeddingKernel(const Context& dev_ctx,
     PADDLE_THROW(common::errors::Unavailable(
         "XPU c_embedding ids only support int32 or int64."));
   }
+#else
+  PADDLE_THROW(common::errors::PreconditionNotMet(
+      "PaddlePaddle is not compiled with DWITH_XPU_BKCL, please recompile with "
+      "DWITH_XPU_BKCL for using c_embedding."));
+#endif
 }
 }  // namespace phi
 
diff --git a/paddle/phi/kernels/xpu/c_embedding_kernel_grad.cc b/paddle/phi/kernels/xpu/c_embedding_kernel_grad.cc
@@ -28,6 +28,7 @@ void CEmbeddingGradKernel(const Context& dev_ctx,
                           const DenseTensor& out_grad,
                           int64_t start_index,
                           DenseTensor* w_grad) {
+#if defined(PADDLE_WITH_XPU_BKCL)
   w_grad->Resize(w.dims());
   dev_ctx.Alloc(w_grad, w.dtype());
   T* table_grad_data = static_cast<T*>(w_grad->data());
@@ -77,6 +78,11 @@ void CEmbeddingGradKernel(const Context& dev_ctx,
         "XPU c_embedding ids only support int32 or int64."));
   }
   PADDLE_ENFORCE_XDNN_SUCCESS(r, "embedding_grad");
+#else
+  PADDLE_THROW(common::errors::PreconditionNotMet(
+      "PaddlePaddle is not compiled with DWITH_XPU_BKCL, please recompile with "
+      "DWITH_XPU_BKCL for using c_embedding_grad."));
+#endif
 }
 
 }  // namespace phi
diff --git a/paddle/phi/kernels/xpu/c_softmax_with_cross_entropy_grad_kernel.cc b/paddle/phi/kernels/xpu/c_softmax_with_cross_entropy_grad_kernel.cc
@@ -35,6 +35,7 @@ void CSoftmaxWithCrossEntropyGradKernel(const Context& dev_ctx,
                                         int rank,
                                         int nranks,
                                         DenseTensor* logits_grad) {
+#if defined(PADDLE_WITH_BKCL)
   using XPUType = typename XPUTypeTrait<T>::Type;
   const phi::DenseTensor* labels = &label_in;
   const phi::DenseTensor* loss_grad = &loss_grad_in;
@@ -78,6 +79,11 @@ void CSoftmaxWithCrossEntropyGradKernel(const Context& dev_ctx,
         ignore_index);
   }
   PADDLE_ENFORCE_XDNN_SUCCESS(ret, "mask_label_by_index_grad");
+#else
+  PADDLE_THROW(common::errors::PreconditionNotMet(
+      "PaddlePaddle is not compiled with DWITH_XPU_BKCL, please recompile with "
+      "DWITH_XPU_BKCL for using c_softmax_with_cross_entropy_grad."));
+#endif
 }
 
 }  // namespace phi
diff --git a/paddle/phi/kernels/xpu/c_softmax_with_cross_entropy_kernel.cc b/paddle/phi/kernels/xpu/c_softmax_with_cross_entropy_kernel.cc
@@ -143,6 +143,7 @@ struct CSoftmaxWithCrossEntropyFunctor<phi::XPUContext, T> {
                   int nranks,
                   DenseTensor* softmax,
                   DenseTensor* loss) {
+#if defined(PADDLE_WITH_XPU_BKCL)
     using XPUType = typename XPUTypeTrait<T>::Type;
     const phi::DenseTensor* logits = &logits_in;
     const phi::DenseTensor* labels = &label_in;
@@ -327,6 +328,12 @@ struct CSoftmaxWithCrossEntropyFunctor<phi::XPUContext, T> {
     // 将label和ignore_index相同的那些loss，置为0
     FixLossAccordingToIgnoreIndex<T>(
         dev_ctx, labels, &predicted_logits, loss, N, ignore_index);
+#else
+    PADDLE_THROW(common::errors::PreconditionNotMet(
+        "PaddlePaddle is not compiled with DWITH_XPU_BKCL, please recompile "
+        "with "
+        "DWITH_XPU_BKCL for using c_softmax_with_cross_entropy."));
+#endif
   }
 };
 
diff --git a/paddle/phi/kernels/xpu/c_split_kernel.cc b/paddle/phi/kernels/xpu/c_split_kernel.cc
@@ -26,6 +26,7 @@ void CSplitKernel(const Context& dev_ctx,
                   int nranks,
                   bool use_model_parallel,
                   DenseTensor* out) {
+#if defined(PADDLE_WITH_XPU_BKCL)
   using XPUType = typename XPUTypeTrait<T>::Type;
 
   PADDLE_ENFORCE_GE(rank,
@@ -73,6 +74,11 @@ void CSplitKernel(const Context& dev_ctx,
                         split_list,
                         axis);
   PADDLE_ENFORCE_XDNN_SUCCESS(ret, "split");
+#else
+  PADDLE_THROW(common::errors::PreconditionNotMet(
+      "PaddlePaddle is not compiled with DWITH_XPU_BKCL, please recompile with "
+      "DWITH_XPU_BKCL for using p_send_kernel."));
+#endif
 }
 }  // namespace phi
 
diff --git a/paddle/phi/kernels/xpu/weight_only_linear_kernel.cc b/paddle/phi/kernels/xpu/weight_only_linear_kernel.cc
@@ -12,7 +12,6 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include <xft/xdnn_plugin.h>
 #include "paddle/phi/backends/xpu/enforce_xpu.h"
 #include "paddle/phi/core/kernel_registry.h"
 #ifdef PADDLE_WITH_XPU_XRE5
