[Distributed] fix recreate nccl comm bug (#73625)

Author: tianhaodongbd

paddle/fluid/distributed/collective/process_group_nccl.cc

@@ -991,8 +991,8 @@ void ProcessGroupNCCL::Restart() {
     phi::distributed::P2POption p2p_opts = place_to_p2p_opts_.at(place_key);
     phi::distributed::CommContextManager::RecreateNCCLComm(
         store_, store_key, rank_, std::to_string(create_count_), &p2p_opts);
-    create_count_++;
   }
+  create_count_++;
 }
 phi::CUDAStream ProcessGroupNCCL::GetStream(const Place& place) {
   const auto& place_key = GetKeyFromPlace(place);

paddle/fluid/distributed/collective/process_group_nccl.h

@@ -15,6 +15,7 @@
 #pragma once
 
 #include <chrono>
+#include <map>
 #include <memory>
 #include <string>
 #include <unordered_map>
@@ -285,7 +286,7 @@ class ProcessGroupNCCL final : public ProcessGroupWithStream {
 
   uint64_t comm_seq_{0};
   std::unordered_map<std::string, uint64_t> p2p_comm_seq_;
-  std::unordered_map<std::string, std::string> place_to_group_key_;
+  std::map<std::string, std::string> place_to_group_key_;
 
   // TODO(sunyilun): attrs below will be removed later
   std::mutex mutex_;

paddle/phi/core/distributed/comm_context_manager.cc

@@ -131,7 +131,7 @@ void CommContextManager::CreateNCCLCommContext(
 void CommContextManager::RecreateNCCLComm(const std::shared_ptr<Store>& store,
                                           const std::string& unique_comm_key,
                                           int rank,
-                                          const std::string& hash_key,
+                                          const std::string& recreate_key,
                                           const P2POption* p2p_opt) {
   auto& comm_context_manager = CommContextManager::GetInstance();
 
@@ -140,7 +140,8 @@ void CommContextManager::RecreateNCCLComm(const std::shared_ptr<Store>& store,
     PADDLE_ENFORCE_GPU_SUCCESS(phi::dynload::ncclGetUniqueId(&nccl_id));
   }
 
-  std::string unique_key = "NCCLCommContext/" + unique_comm_key + hash_key;
+  std::string unique_key =
+      "NCCLCommContext/" + unique_comm_key + "/" + recreate_key;
   if (rank == 0 || (p2p_opt && p2p_opt->is_p2p_op && p2p_opt->p2p_rank == 0)) {
     std::vector<uint8_t> nccl_id_wrapper(
         reinterpret_cast<uint8_t*>(&nccl_id),