[cherry-pick]fix DeepEP compile in infer mode (#71590) (#71656)

* fix DeepEP compile in infer mode (#71590)

* fix DeepEP compile in infer mode

* add compile judge

* fix include optional

* fix DeepEP with ON_INFER (#71701)

Author: Hongqing-work

paddle/fluid/distributed/collective/deep_ep/deep_ep.cpp

@@ -18,7 +18,6 @@
 // https://github.com/deepseek-ai/DeepEP/blob/main/LICENSE
 
 #include <cuda_runtime.h>
-#include <pybind11/functional.h>
 #include <atomic>
 #include <chrono>
 #include <memory>
@@ -209,6 +208,7 @@ int Buffer::get_root_rdma_rank(bool global) const {
 
 int Buffer::get_local_device_id() const { return device_id; }
 
+#ifndef PADDLE_NO_PYTHON
 pybind11::bytearray Buffer::get_local_ipc_handle() const {
   return {ipc_handles[nvl_rank].reserved, CUDA_IPC_HANDLE_SIZE};
 }
@@ -301,6 +301,7 @@ void Buffer::sync(
   // Ready to use
   available = true;
 }
+#endif
 
 std::tuple<deep_ep::detail::Tensor,
            std::optional<deep_ep::detail::Tensor>,

paddle/fluid/distributed/collective/deep_ep/deep_ep.hpp

@@ -24,8 +24,11 @@
 #undef NDEBUG
 #endif
 
+#ifndef PADDLE_NO_PYTHON
 #include <pybind11/pybind11.h>
 #include <pybind11/pytypes.h>
+#endif
+#include <optional>
 #include <tuple>
 #include <vector>
 #include "paddle/fluid/distributed/collective/deep_ep/include/types.h"
@@ -119,6 +122,7 @@ struct Buffer {
 
   int get_local_device_id() const;
 
+#ifndef PADDLE_NO_PYTHON
   pybind11::bytearray get_local_ipc_handle() const;
 
   pybind11::bytearray get_local_nvshmem_unique_id() const;
@@ -127,6 +131,7 @@ struct Buffer {
             const std::vector<std::optional<pybind11::bytearray>>&
                 all_gathered_handles,
             const std::optional<pybind11::bytearray>& root_unique_id_opt);
+#endif
 
   std::tuple<deep_ep::detail::Tensor,
              std::optional<deep_ep::detail::Tensor>,