refactor: clean up device runtime interfaces and initialization semantics

- Split internal implementation headers into a separate include group
- Drop redundant explicit default initialization for Device
- Add `impl` suffix to CUDA guard implementation files
- Unify Arange initialization via DeviceGuardImpl

Author: kilinchange

infini_train/include/autograd/comm.h

@@ -33,7 +33,7 @@ class Scatter : public autograd::Function {
 private:
     const infini_train::nn::parallel::ProcessGroup *pg_ = nullptr;
     std::vector<Device> target_gpus_;
-    Device input_device_ = Device();
+    Device input_device_;
     int64_t dim_ = 0;
 };
 
@@ -52,7 +52,7 @@ class Gather : public autograd::Function {
 
 private:
     const infini_train::nn::parallel::ProcessGroup *pg_ = nullptr;
-    Device target_device_ = Device();
+    Device target_device_;
     std::vector<Device> input_gpus_;
     int64_t dim_ = 0;
     bool unsqueezed_scalar_ = false;
@@ -76,7 +76,7 @@ class Broadcast : public autograd::Function {
     const infini_train::nn::parallel::ProcessGroup *pg_ = nullptr;
     std::vector<Device> target_gpus_;
     int64_t num_inputs_ = 0;
-    Device input_device_ = Device();
+    Device input_device_;
 };
 
 class ReduceAddCoalesced : public autograd::Function {
@@ -95,7 +95,7 @@ class ReduceAddCoalesced : public autograd::Function {
 
 private:
     const infini_train::nn::parallel::ProcessGroup *pg_ = nullptr;
-    Device destination_ = Device();
+    Device destination_;
     std::vector<Device> target_gpus_;
     int64_t num_inputs_ = 0;
 };

infini_train/include/core/device_guard.h

@@ -11,6 +11,7 @@ namespace infini_train::core {
 class Stream;
 class BlasHandle;
 
+// Note(dcj): In the CPU backend, kD2D corresponds to a regular memcpy.
 enum class MemcpyKind : int8_t {
     kH2D = 0,
     kD2H = 1,
@@ -161,7 +162,7 @@ class DeviceGuardImplRegistry {
     DeviceGuardImpl *Get(Device::DeviceType type) const;
 
 private:
-    DeviceGuardImplRegistry();
+    DeviceGuardImplRegistry() = default;
     DeviceGuardImplRegistry(const DeviceGuardImplRegistry &) = delete;
     DeviceGuardImplRegistry &operator=(const DeviceGuardImplRegistry &) = delete;
 

infini_train/include/nn/parallel/data_parallel.h

@@ -20,7 +20,7 @@ class DataParallel : public Module {
 private:
     int dim_ = 0;
     std::vector<Device> devices_;
-    Device output_device_ = Device();
-    Device src_device_ = Device();
+    Device output_device_;
+    Device src_device_;
 };
 } // namespace infini_train::nn::parallel

infini_train/include/nn/parallel/pp/pipeline_stage.h

@@ -42,7 +42,7 @@ class PipelineStage {
     int num_stages_ = -1;
     int prev_rank_ = -1;
     int next_rank_ = -1;
-    Device device_ = Device();
+    Device device_;
     std::vector<std::shared_ptr<Module>> chunks_;
     std::vector<std::vector<int64_t>> recv_shape_;
 };

infini_train/include/nn/parallel/work.h

@@ -58,7 +58,7 @@ class WorkNccl final : public Work {
     void SetException(std::exception_ptr e);
 
 private:
-    Device device_ = Device();
+    Device device_;
     cudaEvent_t ready_event_;
     cudaEvent_t done_event_;
     ncclComm_t comm_;

infini_train/include/tensor.h

@@ -48,7 +48,7 @@ class TensorBuffer {
     size_t Size() const;
 
 private:
-    Device device_ = Device();
+    Device device_;
     size_t size_ = 0;
     void *data_ = nullptr;
 };

infini_train/src/core/cpu/cpu_guard.cc …ini_train/src/core/cpu/cpu_guard_impl.ccinfini_train/src/core/cpu/cpu_guard.cc renamed to infini_train/src/core/cpu/cpu_guard_impl.cc infini_train/src/core/cpu/cpu_guard.cc renamed to infini_train/src/core/cpu/cpu_guard_impl.cc

@@ -1,8 +1,12 @@
-#include "infini_train/src/core/cpu/cpu_guard.h"
+#include "infini_train/src/core/cpu/cpu_guard_impl.h"
 
 #include <cstdlib>
 #include <cstring>
 
+#include "glog/logging.h"
+
+#include "infini_train/include/core/device_guard.h"
+
 namespace infini_train::core::cpu {
 
 CpuGuardImpl::CpuGuardImpl() {}
@@ -15,6 +19,13 @@ void CpuGuardImpl::Malloc(void **dev_ptr, size_t size) { *dev_ptr = std::malloc(
 
 void CpuGuardImpl::Free(void *dev_ptr) { std::free(dev_ptr); }
 
-void CpuGuardImpl::Memcpy(void *dst, const void *src, size_t count, MemcpyKind kind) { std::memcpy(dst, src, count); }
+void CpuGuardImpl::Memcpy(void *dst, const void *src, size_t count, MemcpyKind kind) {
+    CHECK(kind == MemcpyKind::kD2D) << "CpuGuardImpl::Memcpy only supports kD2D (host-to-host) memcpy, "
+                                    << "but got MemcpyKind=" << static_cast<int>(kind);
+
+    std::memcpy(dst, src, count);
+}
+
+INFINI_TRAIN_REGISTER_DEVICE_GUARD_IMPL(Device::DeviceType::kCPU, CpuGuardImpl)
 
 } // namespace infini_train::core::cpu

infini_train/src/core/cpu/cpu_guard.h …fini_train/src/core/cpu/cpu_guard_impl.hinfini_train/src/core/cpu/cpu_guard.h renamed to infini_train/src/core/cpu/cpu_guard_impl.h infini_train/src/core/cpu/cpu_guard.h renamed to infini_train/src/core/cpu/cpu_guard_impl.h

@@ -1,4 +1,4 @@
-#include "infini_train/src/core/cuda/cuda_guard.h"
+#include "infini_train/src/core/cuda/cuda_guard_impl.h"
 
 #include <array>
 #include <cstdint>

infini_train/src/core/cuda/cuda_guard.cc …i_train/src/core/cuda/cuda_guard_impl.ccinfini_train/src/core/cuda/cuda_guard.cc renamed to infini_train/src/core/cuda/cuda_guard_impl.cc infini_train/src/core/cuda/cuda_guard.cc renamed to infini_train/src/core/cuda/cuda_guard_impl.cc

@@ -2,11 +2,14 @@
 
 #include <cstdint>
 
-#include "infini_train/include/core/blas_handle.h"
 #include "infini_train/include/core/device_guard.h"
-#include "infini_train/include/core/stream.h"
 #include "infini_train/include/device.h"
 
+namespace infini_train::core {
+class Stream;
+class BlasHandle;
+} // namespace infini_train::core
+
 namespace infini_train::core::cuda {
 
 class CudaGuardImpl : public DeviceGuardImpl {