refactor: rename files (#49117)

paddle/fluid/distributed/collective/CMakeLists.txt

@@ -3,13 +3,13 @@ cc_library(
   SRCS ProcessGroup.cc
   DEPS dense_tensor)
 cc_library(
-  processgroup_stream
-  SRCS ProcessGroupStream.cc
+  process_group_stream
+  SRCS process_group_stream.cc
   DEPS dense_tensor)
 cc_library(
   eager_reducer
   SRCS reducer.cc
-  DEPS eager_api processgroup processgroup_stream phi_api string_helper)
+  DEPS eager_api processgroup process_group_stream phi_api string_helper)
 
 if(WITH_DISTRIBUTE)
   cc_library(
@@ -20,10 +20,10 @@ endif()
 
 if(WITH_NCCL OR WITH_RCCL)
   cc_library(
-    processgroup_nccl
-    SRCS ProcessGroupNCCL.cc NCCLTools.cc Common.cc check.cc
+    process_group_nccl
+    SRCS process_group_nccl.cc nccl_tools.cc Common.cc check.cc
     DEPS processgroup
-         processgroup_stream
+         process_group_stream
          place
          enforce
          collective_helper

paddle/fluid/distributed/collective/ProcessGroup.h

@@ -97,6 +97,17 @@ class ProcessGroup {
         GetBackendName()));
   }
 
+  virtual std::shared_ptr<ProcessGroup::Task> AllGather(
+      phi::DenseTensor* out_tensor,
+      const phi::DenseTensor& in_tensor,
+      bool sync_op) {
+    return AllGather(out_tensor,
+                     in_tensor,
+                     /*offset*/ 0,
+                     /*numel*/ -1,  // -1 indicates the whole tensor
+                     sync_op);
+  }
+
   virtual std::shared_ptr<ProcessGroup::Task> AllGather(
       phi::DenseTensor* out_tensor,
       const phi::DenseTensor& in_tensor,
@@ -175,6 +186,16 @@ class ProcessGroup {
         GetBackendName()));
   }
 
+  virtual std::shared_ptr<ProcessGroup::Task> Recv(phi::DenseTensor* tensor,
+                                                   int src_rank,
+                                                   bool sync_op) {
+    return Recv(tensor,
+                src_rank,
+                /*offset*/ 0,
+                /*numel*/ -1,  // -1 indicates the whole tensor
+                sync_op);
+  }
+
   virtual std::shared_ptr<ProcessGroup::Task> Recv(phi::DenseTensor* tensor,
                                                    int src_rank,
                                                    int64_t offset,
@@ -185,6 +206,15 @@ class ProcessGroup {
         GetBackendName()));
   }
 
+  virtual std::shared_ptr<ProcessGroup::Task> Send(
+      const phi::DenseTensor& tensor, int dst_rank, bool sync_op) {
+    return Send(tensor,
+                dst_rank,
+                /*offset*/ 0,
+                /*numel*/ -1,  // -1 indicates the whole tensor
+                sync_op);
+  }
+
   virtual std::shared_ptr<ProcessGroup::Task> Send(
       const phi::DenseTensor& tensor,
       int dst_rank,

paddle/fluid/distributed/collective/ProcessGroupBKCL.h

@@ -20,7 +20,7 @@
 #include <string>
 #include <unordered_map>
 
-#include "paddle/fluid/distributed/collective/ProcessGroupStream.h"
+#include "paddle/fluid/distributed/collective/process_group_stream.h"
 #include "paddle/fluid/distributed/store/store.h"
 #include "paddle/fluid/platform/device/xpu/xpu_header.h"
 #include "paddle/fluid/platform/enforce.h"

paddle/fluid/distributed/collective/check.cc

@@ -14,7 +14,7 @@
 
 #include "paddle/fluid/distributed/collective/check.h"
 
-#include "paddle/fluid/distributed/collective/NCCLTools.h"
+#include "paddle/fluid/distributed/collective/nccl_tools.h"
 #include "paddle/fluid/platform/enforce.h"
 #include "paddle/fluid/platform/place.h"
 #include "paddle/phi/core/dense_tensor.h"

paddle/fluid/distributed/collective/NCCLTools.cc → paddle/fluid/distributed/collective/nccl_tools.cc

@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/distributed/collective/NCCLTools.h"
+#include "paddle/fluid/distributed/collective/nccl_tools.h"
 
 #include "paddle/fluid/platform/enforce.h"
 

paddle/fluid/distributed/collective/ProcessGroupNCCL.cc → paddle/fluid/distributed/collective/process_group_nccl.cc

@@ -12,11 +12,11 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/distributed/collective/ProcessGroupNCCL.h"
+#include "paddle/fluid/distributed/collective/process_group_nccl.h"
 
 #include "paddle/fluid/distributed/collective/Common.h"
-#include "paddle/fluid/distributed/collective/NCCLTools.h"
 #include "paddle/fluid/distributed/collective/check.h"
+#include "paddle/fluid/distributed/collective/nccl_tools.h"
 #include "paddle/fluid/distributed/collective/utils.h"
 #include "paddle/fluid/platform/device/gpu/nccl_helper.h"
 #include "paddle/fluid/platform/place.h"

paddle/fluid/distributed/collective/ProcessGroupNCCL.h → paddle/fluid/distributed/collective/process_group_nccl.h

@@ -20,7 +20,7 @@
 #include <unordered_map>
 #include <vector>
 
-#include "paddle/fluid/distributed/collective/ProcessGroupStream.h"
+#include "paddle/fluid/distributed/collective/process_group_stream.h"
 #include "paddle/fluid/distributed/store/store.h"
 #include "paddle/fluid/platform/cuda_device_guard.h"
 #include "paddle/fluid/platform/device_event.h"
@@ -29,7 +29,7 @@
 #include "paddle/phi/core/device_context.h"
 
 #if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
-#include "paddle/fluid/distributed/collective/NCCLTools.h"
+#include "paddle/fluid/distributed/collective/nccl_tools.h"
 #endif
 
 #ifdef PADDLE_WITH_RCCL

paddle/fluid/distributed/collective/ProcessGroupStream.cc → paddle/fluid/distributed/collective/process_group_stream.cc

@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/distributed/collective/ProcessGroupStream.h"
+#include "paddle/fluid/distributed/collective/process_group_stream.h"
 
 namespace paddle {
 namespace distributed {
@@ -40,6 +40,19 @@ std::shared_ptr<ProcessGroup::Task> ProcessGroupStream::AllGather(
                    /*use_calc_stream*/ false);
 }
 
+std::shared_ptr<ProcessGroup::Task> ProcessGroupStream::AllGather(
+    phi::DenseTensor* out_tensor,
+    const phi::DenseTensor& in_tensor,
+    bool sync_op,
+    bool use_calc_stream) {
+  return AllGather(out_tensor,
+                   in_tensor,
+                   /*offset*/ 0,
+                   /*numel*/ -1,  // -1 indicates the whole tensor
+                   sync_op,
+                   use_calc_stream);
+}
+
 std::shared_ptr<ProcessGroup::Task> ProcessGroupStream::AllGather(
     phi::DenseTensor* out_tensor,
     const phi::DenseTensor& in_tensor,
@@ -200,6 +213,19 @@ std::shared_ptr<ProcessGroup::Task> ProcessGroupStream::Recv(
               /*use_calc_stream*/ false);
 }
 
+std::shared_ptr<ProcessGroup::Task> ProcessGroupStream::Recv(
+    phi::DenseTensor* tensor,
+    int src_rank,
+    bool sync_op,
+    bool use_calc_stream) {
+  return Recv(tensor,
+              src_rank,
+              /*offset*/ 0,
+              /*numel*/ -1,  // -1 indicates sending the whole tensor
+              sync_op,
+              use_calc_stream);
+}
+
 std::shared_ptr<ProcessGroup::Task> ProcessGroupStream::Recv(
     phi::DenseTensor* tensor,
     int src_rank,
@@ -225,6 +251,19 @@ std::shared_ptr<ProcessGroup::Task> ProcessGroupStream::Send(
               /*use_calc_stream*/ false);
 }
 
+std::shared_ptr<ProcessGroup::Task> ProcessGroupStream::Send(
+    const phi::DenseTensor& tensor,
+    int dst_rank,
+    bool sync_op,
+    bool use_calc_stream) {
+  return Send(tensor,
+              dst_rank,
+              /*offset*/ 0,
+              /*numel*/ -1,  // -1 indicates receiving the whole tensor
+              sync_op,
+              use_calc_stream);
+}
+
 std::shared_ptr<ProcessGroup::Task> ProcessGroupStream::Send(
     const phi::DenseTensor& tensor,
     int dst_rank,

paddle/fluid/distributed/collective/ProcessGroupStream.h → paddle/fluid/distributed/collective/process_group_stream.h

@@ -69,6 +69,12 @@ class ProcessGroupStream : public ProcessGroup {
       int64_t numel,
       bool sync_op) override;
 
+  virtual std::shared_ptr<ProcessGroup::Task> AllGather(
+      phi::DenseTensor* out_tensor,
+      const phi::DenseTensor& in_tensor,
+      bool sync_op,
+      bool use_calc_stream);
+
   virtual std::shared_ptr<ProcessGroup::Task> AllGather(
       phi::DenseTensor* out_tensor,
       const phi::DenseTensor& in_tensor,
@@ -161,6 +167,11 @@ class ProcessGroupStream : public ProcessGroup {
                                            int64_t numel,
                                            bool sync_op) override;
 
+  virtual std::shared_ptr<ProcessGroup::Task> Recv(phi::DenseTensor* tensor,
+                                                   int src_rank,
+                                                   bool sync_op,
+                                                   bool use_calc_stream);
+
   virtual std::shared_ptr<ProcessGroup::Task> Recv(phi::DenseTensor* tensor,
                                                    int src_rank,
                                                    int64_t offset,
@@ -174,6 +185,11 @@ class ProcessGroupStream : public ProcessGroup {
                                            int64_t numel,
                                            bool sync_op) override;
 
+  std::shared_ptr<ProcessGroup::Task> Send(const phi::DenseTensor& tensor,
+                                           int dst_rank,
+                                           bool sync_op,
+                                           bool use_calc_stream);
+
   virtual std::shared_ptr<ProcessGroup::Task> Send(
       const phi::DenseTensor& tensor,
       int dst_rank,

paddle/fluid/operators/fused/fused_attention_op.cu

@@ -30,7 +30,7 @@ limitations under the License. */
 #include "paddle/phi/kernels/funcs/math_function.h"
 
 #if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
-#include "paddle/fluid/distributed/collective/ProcessGroupNCCL.h"
+#include "paddle/fluid/distributed/collective/process_group_nccl.h"
 #include "paddle/fluid/platform/collective_helper.h"
 #include "paddle/fluid/platform/device/gpu/nccl_helper.h"
 #endif

paddle/fluid/operators/fused/fused_feedforward_op.cu

@@ -23,7 +23,7 @@ limitations under the License. */
 #include "paddle/phi/kernels/funcs/elementwise_functor.h"
 
 #if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
-#include "paddle/fluid/distributed/collective/ProcessGroupNCCL.h"
+#include "paddle/fluid/distributed/collective/process_group_nccl.h"
 #include "paddle/fluid/platform/collective_helper.h"
 #include "paddle/fluid/platform/device/gpu/nccl_helper.h"
 #endif

paddle/fluid/pybind/CMakeLists.txt

@@ -157,7 +157,7 @@ endif()
 if(WITH_PYTHON)
   set(PYBIND_DEPS ${PYBIND_DEPS} processgroup eager_reducer)
   if(WITH_NCCL OR WITH_RCCL)
-    set(PYBIND_DEPS ${PYBIND_DEPS} processgroup_nccl)
+    set(PYBIND_DEPS ${PYBIND_DEPS} process_group_nccl)
   endif()
   if(WITH_XPU_BKCL)
     set(PYBIND_DEPS ${PYBIND_DEPS} processgroup_bkcl)

paddle/fluid/pybind/distributed_py.cc

@@ -22,8 +22,8 @@ limitations under the License. */
 #endif
 
 #include "paddle/fluid/distributed/collective/ProcessGroup.h"
-#include "paddle/fluid/distributed/collective/ProcessGroupStream.h"
 #include "paddle/fluid/distributed/collective/Types.h"
+#include "paddle/fluid/distributed/collective/process_group_stream.h"
 #include "paddle/fluid/distributed/collective/reducer.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/tensor.h"
@@ -34,7 +34,7 @@ limitations under the License. */
 #include "paddle/phi/api/all.h"
 
 #if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
-#include "paddle/fluid/distributed/collective/ProcessGroupNCCL.h"
+#include "paddle/fluid/distributed/collective/process_group_nccl.h"
 #endif
 
 #if defined(PADDLE_WITH_MPI)
@@ -169,9 +169,7 @@ void BindDistributed(py::module *m) {
                 auto p_dense =
                     std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
                 auto out_dense = *p_dense;
-                // numel == -1 indicates sending the whole tensor
-                return self.Send(
-                    out_dense, dst, /*offset*/ 0, /*numel*/ -1, sync_op);
+                return self.Send(out_dense, dst, sync_op);
               },
               py::arg("tensor"),
               py::arg("dst"),
@@ -215,9 +213,7 @@ void BindDistributed(py::module *m) {
                 auto p_dense =
                     std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
                 auto *in_dense = p_dense.get();
-                // numel == -1 indicates receiving the whole tensor
-                return self.Recv(
-                    in_dense, src, /*offset*/ 0, /*numel*/ -1, sync_op);
+                return self.Recv(in_dense, src, sync_op);
               },
               py::arg("tensor"),
               py::arg("src"),
@@ -270,11 +266,7 @@ void BindDistributed(py::module *m) {
                 auto in_dense = *p_in_tensor;
 
                 auto *dev_ctx = self.GetDeviceContext(in_tensor.place());
-                auto task = self.AllGather(out_dense,
-                                           in_dense,
-                                           /*offset*/ 0,
-                                           /*numel*/ -1,
-                                           sync_op);
+                auto task = self.AllGather(out_dense, in_dense, sync_op);
                 SplitTensor(*dev_ctx, *out_dense, &out_tensor_list);
                 task->UpdateWaitChain(*dev_ctx);
                 return task;
@@ -300,11 +292,7 @@ void BindDistributed(py::module *m) {
                     in_tensor.impl());
                 auto in_dense = *p_in_tensor;
 
-                return self.AllGather(out_dense,
-                                      in_dense,
-                                      /*offset*/ 0,
-                                      /*numel*/ -1,
-                                      sync_op);
+                return self.AllGather(out_dense, in_dense, sync_op);
               },
               py::arg("out"),
               py::arg("in"),
@@ -771,8 +759,6 @@ void BindDistributed(py::module *m) {
                 auto *dev_ctx = self.GetDeviceContext(in_tensor.place(), true);
                 auto task = self.AllGather(out_dense,
                                            in_dense,
-                                           /*offset*/ 0,
-                                           /*numel*/ -1,
                                            /*sync_op*/ true,
                                            /*use_calc_stream*/ true);
                 SplitTensor(*dev_ctx, *out_dense, &out_tensor_list);
@@ -799,8 +785,6 @@ void BindDistributed(py::module *m) {
 
                 return self.AllGather(out_dense,
                                       in_dense,
-                                      /*offset*/ 0,
-                                      /*numel*/ -1,
                                       /*sync_op*/ true,
                                       /*use_calc_stream*/ true);
               },
@@ -1127,11 +1111,8 @@ void BindDistributed(py::module *m) {
                 auto p_dense =
                     std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
                 auto out_dense = *p_dense;
-                // numel == -1 indicates sending the whole tensor
                 return self.Send(out_dense,
                                  dst,
-                                 /*offset*/ 0,
-                                 /*numel*/ -1,
                                  /*sync_op*/ true,
                                  /*use_calc_stream*/ true);
               },
@@ -1177,11 +1158,8 @@ void BindDistributed(py::module *m) {
                 auto p_dense =
                     std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
                 auto *in_dense = p_dense.get();
-                // numel == -1 indicates receiving the whole tensor
                 return self.Recv(in_dense,
                                  src,
-                                 /*offset*/ 0,
-                                 /*numel*/ -1,
                                  /*sync_op*/ true,
                                  /*use_calc_stream*/ true);
               },

paddle/phi/backends/CMakeLists.txt

@@ -67,7 +67,7 @@ endif()
 
 set(COMM_UTILS_DEPS processgroup)
 if(WITH_NCCL OR WITH_RCCL)
-  set(COMM_UTILS_DEPS ${PROCESS_GROUP_UTILS_DEPS} processgroup_nccl)
+  set(COMM_UTILS_DEPS ${PROCESS_GROUP_UTILS_DEPS} process_group_nccl)
 endif()
 if(WITH_CUSTOM_DEVICE)
   set(COMM_UTILS_DEPS ${PROCESS_GROUP_UTILS_DEPS} processgroup_custom)

paddle/phi/backends/processgroup_comm_utils.cc

@@ -15,7 +15,7 @@
 #include "paddle/fluid/distributed/collective/ProcessGroup.h"
 #include "paddle/phi/backends/c_comm_lib.h"
 #if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
-#include "paddle/fluid/distributed/collective/ProcessGroupNCCL.h"
+#include "paddle/fluid/distributed/collective/process_group_nccl.h"
 #endif
 #if defined(PADDLE_WITH_CUSTOM_DEVICE)
 #include "paddle/fluid/distributed/collective/ProcessGroupCustom.h"

paddle/phi/kernels/CMakeLists.txt

@@ -80,7 +80,7 @@ set(COMMON_KERNEL_DEPS
 
 set(COMMON_KERNEL_DEPS ${COMMON_KERNEL_DEPS} processgroup)
 if(WITH_NCCL OR WITH_RCCL)
-  set(COMMON_KERNEL_DEPS ${COMMON_KERNEL_DEPS} processgroup_nccl)
+  set(COMMON_KERNEL_DEPS ${COMMON_KERNEL_DEPS} process_group_nccl)
 endif()
 set(COMMON_KERNEL_DEPS ${COMMON_KERNEL_DEPS} processgroup_comm_utils)
 if(WITH_CUDNN_FRONTEND)

paddle/phi/kernels/gpu/sync_batch_norm_utils.h

@@ -28,7 +28,7 @@ namespace cub = hipcub;
 #endif
 #include "paddle/fluid/distributed/collective/ProcessGroup.h"
 #if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
-#include "paddle/fluid/distributed/collective/ProcessGroupNCCL.h"
+#include "paddle/fluid/distributed/collective/process_group_nccl.h"
 #endif
 #include "paddle/fluid/memory/malloc.h"
 #include "paddle/fluid/platform/device/gpu/nccl_helper.h"