!10145 LoadTensors was using a 4-byte size field instead of an 8-byte size field

From: @john_tzanakakis
Reviewed-by: @wenkai_dist,@mikef
Signed-off-by:

mindspore/ccsrc/debug/debug_services.cc

@@ -207,7 +207,7 @@ void DebugServices::CheckWatchpoints(std::vector<std::string> *name, std::vector
 }
 
 void DebugServices::ReadNodesTensors(std::vector<std::string> name, std::vector<std::string> *ret_name,
-                                     std::vector<char *> *data_ptr, std::vector<unsigned int> *data_size,
+                                     std::vector<char *> *data_ptr, std::vector<ssize_t> *data_size,
                                      std::vector<TypePtr> *dtype, std::vector<std::vector<int64_t>> *shape) {
   std::vector<std::tuple<std::string, std::shared_ptr<TensorData>>> result_list;
   tensor_loader_->SearchTensors(name, &result_list);

mindspore/ccsrc/debug/debug_services.h

@@ -180,8 +180,8 @@ class DebugServices {
                         const bool step_end, const bool recheck);
 
   void ReadNodesTensors(std::vector<std::string> name, std::vector<std::string> *ret_name,
-                        std::vector<char *> *data_ptr, std::vector<unsigned int> *data_size,
-                        std::vector<TypePtr> *dtype, std::vector<std::vector<int64_t>> *shape);
+                        std::vector<char *> *data_ptr, std::vector<ssize_t> *data_size, std::vector<TypePtr> *dtype,
+                        std::vector<std::vector<int64_t>> *shape);
 
   bool IsWatchPoint(const std::string &kernel_name, const CNodePtr &kernel = nullptr) const;
 

mindspore/ccsrc/debug/debugger/debugger.cc

@@ -736,7 +736,7 @@ std::list<TensorProto> Debugger::LoadTensors(const ProtoVector<TensorProto> &ten
   std::vector<std::string> name;
   std::vector<std::string> ret_name;
   std::vector<char *> data_ptr;
-  std::vector<unsigned int> data_size;
+  std::vector<ssize_t> data_size;
   std::vector<TypePtr> dtype;
   std::vector<std::vector<int64_t>> shape;
 
@@ -749,7 +749,7 @@ std::list<TensorProto> Debugger::LoadTensors(const ProtoVector<TensorProto> &ten
   unsigned int result_index = 0;
 
   for (auto tensor : tensors) {
-    int size_iter = 0;
+    ssize_t size_iter = 0;
     if (result_index >= ret_name.size() || ret_name[result_index] != GetTensorFullName(tensor)) {
       TensorProto tensor_item;
       tensor_item.set_finished(true);
@@ -757,9 +757,9 @@ std::list<TensorProto> Debugger::LoadTensors(const ProtoVector<TensorProto> &ten
       tensor_list.push_back(tensor_item);
       continue;
     }
-    int tensor_size = data_size[result_index];
+    ssize_t tensor_size = data_size[result_index];
     while (size_iter < tensor_size) {
-      int chunk_size = CHUNK_SIZE;
+      ssize_t chunk_size = CHUNK_SIZE;
       TensorProto tensor_item;
       tensor_item.set_finished(false);
       if (tensor_size - size_iter <= CHUNK_SIZE) {

mindspore/ccsrc/runtime/device/gpu/gpu_kernel_runtime.cc

@@ -131,6 +131,10 @@ void LoadKernelData(Debugger *debugger, const CNodePtr &kernel,
       std::string input_kernel_name = input_kernel->fullname_with_scope();
       auto addr = kernel_inputs[j];
       auto type = AnfAlgo::GetOutputInferDataType(input_kernel, PARAMETER_OUTPUT_INDEX);
+      // For example, this happens with the Depend op
+      if (type == kMetaTypeNone) {
+        continue;
+      }
       auto format = kOpFormat_DEFAULT;
       auto gpu_addr = std::make_unique<GPUDeviceAddress>(addr->addr, addr->size, format, type);
       string input_tensor_name = input_kernel_name + ':' + "0";
@@ -157,6 +161,10 @@ void LoadKernelData(Debugger *debugger, const CNodePtr &kernel,
     for (int j : real_outputs) {
       auto addr = kernel_outputs[j];
       auto type = AnfAlgo::GetOutputInferDataType(kernel, j);
+      // For example, this happens with the Depend op
+      if (type == kMetaTypeNone) {
+        continue;
+      }
       auto format = kOpFormat_DEFAULT;
       auto gpu_addr = std::make_unique<GPUDeviceAddress>(addr->addr, addr->size, format, type);
       string tensor_name = kernel_name + ':' + std::to_string(j);