Fix

Author: co63oc

python/paddle/distributed/auto_parallel/static/cluster.py

@@ -741,7 +741,7 @@ def _build_from_dict(self, cluster_info):
  cluster_info.get("alpha_latency")
  )
  else:
- self._alpha_latecy = None
+ self._alpha_latency = None
 
  def build_from_file(self, json_file_path):
  with open(json_file_path) as json_file:

python/paddle/distributed/auto_parallel/static/completion.py

@@ -1643,7 +1643,7 @@ def _is_grad_var_name(name):
  def _get_forward_varname_from_grad_varname(grad_var_name):
  assert _is_grad_var_name(
  grad_var_name
- ), f"[{grad_var_name}] is not a grad varnme."
+ ), f"[{grad_var_name}] is not a grad var name."
  return grad_var_name[: grad_var_name.find("@GRAD")]
 
  def _get_op_by_id(ops, id):
@@ -1769,7 +1769,7 @@ def _complete_grad_op_with_forward_op(forward_op, grad_op, vars):
  def infer_backward_op_partial_status(
  vars, grad_op, grad_op_dist_attr
  ):
- # NOTE Since we use composite op in static mode which might have implicit Reduction of broadcast axes for caculating parameter's gradient.
+ # NOTE Since we use composite op in static mode which might have implicit Reduction of broadcast axes for calculating parameter's gradient.
  # Those implicit Reduction hinder the Partial inference in a normal way, and we need a special method to handle it.
  param_grads = []
  activation_grad = None
@@ -1993,7 +1993,7 @@ def infer_backward_op_partial_status(
  output_name, ref_fwd_dims_mapping
  )
  # NOTE(zhaoyingli):
- # The sum op is used to accmulate the grads' value of the same forward var,
+ # The sum op is used to accumulate the grads' value of the same forward var,
  # sum op's chunk_id is same with the last op which generate the grad.
  ref_chunk_id = None
  ref_process_mesh = None
@@ -2336,7 +2336,7 @@ def _init_global_mesh_for_program(self):
  assert dist_op is not None
  dist_op.dist_attr.process_mesh = ProcessMesh(world_ranks)
 
- # Find the most compatible implemenetations from the distributed operator
+ # Find the most compatible implementations from the distributed operator
  op_dist_impls = find_compatible_distributed_operator_impls(
  dist_op, fwd=True
  )

python/paddle/distributed/auto_parallel/static/dist_tensor.py

@@ -112,8 +112,8 @@ def get_local_offsets(
  global_sizes, dims_mapping, topology, processes, rank, shard_sizes
  )
  local_offsets = []
- rank_relatvie = processes.index(rank)
- coordinate = _linear_idx2coordinate(topology, rank_relatvie)
+ rank_relative = processes.index(rank)
+ coordinate = _linear_idx2coordinate(topology, rank_relative)
 
  for i in range(len(global_sizes)):
  if dims_mapping[i] == -1:

python/paddle/distributed/auto_parallel/static/helper.py

@@ -36,7 +36,7 @@
 class ProxyLayer(Layer):
  """
  ProxyLayer implements all logic for converting dygraph model into
- static Program IR. Meanwhile, it provides conviential interfaces for
+ static Program IR. Meanwhile, it provides conventional interfaces for
  auto parallel to visit feed/fetch/loss/metric variables.
  """
 

python/paddle/distributed/auto_parallel/static/utils.py

@@ -297,8 +297,8 @@ def _get_comm_group(processes, shape, axis, rank):
  assert (
  rank in processes
  ), f"rank [{rank}] is NOT in processes group {processes}"
- rank_relatvie = processes.index(rank)
- coordinate = _linear_idx2coordinate(shape, rank_relatvie)
+ rank_relative = processes.index(rank)
+ coordinate = _linear_idx2coordinate(shape, rank_relative)
  coordinates_in_group = [coordinate[:] for i in range(shape[axis])]
 
  # select comm group
@@ -328,8 +328,8 @@ def _get_idx_in_axis(processes, shape, axis, rank):
 
  # NOTE _linear_idx2coordinate assume processes mesh start with 0 and continuous
  # tricks to support processes mesh when it is not start with 0 or continuous
- rank_relatvie = processes.index(rank)
- coordinate = _linear_idx2coordinate(shape, rank_relatvie)
+ rank_relative = processes.index(rank)
+ coordinate = _linear_idx2coordinate(shape, rank_relative)
  return coordinate[axis]