Fix some typos (RebuildGruops, primtive, etc.) (#61762)

co63oc · web-flow · commit c2944293b219 · 2024-02-19T16:28:34.000+08:00
diff --git a/paddle/fluid/imperative/reducer.cc b/paddle/fluid/imperative/reducer.cc
@@ -884,7 +884,7 @@ void Reducer::FusedAllReduceSchedule(const int run_order,
   }
 }
 
-std::vector<std::vector<size_t>> Reducer::RebuildGruops() {
+std::vector<std::vector<size_t>> Reducer::RebuildGroups() {
   VLOG(3) << "The order of parameter arrival: "
           << string::join_strings(rebuild_var_indices_, ',');
 
@@ -960,7 +960,7 @@ void Reducer::ProcessUnusedDenseVars() {
       auto grad_var_base_tmp = dest_var_base->MutableGradVarBase();
       // NOTE(haohongxiang): Calling SetIsEmpty here is to make sure that
       // gradient accumulation can continue normally after clear_gradients()
-      // especiall in cases including complex control flow.
+      // especially in cases including complex control flow.
       grad_var_base_tmp->SharedVar()->SetIsEmpty(false);
 
       // 4. set grad tensor
@@ -1013,13 +1013,13 @@ void Reducer::FinalizeBackward() {
 
   if (NeedRebuildGroup()) {
     VLOG(3) << "Start rebuilding the groups";
-    auto rebuild_group_indices = RebuildGruops();
+    auto rebuild_group_indices = RebuildGroups();
     group_indices_ = std::move(rebuild_group_indices);
     InitializeGroups(group_indices_);
   }
 
   if (find_unused_vars_each_step_) {
-// TODO(liuyuhui) support xpu about Tensorcopy/TensorFromVector/TensorToVector
+// TODO(liuyuhui) support xpu about TensorCopy/TensorFromVector/TensorToVector
 #if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL) || \
     defined(PADDLE_WITH_GLOO)
     ProcessUnusedDenseVars();
diff --git a/paddle/fluid/imperative/reducer.h b/paddle/fluid/imperative/reducer.h
@@ -160,7 +160,7 @@ class Reducer {
 
   void FinalizeBackward();
 
-  std::vector<std::vector<size_t>> RebuildGruops();
+  std::vector<std::vector<size_t>> RebuildGroups();
 
   inline bool NeedRebuildGroup() {
     return !has_rebuilt_group_ && !find_unused_vars_each_step_;
diff --git a/python/paddle/distributed/auto_parallel/static/completion.py b/python/paddle/distributed/auto_parallel/static/completion.py
@@ -1150,7 +1150,7 @@ def set_process_mesh(block, op, process_mesh, var_to_process_mesh):
         num_chunks = pp_degree * vpp_degree
         assert (
             len(seg_op_deps) % num_chunks == 0
-        ), "The number of layers[{}] ({}) should be devided by part number ({}).".format(
+        ), "The number of layers[{}] ({}) should be divided by part number ({}).".format(
             seg_method, len(seg_op_deps), num_chunks
         )
 
@@ -1621,7 +1621,7 @@ def _get_op_by_id(ops, id):
                     continue
 
                 else:
-                    raise ValueError(f"got unexpect op [{str(grad_op.type)}]")
+                    raise ValueError(f"got unexpected op [{str(grad_op.type)}]")
 
                 self._dist_context.set_op_dist_attr_for_program(
                     grad_op, grad_op_dist_attr
@@ -1818,9 +1818,9 @@ def infer_backward_op_partial_status(
                         f"Backward Partial is not adapted for {str(grad_op)}"
                     )
 
-                # resulote partial
+                # resolute partial
                 # NOTE We set the Partial status in op_dist_attr instead tensor_dist_attr
-                # since the Partial will be reshard as Replicated immedidately after op output in static mode.
+                # since the Partial will be reshard as Replicated immediately after op output in static mode.
                 if len(param_grads) > 0:
                     activation_grad_dims_mapping = (
                         grad_op_dist_attr.get_input_dims_mapping(
@@ -2059,7 +2059,7 @@ def infer_backward_op_partial_status(
                         grad_op, grad_op_dist_attr
                     )
                 else:
-                    raise ValueError(f"got unexpect op [{str(grad_op.type)}]")
+                    raise ValueError(f"got unexpected op [{str(grad_op.type)}]")
 
     def complete_update_annotation(self, serial_main_program):
         """Complete the annotation of vars and ops in the update phase for parallel program."""
diff --git a/python/paddle/distributed/auto_parallel/static/engine.py b/python/paddle/distributed/auto_parallel/static/engine.py
@@ -1816,7 +1816,7 @@ def _validate_batch(self, batch):
             return [None]
 
         if self._strategy.pipeline.enable or self._acc_steps == 1:
-            # pp with schedule or navie-pp
+            # pp with schedule or naive-pp
             return batch
         else:
             # split feed data with gradient_merge k_steps
@@ -2073,7 +2073,7 @@ def cost(self, inputs_spec=None, labels_spec=None, mode=None):
         # Check parallel mode
         if self._strategy.auto_mode == "full":
             self._logger.info(
-                "The cost will be calcudated in the search process when the auto mode is full."
+                "The cost will be calculated in the search process when the auto mode is full."
             )
             return
 
diff --git a/python/paddle/distributed/auto_parallel/static/operators/dist_default.py b/python/paddle/distributed/auto_parallel/static/operators/dist_default.py
@@ -61,7 +61,7 @@ def prim_operator_data_parallel_functor(ctx, src_op):
     if var_name in ctx.grads_params:
         assert (
             var_name not in ctx.synced_gradient
-        ), f"in primtive mode, grad is already {var_name} synced"
+        ), f"in primitive mode, grad is already {var_name} synced"
         ctx.synced_gradient.add(var_name)
         sync_group = new_process_group(ctx.data_parallel_group)
 
@@ -554,7 +554,7 @@ def forward(ctx, *args, **kwargs):
                         )
             dst_op.desc._set_attr('shape', shape_list)
 
-        # data parallel synchronization for primtive operators
+        # data parallel synchronization for primitive operators
         from paddle.incubate.autograd import prim_enabled
 
         if prim_enabled():
diff --git a/python/paddle/distributed/auto_parallel/static/reshard.py b/python/paddle/distributed/auto_parallel/static/reshard.py
@@ -1113,7 +1113,7 @@ def __init__(
         self._has_sent = {}
         self._has_recv = {}
         self._has_allgather = {}
-        # to avoid reshard repeatly
+        # to avoid reshard repeatedly
         self._has_resharded = {}
 
     @property
diff --git a/python/paddle/distributed/auto_parallel/static/tuner/algorithms.py b/python/paddle/distributed/auto_parallel/static/tuner/algorithms.py
@@ -58,7 +58,7 @@ def collect_model_info(self, main_prog, startup_prog):
         Collect the model static info (from programs) that could be used to
         pruning candidate trials and saving tuning time. For instance,
         model info like number of model parameters and activation memory could be
-        used to prune candidated trial and decide the next trial.
+        used to prune candidate trial and decide the next trial.
         """
         pass
 
@@ -158,7 +158,7 @@ def update(self, results):
 
 
 @register_algor("recompute")
-class ReccomputeCheckpointAlgorithm(AlgorithmBase):
+class RecomputeCheckpointAlgorithm(AlgorithmBase):
     def __init__(self, config):
         super().__init__(config)
         self._changed_configs = ["recompute"]
diff --git a/python/paddle/distributed/auto_parallel/static/tuner/rule_based_tuner.py b/python/paddle/distributed/auto_parallel/static/tuner/rule_based_tuner.py
@@ -250,7 +250,7 @@ def build(self):
 
 
 @register_pattern
-class FFNPattrern(BasePattern):
+class FFNPattern(BasePattern):
     """FFN pattern defined by GPT model in PaddleFleetX."""
 
     name = "ffn"
@@ -971,7 +971,7 @@ def partition_cluster(
         ],
     ) -> list:
         """
-        Partiton cluster into possible device meshes.
+        Partition cluster into possible device meshes.
         Args:
             n (int): The number of nodes.
             m (int): The number of single devices on each node.
@@ -1044,7 +1044,7 @@ class RuleBasedTuner:
         mode (str): The mode of current task, it can be train or eval. Default: train.
         level (str): The level of this tuner, it can be o1 or o2.
                      o2 level may find better strategy but need more time than o1.
-                     If level is o1, it means all layers within same parallelism and place layers evenly when in pipeline parallism.
+                     If level is o1, it means all layers within same parallelism and place layers evenly when in pipeline parallelism.
                      If level is o2, it means layers can has own parallelism and place layers may not evenly.
                      Default: o1.
     """
@@ -1363,8 +1363,8 @@ def _gen_fwd_sub_program_by_clone(self, ops):
 
         return program
 
-    def _compelte_sub_fwd_program(self, idx, sub_fwd_program, process_mesh):
-        """Compelete forward sub  program."""
+    def _complete_sub_fwd_program(self, idx, sub_fwd_program, process_mesh):
+        """Complete forward sub  program."""
         selective_parallelisms = (
             ["dp", "mp"] if len(process_mesh.shape) == 1 else ["dp_mp", "mp_dp"]
         )
@@ -1424,7 +1424,7 @@ def complete_sub_fwd_programs(self, process_mesh):
             sub_fwd_program = self.fwd_sub_programs[idx]
             if idx not in self.sub_programs_dist_context:
                 self.sub_programs_dist_context[idx] = {}
-            self._compelte_sub_fwd_program(idx, sub_fwd_program, process_mesh)
+            self._complete_sub_fwd_program(idx, sub_fwd_program, process_mesh)
 
     def _complete_sub_bwd_program(self, sub_program_dist_context):
         """
@@ -1926,7 +1926,7 @@ def _local_stage_pass(self, start, end, process_mesh):
             ]
 
         # Use beam search, the beam size is 2.
-        # When the process mesh is 1-D, the selecetive parallelsim can be dp or mp.
+        # When the process mesh is 1-D, the selective parallelism can be dp or mp.
         # Because the first layer often contains more ops than other layer, using beam search can find more accurate strategy.
         count = 0
         for dist_context_x in dist_contexts_x:
@@ -2274,7 +2274,7 @@ def tune_o1(self):
 
             # For example, device_mesh is [1, 8] and process_mesh is [8].
             # The selective parallelism is dp or mp
-            # Get dp8 or mp8 cost and compare them to get best sreategy.
+            # Get dp8 or mp8 cost and compare them to get best strategy.
             for parallelism in ["dp", "mp", "dp_mp", "mp_dp"]:
                 for process_mesh_shape in process_mesh_shapes:
                     dist_context_of_device_meshes = None
@@ -2421,7 +2421,7 @@ def tune(self):
 
         elif self.level == "o1":
             # If level is o1, it means all layers within same parallelism.
-            # When in pipeline parallism, it means that place layers evenly.
+            # When in pipeline parallelism, it means that place layers evenly.
             use_o2_level = False
             for device_meshes in self.device_meshes_list:
                 if len(device_meshes) > 1:
@@ -2433,7 +2433,7 @@ def tune(self):
                         else:
                             if shape != device_mesh.shape:
                                 self._logger.info(
-                                    "Warning: The o1 level is not be supported when the number of machines is prime numer which greaters than 1. We will use o2 level to tune."
+                                    "Warning: The o1 level is not be supported when the number of machines is prime number which greater than 1. We will use o2 level to tune."
                                 )
                                 use_o2_level = True
                                 break
diff --git a/python/paddle/distributed/auto_parallel/static/utils.py b/python/paddle/distributed/auto_parallel/static/utils.py
@@ -1012,7 +1012,7 @@ def _merge_parameter(
     partition_param_list, param, partition_index, complete_shape
 ):
     """
-    Merge partitial parameters to a complete one.
+    Merge partial parameters to a complete one.
 
     Returns:
         None
@@ -1355,7 +1355,7 @@ def update_op_dims_mapping_by_default_dist_impl(dist_op):
     changed = False
     op_dist_attr = dist_op.dist_attr
     op_desc = dist_op.serial_op.desc
-    # The following statement will be replaced by a more elegent way
+    # The following statement will be replaced by a more elegant way
     if op_desc.type() == "shape" or op_desc.type() == "slice":
         return False
     output_names = op_desc.output_names()
diff --git a/python/paddle/distributed/checkpoint/load_state_dict.py b/python/paddle/distributed/checkpoint/load_state_dict.py
@@ -67,7 +67,7 @@ def get_rank_to_files(path, state_dict, process_group, use_dist):
     Get the mapping of rank to its accessible files.
     """
     metadata_files, local_data_files = get_checkpoint_files(path)
-    # The neccesary files to be read
+    # The necessary files to be read
     tensor_key_list = []
     necessary_files = []
     for metadata_file in metadata_files:
@@ -102,7 +102,7 @@ def get_rank_to_files(path, state_dict, process_group, use_dist):
     logger.debug(
         f"necessary_data_files_set:{necessary_data_files_set}, global_data_files_set:{global_data_files_set}"
     )
-    # check neccesary files in global_data_files
+    # check necessary files in global_data_files
     assert (
         global_data_files_set & necessary_data_files_set
         == necessary_data_files_set
@@ -251,14 +251,14 @@ def compute_overlap(
     cur_offsets = []
     storage_offsets = []
     lengths = []
-    for cur_len, cur_offset, strorage_len, storage_offset in zip(
+    for cur_len, cur_offset, storage_len, storage_offset in zip(
         cur_chunk_metadata.local_shape,
         cur_chunk_metadata.global_offset,
         storage_local_tensor_metadata.local_shape,
         storage_local_tensor_metadata.global_offset,
     ):
         begin_offset = max(cur_offset, storage_offset)
-        end_offset = min(cur_offset + cur_len, storage_offset + strorage_len)
+        end_offset = min(cur_offset + cur_len, storage_offset + storage_len)
         if begin_offset == cur_offset:
             cur_offsets.append(0)
             storage_offsets.append(begin_offset - storage_offset)
@@ -280,14 +280,14 @@ def not_overlap(
     cur_chunk_metadata: LocalTensorMetadata,
     storage_local_tensor_metadata: LocalTensorMetadata,
 ):
-    for cur_len, cur_offset, strorage_len, storage_offset in zip(
+    for cur_len, cur_offset, storage_len, storage_offset in zip(
         cur_chunk_metadata.local_shape,
         cur_chunk_metadata.global_offset,
         storage_local_tensor_metadata.local_shape,
         storage_local_tensor_metadata.global_offset,
     ):
         if (
-            cur_offset >= (storage_offset + strorage_len)
+            cur_offset >= (storage_offset + storage_len)
             or (cur_offset + cur_len) <= storage_offset
         ):
             return True
diff --git a/python/paddle/distributed/checkpoint/save_state_dict.py b/python/paddle/distributed/checkpoint/save_state_dict.py
@@ -27,12 +27,12 @@
 
 def check_state_dict(state_dict, process_group):
     local_keys = list(state_dict.keys())
-    gloabl_keys = []
-    paddle.distributed.all_gather_object(gloabl_keys, local_keys, process_group)
-    for keys in gloabl_keys[1:]:
+    global_keys = []
+    paddle.distributed.all_gather_object(global_keys, local_keys, process_group)
+    for keys in global_keys[1:]:
         assert (
-            keys == gloabl_keys[0]
-        ), f"keys:{keys} != first_keys: {gloabl_keys[0]}"
+            keys == global_keys[0]
+        ), f"keys:{keys} != first_keys: {global_keys[0]}"
 
 
 def check_file_name(file_name, process_group):
