fix DeepEP with ON_INFER (#71701)

Author: Hongqing-work

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

@@ -208,7 +208,7 @@ int Buffer::get_root_rdma_rank(bool global) const {
 
 int Buffer::get_local_device_id() const { return device_id; }
 
-#if !defined(PADDLE_ON_INFERENCE) && !defined(PADDLE_NO_PYTHON)
+#ifndef PADDLE_NO_PYTHON
 pybind11::bytearray Buffer::get_local_ipc_handle() const {
   return {ipc_handles[nvl_rank].reserved, CUDA_IPC_HANDLE_SIZE};
 }

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

@@ -24,7 +24,7 @@
 #undef NDEBUG
 #endif
 
-#if !defined(PADDLE_ON_INFERENCE) && !defined(PADDLE_NO_PYTHON)
+#ifndef PADDLE_NO_PYTHON
 #include <pybind11/pybind11.h>
 #include <pybind11/pytypes.h>
 #endif
@@ -122,7 +122,7 @@ struct Buffer {
 
   int get_local_device_id() const;
 
-#if !defined(PADDLE_ON_INFERENCE) && !defined(PADDLE_NO_PYTHON)
+#ifndef PADDLE_NO_PYTHON
   pybind11::bytearray get_local_ipc_handle() const;
 
   pybind11::bytearray get_local_nvshmem_unique_id() const;