bugfix: fix build error on cuda and ilu device. (#1252)

Author: XuZhang99

xllm/core/framework/parallel_state/dit_collective_communicator.cpp

@@ -92,6 +92,7 @@ void DiTCollectiveCommunicator::create_process_groups(
     auto local_rank = tp_parallel_info.rank();
     auto& rank_per_group = tp_parallel_info.rank_per_group()[group_id];
     int port_offset = group_id + 1;
+#if defined(USE_NPU) || defined(USE_MLU)
     dit_tp_group_ = create_process_group(global_rank,
                                          local_rank,
                                          rank_per_group,
@@ -101,6 +102,7 @@ void DiTCollectiveCommunicator::create_process_groups(
                                          host,
                                          "tp_group",
                                          device);
+#endif
     parallel_args_->dit_tp_group_ = dit_tp_group_.get();
     port += num_group;
   }
@@ -112,6 +114,7 @@ void DiTCollectiveCommunicator::create_process_groups(
     auto local_rank = sp_parallel_info.rank();
     auto& rank_per_group = sp_parallel_info.rank_per_group()[group_id];
     int port_offset = group_id + 1;
+#if defined(USE_NPU) || defined(USE_MLU)
     dit_sp_group_ = create_process_group(global_rank,
                                          local_rank,
                                          rank_per_group,
@@ -121,6 +124,7 @@ void DiTCollectiveCommunicator::create_process_groups(
                                          host,
                                          "sp_group",
                                          device);
+#endif
     parallel_args_->dit_sp_group_ = dit_sp_group_.get();
     port += num_group;
   }
@@ -132,6 +136,7 @@ void DiTCollectiveCommunicator::create_process_groups(
     auto local_rank = cfg_parallel_info.rank();
     auto& rank_per_group = cfg_parallel_info.rank_per_group()[group_id];
     int port_offset = group_id + 1;
+#if defined(USE_NPU) || defined(USE_MLU)
     dit_cfg_group_ = create_process_group(global_rank,
                                           local_rank,
                                           rank_per_group,
@@ -141,6 +146,7 @@ void DiTCollectiveCommunicator::create_process_groups(
                                           host,
                                           "cfg_group",
                                           device);
+#endif
     parallel_args_->dit_cfg_group_ = dit_cfg_group_.get();
     port += num_group;
   }
@@ -152,6 +158,7 @@ void DiTCollectiveCommunicator::create_process_groups(
     auto local_rank = dp_parallel_info.rank();
     auto& rank_per_group = dp_parallel_info.rank_per_group()[group_id];
     int port_offset = group_id + 1;
+#if defined(USE_NPU) || defined(USE_MLU)
     dit_dp_group_ = create_process_group(global_rank,
                                          local_rank,
                                          rank_per_group,
@@ -161,6 +168,7 @@ void DiTCollectiveCommunicator::create_process_groups(
                                          host,
                                          "dp_group",
                                          device);
+#endif
     parallel_args_->dit_dp_group_ = dit_dp_group_.get();
     port += num_group;
   }

xllm/core/framework/parallel_state/process_group.cpp

@@ -198,6 +198,8 @@ std::unique_ptr<ProcessGroup> create_process_group(
       rank, world_size, rank_size, port, trans, host, group_name, device);
 }
 
+#if defined(USE_NPU) || defined(USE_MLU)
+// we only support DiT models onNPU and MLU for now.
 // TODO: This function is used by DiT models, since the DiT communication group
 // info have already been calculated by rank_generator, we only need to pass the
 // info to create the process groups. For any device that want to reuse the
@@ -223,4 +225,5 @@ std::unique_ptr<ProcessGroup> create_process_group(
                                             group_name,
                                             device);
 }
+#endif
 }  // namespace xllm

xllm/core/framework/parallel_state/process_group.h

@@ -133,6 +133,8 @@ std::unique_ptr<xllm::ProcessGroup> create_process_group(
     const std::string& group_name,
     const torch::Device& device);
 
+#if defined(USE_NPU) || defined(USE_MLU)
+// for DiT models
 std::unique_ptr<xllm::ProcessGroup> create_process_group(
     int32_t global_rank,
     int32_t local_rank,
@@ -143,5 +145,6 @@ std::unique_ptr<xllm::ProcessGroup> create_process_group(
     const std::string& host,
     const std::string& group_name,
     const torch::Device& device);
+#endif
 
 }  // namespace xllm