refactor: cleanup headers and optimize branch checks.

Author: yingxudeng

xllm/core/common/CMakeLists.txt

@@ -30,7 +30,6 @@ cc_library(
     absl::random_random
     absl::strings
     torch
-    $<$<BOOL:${USE_NPU}>:torch_python>
     $<$<BOOL:${USE_NPU}>:torch_npu>
     $<$<BOOL:${USE_MSPTI}>:mspti>
     $<$<BOOL:${USE_NPU}>:ms_tools_ext>

xllm/core/common/global_flags.cpp

@@ -466,7 +466,9 @@ DEFINE_bool(enable_constrained_decoding,
             "through pre-defined rules.");
 
 
+#if defined(USE_NPU)
 DEFINE_string(
     npu_kernel_backend,
     "ATB",
     "NPU kernel backend. Supported options: ATB, TORCH. Default is ATB.");
+#endif

xllm/core/common/global_flags.h

@@ -227,4 +227,6 @@ DECLARE_double(dit_cache_residual_diff_threshold);
 
 DECLARE_bool(enable_constrained_decoding);
 
+#if defined(USE_NPU)
 DECLARE_string(npu_kernel_backend);
+#endif

xllm/core/framework/parallel_state/collective_communicator.cpp

@@ -21,7 +21,6 @@ limitations under the License.
 #include "npu_process_group.h"
 #include "xllm_kernels/core/include/atb_speed/base/external_comm_manager.h"
 #include "xllm_kernels/core/include/atb_speed/utils/singleton.h"
-#include "xllm_kernels/models/base/param/mapping.h"
 #elif defined(USE_MLU)
 #include "mlu_process_group.h"
 #elif defined(USE_CUDA)

xllm/core/framework/parallel_state/cuda_process_group.h

@@ -34,7 +34,8 @@ class ProcessGroupNccl : public ProcessGroup {
       : ProcessGroup(device) {
     c10::intrusive_ptr<c10d::ProcessGroupNCCL::Options> pg_options =
         c10d::ProcessGroupNCCL::Options::create();
-#if TORCH_VERSION_MAJOR >= 2 && TORCH_VERSION_MINOR >= 7
+#if TORCH_VERSION_MAJOR > 2 || \
+    (TORCH_VERSION_MAJOR == 2 && TORCH_VERSION_MINOR >= 7)
     pg_options->group_name = group_name;
 #endif
     int rank = global_rank;

xllm/core/framework/parallel_state/npu_process_group.cpp

@@ -14,12 +14,6 @@ limitations under the License.
 ==============================================================================*/
 
 #include "npu_process_group.h"
-#ifdef TORCH_HIGHER_THAN_PTA6
-#include <torch_npu/csrc/framework/OpCommand.h>
-#else
-#include <torch_npu/csrc/aten/NPUNativeFunctions.h>
-#include <torch_npu/csrc/framework/utils/OpPreparation.h>
-#endif
 
 #include <c10d/ProcessGroup.hpp>
 #include <c10d/TCPStore.hpp>
@@ -49,7 +43,10 @@ ProcessGroupHCCL::ProcessGroupHCCL(int global_rank,
     : ProcessGroup(device) {
   c10::intrusive_ptr<c10d_npu::ProcessGroupHCCL::Options> hccl_pg_options =
       c10d_npu::ProcessGroupHCCL::Options::create();
-  // hccl_pg_options->group_name = group_name;
+#if TORCH_VERSION_MAJOR > 2 || \
+    (TORCH_VERSION_MAJOR == 2 && TORCH_VERSION_MINOR >= 7)
+  hccl_pg_options->group_name = group_name;
+#endif
   int rank = global_rank;
   if (world_size != rank_size) {
     auto [local_rank, group_ranks] =

xllm/core/layers/common/tests/tests_utils.h

@@ -125,7 +125,8 @@ class MockBackend : public c10d::Backend {
 
   int64_t getSize() const { return world_size_; }
 
-#if TORCH_VERSION_MAJOR >= 2 && TORCH_VERSION_MINOR >= 7
+#if TORCH_VERSION_MAJOR > 2 || \
+    (TORCH_VERSION_MAJOR == 2 && TORCH_VERSION_MINOR >= 7)
   void shutdown() override {
     // Mock implementation - do nothing
   }