[NewFeature]custom_allreduce support cudagraph recapture (#4305)

* custom_allreduce support cudagraph recapture

* add shut_down/restart default group

custom_ops/gpu_ops/cpp_extensions.cc

@@ -623,6 +623,8 @@ int64_t open_mem_handle(paddle::Tensor& mem_handle);
 
 void free_shared_buffer(int64_t buffer);
 
+void clear_ipc_handles(int64_t _fa);
+
 // speculative decoding Kernel
 std::vector<paddle::Tensor> SpeculateGetPaddingOffset(
     const paddle::Tensor& input_ids,
@@ -1229,6 +1231,8 @@ PYBIND11_MODULE(fastdeploy_ops, m) {
 
   m.def("free_shared_buffer", &free_shared_buffer, "free_shared_buffer");
 
+  m.def("clear_ipc_handles", &clear_ipc_handles, "clear_ipc_handles");
+
   m.def("open_mem_handle", &open_mem_handle, "open_mem_handle");
 
   m.def("get_graph_buffer_ipc_meta", &get_graph_buffer_ipc_meta, "get_graph_buffer_ipc_meta");

custom_ops/gpu_ops/custom_all_reduce/all_reduce.cu

@@ -122,10 +122,14 @@ void register_graph_buffers(fptr_t _fa,
   for (int i = 0; i < handles.size(); i++) {
     bytes.emplace_back(handles[i].begin(), handles[i].end());
   }
-  bytes.reserve(handles.size());
   fa->register_graph_buffers(bytes, offsets);
 }
 
+void clear_ipc_handles(fptr_t _fa) {
+  auto fa = reinterpret_cast<paddle::CustomAllreduce*>(_fa);
+  fa->clear_ipc_handles();
+}
+
 std::tuple<fptr_t, paddle::Tensor> allocate_shared_buffer_and_handle(
     int64_t size) {
 

custom_ops/gpu_ops/custom_all_reduce/all_reduce.cuh

@@ -517,10 +517,15 @@ class CustomAllreduce {
 #undef KL
   }
 
-  ~CustomAllreduce() {
+  void clear_ipc_handles(){
     for (auto [_, ptr] : ipc_handles_) {
       CUDACHECK(cudaIpcCloseMemHandle(ptr));
     }
+    ipc_handles_.clear();
+  }
+
+  ~CustomAllreduce() {
+    clear_ipc_handles();
   }
 };
 }  // namespace paddle