!7946 fix auto parallel optimizer weight shard

Merge pull request !7946 from gziyan/fix_auto_optim_shard

mindspore/ccsrc/frontend/parallel/ops_info/operator_info.h

@@ -164,8 +164,6 @@ class OperatorInfo {
   const std::unordered_map<std::string, ValuePtr> &attrs() const { return attrs_; }
   void set_stage_id(int32_t stage_id) { stage_id_ = stage_id; }
   int32_t stage_id() const { return stage_id_; }
-  void set_opt_shard_flag(bool flag) { opt_shard_flag_ = flag; }
-  bool opt_shard_flag() { return opt_shard_flag_; }
   Status CreateGroupByTensorMap(const Shape &tensor_map, std::vector<Group> *group);
 
   // Key for user data.
@@ -269,7 +267,6 @@ class OperatorInfo {
  private:
   OperatorCostPtr operator_cost_;
   std::vector<TypePtr> outputs_type_;
-  bool opt_shard_flag_ = false;
 };
 
 Shape GetSliceShape(const Shape &tensor_shape, const Dimensions &strategy);

mindspore/ccsrc/frontend/parallel/step_parallel.cc

@@ -808,11 +808,17 @@ void InsertVirtualDivOp(const VirtualDivOp &virtual_div_op, const CNodePtr &node
   }
 }
 
+// Only used for InsertMirrorOps
 std::pair<AnfNodePtr, bool> FindParameter(const AnfNodePtr &node, const FuncGraphPtr &func_graph) {
   if (!node->isa<Parameter>() && !node->isa<CNode>() && !node->isa<ValueNode>()) {
     return std::make_pair(nullptr, false);
   } else if (node->isa<Parameter>()) {
+    auto param_ptr = node->user_data<parallel::TensorLayout>();
+    if (param_ptr != nullptr && !param_ptr->opt_shard_group().empty()) {
+      return std::make_pair(nullptr, false);
+    } else {
       return std::make_pair(node, false);
+    }
   } else if (node->isa<ValueNode>()) {
     if (IsValueNode<RefKey>(node)) {
       std::vector<AnfNodePtr> param_v = FindParameterByRefKeyNode(node, func_graph);
@@ -820,8 +826,13 @@ std::pair<AnfNodePtr, bool> FindParameter(const AnfNodePtr &node, const FuncGrap
         MS_LOG(EXCEPTION) << "FindParameterByRefKeyNode failed, return vector size must be 1, real is  "
                           << param_v.size();
       }
+      auto param_ptr = param_v[0]->user_data<parallel::TensorLayout>();
+      if (param_ptr != nullptr && !param_ptr->opt_shard_group().empty()) {
+        return std::make_pair(nullptr, true);
+      } else {
         return std::make_pair(node, true);
       }
+    }
     return std::make_pair(nullptr, false);
   } else {
     CNodePtr cnode = node->cast<CNodePtr>();
@@ -1002,7 +1013,7 @@ void BackwardCommunication(const OperatorInfoPtr &distribute_operator, const CNo
   MirrorOps mirror_ops = distribute_operator->mirror_ops();
   VirtualDivOp virtual_div_op = distribute_operator->virtual_div_op();
   // insert mirror op
-  if (!mirror_ops.empty() && !distribute_operator->opt_shard_flag()) {
+  if (!mirror_ops.empty()) {
     MS_LOG(INFO) << "insert mirror op for " << distribute_operator->name();
     InsertMirrorOps(mirror_ops, node);
   }
@@ -1374,39 +1385,51 @@ std::pair<AnfNodePtr, int> FindSubGraph(const FuncGraphPtr &graph, const AnfNode
   return std::make_pair(nullptr, 0);
 }
 
-void ApplyParallelOptOnParam(TensorLayout *tensor_layout, const OperatorInfoPtr &distribute_operator,
+void InsertAllGatherOp(const std::string &group, const std::pair<AnfNodePtr, int> &res, const AnfNodePtr &parameter) {
-                             const CNodePtr &cnode, const AnfNodePtr &parameter, size_t index) {
+  Operator op = CreateAllGatherOp(group);
-  MS_EXCEPTION_IF_NULL(distribute_operator);
+  MS_EXCEPTION_IF_NULL(res.first);
-  MS_EXCEPTION_IF_NULL(cnode);
   MS_EXCEPTION_IF_NULL(parameter);
-  std::vector<Group> dev_group;
+  auto cnode = res.first->cast<CNodePtr>();
-  // create communication group for allgather operator
-  if (distribute_operator->CreateGroupByTensorMap(tensor_layout->origin_tensor_map().array(), &dev_group) ==
-        Status::SUCCESS &&
-      !dev_group.empty()) {
-    // set optimizer shard split flag to avoid inserting mirror_ops
-    distribute_operator->set_opt_shard_flag(true);
-    // insert allgather operator between shard parameter and cnode
-    Operator op = CreateAllGatherOp(dev_group[0].name());
   auto graph = cnode->func_graph();
   MS_EXCEPTION_IF_NULL(graph);
-    InsertNode(op, cnode, index, parameter, graph, PARALLEL_OPTIMIZER_ALLGATHER);
+  InsertNode(op, cnode, res.second, parameter, graph, PARALLEL_OPTIMIZER_ALLGATHER);
-    // set communication group in tensor layout for checkpoint saving
-    tensor_layout->set_opt_shard_group(dev_group[0].name());
   // add fusion flag
-    auto allgather = cnode->input(index)->cast<CNodePtr>();
+  auto allgather = cnode->input(res.second)->cast<CNodePtr>();
   auto prim = GetValueNode<PrimitivePtr>(allgather->input(0));
   auto attrs = prim->attrs();
   // enable fusion flag later when it's supported in backend
   attrs["fusion"] = MakeValue(0);
   prim->SetAttrs(attrs);
-    MS_LOG(INFO) << "Parallel optimizer is applied on " << parameter->ToString();
+}
-  } else {
+
-    MS_LOG(ERROR) << "Parallel optimizer applied on " << parameter->ToString() << "failed!";
+void ApplyParallelOptOnParam(const FuncGraphPtr &root, const AnfNodePtr &parameter,
+                             const std::string &opt_shard_group) {
+  if (opt_shard_group.empty()) {
+    return;
+  }
+  FuncGraphManagerPtr manager = root->manager();
+  MS_EXCEPTION_IF_NULL(manager);
+  auto param_sub_set = manager->node_users()[parameter];
+  for (auto &param_pair : param_sub_set) {
+    auto cnode = param_pair.first->cast<CNodePtr>();
+    MS_EXCEPTION_IF_NULL(cnode);
+    if (cnode->in_forward_flag()) {
+      OperatorInfoPtr distribute_operator = cnode->user_data<OperatorInfo>();
+      if (distribute_operator == nullptr) {
+        MS_LOG(WARNING) << "Parallel optimizer: " << cnode->ToString() << " 's OperatorInfoPtr is nullptr";
+      } else if (IntToSize(param_pair.second - 1) >= distribute_operator->inputs_tensor_info().size()) {
+        MS_LOG(EXCEPTION) << "The index is out of range, index is  " << param_pair.second - 1 << ", vector size is  "
+                          << distribute_operator->inputs_tensor_info().size();
+      }
+      // insert allgather operator between shard parameter and cnode
+      InsertAllGatherOp(opt_shard_group, param_pair, parameter);
+      MS_LOG(INFO) << "Parallel optimizer is applied between " << parameter->ToString() << " and " << cnode->ToString();
+    }
   }
 }
 
-void SetParallelShape(const AnfNodePtr &parameter, const std::pair<AnfNodePtr, int> &res) {
+// When this function returns non-empty string, that means parallel optimizer is applied on this parameter.
+std::string SetParallelShape(const AnfNodePtr &parameter, const std::pair<AnfNodePtr, int> &res) {
   MS_EXCEPTION_IF_NULL(parameter);
   AbstractBasePtr abstract = parameter->abstract();
   MS_EXCEPTION_IF_NULL(abstract);
@@ -1417,26 +1440,40 @@ void SetParallelShape(const AnfNodePtr &parameter, const std::pair<AnfNodePtr, i
   if (distribute_operator == nullptr) {
     MS_LOG(EXCEPTION) << "Failure:node " << cnode->ToString() << " 's OperatorInfoPtr is nullptr";
   }
-
   if (IntToSize(res.second - 1) >= distribute_operator->inputs_tensor_info().size()) {
     MS_LOG(EXCEPTION) << "The index is out of range, index is  " << res.second - 1 << ", vector size is  "
                       << distribute_operator->inputs_tensor_info().size();
   }
   TensorInfo tensorinfo_in = distribute_operator->inputs_tensor_info()[IntToSize(res.second - 1)];
   TensorLayout tensor_layout = tensorinfo_in.tensor_layout();
+  Shape slice_shape = tensor_layout.slice_shape().array();
+  std::string opt_shard_group;
   MS_EXCEPTION_IF_NULL(ParallelContext::GetInstance());
   bool enable_parallel_optimizer = ParallelContext::GetInstance()->enable_parallel_optimizer();
-  Shape slice_shape = tensor_layout.slice_shape().array();
   if (enable_parallel_optimizer) {
     if (!ParameterRequireGrad(parameter)) {
       // only trainable parameters need parallel optimizer
-      MS_LOG(INFO) << "Parallel optimizer is no need for " << parameter->ToString();
+      MS_LOG(INFO) << "Parallel optimizer: " << parameter->ToString() << " is not trainable parameter.";
     } else if (tensor_layout.GenerateOptShardSliceShape() == Status::SUCCESS) {
       // get a totally shard tensor slice shape if the weight is repeated on devices
       // and the shape of the first dimension could be divided
       // apply parallel optimizer on parameters
-      ApplyParallelOptOnParam(&tensor_layout, distribute_operator, cnode, parameter, IntToSize(res.second));
+      // create communication group for allgather operator
       slice_shape = tensor_layout.opt_shard_slice_shape();
+      std::vector<Group> dev_group;
+      if (distribute_operator->CreateGroupByTensorMap(tensor_layout.origin_tensor_map().array(), &dev_group) ==
+            Status::SUCCESS &&
+          !dev_group.empty()) {
+        opt_shard_group = dev_group[0].name();
+        // set communication group in tensor layout for checkpoint saving
+        tensor_layout.set_opt_shard_group(opt_shard_group);
+        MS_LOG(INFO) << "Parallel optimizer: create group " << opt_shard_group << " for " << parameter->ToString()
+                     << " success.";
+      } else {
+        MS_LOG(WARNING) << "Parallel optimizer: create group for " << parameter->ToString() << " failed.";
+      }
+    } else {
+      MS_LOG(INFO) << "Parallel optimizer: " << parameter->ToString() << "'s shape does not satisfy the conditions.";
     }
   }
   MS_LOG(INFO) << "SetParallelShape slice_shape  " << parameter->ToString() << "  shape "
@@ -1451,6 +1488,7 @@ void SetParallelShape(const AnfNodePtr &parameter, const std::pair<AnfNodePtr, i
   ParameterPtr parameter_ptr = parameter->cast<ParameterPtr>();
   MS_EXCEPTION_IF_NULL(parameter_ptr);
   parameter_ptr->set_user_data<TensorLayout>(std::make_shared<TensorLayout>(tensor_layout));
+  return opt_shard_group;
 }
 
 void CoverSliceShape(const FuncGraphPtr &root) {
@@ -1460,14 +1498,18 @@ void CoverSliceShape(const FuncGraphPtr &root) {
     MS_EXCEPTION_IF_NULL(parameter->Shape());
     auto iter = g_RefMap.find(parameter);
     if (iter != g_RefMap.end()) {
-      SetParallelShape(parameter, g_RefMap[parameter]);
+      std::string group = SetParallelShape(parameter, g_RefMap[parameter]);
+      // find all forward nodes that use parameter in graphs and insert allgather if group is not empty
+      ApplyParallelOptOnParam(root, parameter, group);
       continue;
     }
     std::pair<AnfNodePtr, int> res = FindSubGraph(root, parameter);
     if (res.first == nullptr) {
       MS_LOG(INFO) << "Parameter " << parameter->ToString() << " don't need to set parallel shape";
     } else {
-      SetParallelShape(parameter, res);
+      std::string group = SetParallelShape(parameter, res);
+      // find all forward nodes that use parameter in graphs and insert allgather if group is not empty
+      ApplyParallelOptOnParam(root, parameter, group);
       MS_LOG(DEBUG) << "Parameter " << parameter->ToString() << " shape " << parameter->Shape()->ToString();
     }
   }

mindspore/ccsrc/frontend/parallel/step_parallel.h

@@ -109,7 +109,7 @@ std::vector<Shapes> ExtractShape(const CNodePtr &node);
 std::pair<AnfNodePtr, int> FindSubGraph(const FuncGraphPtr &func_graph, const AnfNodePtr &parameter);
 
 // Set distribute shape for parameters abstract
-void SetParallelShape(const AnfNodePtr &parameter, const std::pair<AnfNodePtr, int> &res);
+std::string SetParallelShape(const AnfNodePtr &parameter, const std::pair<AnfNodePtr, int> &res);
 
 // change parameters'shape in resource
 void CoverSliceShape(const FuncGraphPtr &root);

mindspore/common/parameter.py

@@ -418,9 +418,9 @@ class Parameter(MetaTensor_):
                 if _is_role_worker():
                     data = self.init_mode.to_tensor(0, [1])
                 else:
-                    data = self.init_mode.to_tensor(slice_index, layout[2])
+                    data = self.init_mode.to_tensor(slice_index, layout[2], layout[5])
             else:
-                data = self.init_mode.to_tensor(slice_index, layout[2])
+                data = self.init_mode.to_tensor(slice_index, layout[2], layout[5])
         else:
             if (self.init_in_server and self.is_param_ps and isinstance(self.init_mode, MetaTensor)):
                 if _is_role_worker():

mindspore/common/tensor.py

@@ -16,6 +16,7 @@
 import numpy as np
 
 from mindspore import log as logger
+from mindspore.communication.management import get_rank, get_group_size
 from .._c_expression import Tensor as Tensor_
 from .._c_expression import MetaTensor as MetaTensor_
 from .._checkparam import check_type, check_typename
@@ -409,7 +410,7 @@ class MetaTensor(MetaTensor_):
         self.init = init
         MetaTensor_.__init__(self, dtype, shape)
 
-    def to_tensor(self, slice_index=None, shape=None):
+    def to_tensor(self, slice_index=None, shape=None, opt_shard_group=None):
         """
         Get the tensor format data of this MetaTensor.
 
@@ -418,6 +419,8 @@ class MetaTensor(MetaTensor_):
                 It is used when initialize a slice of a parameter, it guarantees that devices
                 using the same slice can generate the same tensor.
             shape (list[int]): Shape of the slice, it is used when initialize a slice of the parameter.
+            opt_shard_group(str): Optimizer shard group which is used in auto or semi auto parallel mode
+                to get one shard of a parameter's slice.
         """
         if self.init is None:
             raise TypeError("to_dense must be set MetaTensor.init, init can't be None")
@@ -453,7 +456,12 @@ class MetaTensor(MetaTensor_):
 
         with seed_context(self.init):
             self.init(arr)
-        return Tensor(arr, dtype=self.dtype)
+        data = np.array(arr)
+        if opt_shard_group:
+            rank = get_rank(opt_shard_group)
+            size = get_group_size(opt_shard_group)
+            data = np.split(data, size)[rank]
+        return Tensor(data, dtype=self.dtype)
 
 
 def _vm_compare(*args):

mindspore/parallel/_cell_wrapper.py

@@ -16,7 +16,7 @@
 
 from mindspore.nn.cell import Cell
 from mindspore.ops.operations.comm_ops import AllGather
-
+from mindspore.communication import GlobalComm
 
 _allgather_cell = None
 
@@ -26,10 +26,10 @@ class AllGatherCell(Cell):
     Allgather cell, used in model parallel scenario.
     To allgather the selected parameter slice from each device.
     """
-    def __init__(self):
+    def __init__(self, group):
         super(AllGatherCell, self).__init__(auto_prefix=False)
 
-        self.allgather = AllGather()
+        self.allgather = AllGather(group)
 
     def construct(self, x):
         x = self.allgather(x)
@@ -58,13 +58,16 @@ class SaveOptShardCkptCell(Cell):
         return x
 
 
-def get_allgather_cell(group):
+def get_allgather_cell(group, need_merge_twice=False):
     """Get AllGatherCell object."""
     global _allgather_cell
-    if group:
+    if need_merge_twice:
         _allgather_cell = SaveOptShardCkptCell(group)
     else:
-        _allgather_cell = AllGatherCell()
+        if group:
+            _allgather_cell = AllGatherCell(group)
+        else:
+            _allgather_cell = AllGatherCell(GlobalComm.WORLD_COMM_GROUP)
     return _allgather_cell
 
 

mindspore/train/serialization.py

@@ -456,13 +456,18 @@ def _get_merged_param_data(net, param_name, param_data):
     # while any dim is not equal to -1, means param is split and needs to be merged
     # pipeline parallel need to be supported here later
     for dim in tensor_map:
-        if dim != -1 or opt_shard_group:
+        if dim != -1:
-            allgather_net = get_allgather_cell(opt_shard_group)
+            if opt_shard_group:
+                allgather_net = get_allgather_cell(opt_shard_group, True)
+            else:
+                allgather_net = get_allgather_cell(opt_shard_group, False)
             param_data = allgather_net(param_data)
             if field_size:
                 return _reshape_param_data_with_weight(param_data, dev_mat, field_size)
             return _reshape_param_data(param_data, dev_mat, tensor_map)
-
+    if opt_shard_group:
+        allgather_net = get_allgather_cell(opt_shard_group, False)
+        param_data = allgather_net(param_data)
     return param_data
 
 

tests/ut/python/parallel/test_parallel_optimizer.py

@@ -77,8 +77,9 @@ def auto_parallel_compile_net(mode, dev_num, strategy1=None, strategy2=None):
     train_network = TrainOneStepCell(net, optimizer)
     train_network.set_auto_parallel()
     train_network.set_train()
-    _executor.compile(train_network, inputs, label)
+    _executor.compile(train_network, inputs, label, phase="train", auto_parallel_mode=True)
     context.reset_auto_parallel_context()
+    return train_network
 
 
 def test_auto_parallel_momentum_1():
@@ -93,7 +94,11 @@ def test_auto_parallel_momentum_2():
 def test_auto_parallel_momentum_3():
     # hybrid parallel case
     # weight1 could not be shard and weight2 is repeated
-    auto_parallel_compile_net("semi_auto_parallel", 32, ((4, 8), (8, 1)), ((4, 4), (4, 2)))
+    train_network = auto_parallel_compile_net("semi_auto_parallel", 32, ((4, 8), (8, 1)), ((4, 4), (4, 2)))
+    param_dict = train_network.parameter_layout_dict
+    # validate opt_shard_group
+    assert not param_dict["weight1"][5]
+    assert param_dict["weight2"][5].startswith("4")
 
 
 def test_auto_parallel_momentum_4():