Fix defition definition (#60679)

Author: co63oc

python/paddle/device/cuda/cuda_graphed_layer.py

@@ -66,9 +66,9 @@ def forward(ctx, context, *args):
  ctx.save_for_backward(context, context.args_static, y)
  return y.detach()
  else:
- for x_staic, x in zip(context.args_static, args):
- if isinstance(x_staic, paddle.Tensor):
- x_staic.copy_(x, True)
+ for x_static, x in zip(context.args_static, args):
+ if isinstance(x_static, paddle.Tensor):
+ x_static.copy_(x, True)
 
  context.forward_graph.replay()
  y = context.y_static

python/paddle/distributed/auto_parallel/interface.py

@@ -237,9 +237,9 @@ def __call__(self, *args, **kwargs):
 
 def exclude_ops_in_recompute(run_function):
  """
- Exclude some operators in recompute segements.
+ Exclude some operators in recompute segments.
  Args:
- run_function (callabe): The callabe function to be excluded.
+ run_function (callable): The callable function to be excluded.
 
  Returns:
  ExcludeOperator: The callable object.

python/paddle/distributed/auto_parallel/random.py

@@ -29,7 +29,7 @@
 _basic_seed = 42
 _basic_name = ""
 
-# use Prime number as offset to avoid confict
+# use Prime number as offset to avoid conflict
 _mesh_offset = 173
 _dim_offsets = [11, 23, 37, 73]
 
@@ -57,7 +57,7 @@ def parallel_manual_seed(seed, name=""):
 
  This function should be called only once before auto parallel compiles the computation graph (e.g. auto_parallel.engine.prepare() or fit()).
 
- This seed only affects how randomness-relative **operators** (dropout, fuse op with dropout inside, etc) are execute amonge mesh, and would NOT affect other processe like Parameter initialization.
+ This seed only affects how randomness-relative **operators** (dropout, fuse op with dropout inside, etc) are execute amonge mesh, and would NOT affect other process like Parameter initialization.
 
  Examples:
  # seed relative to training step
@@ -85,7 +85,7 @@ def determinate_rng(
  ), "Cannot provide dims mapping and placements at same time."
  # TODO(JZ-LIANG) Support Mesh with any high rank
  # use a string to unique integer hashing algorithm for seed computation.
- # instead of using offsets to coodinate seed across devices.
+ # instead of using offsets to coordinate seed across devices.
  if len(process_mesh.shape) > 4:
  raise NotImplementedError(
  "Auto Parallel Random Control for Mesh's rank > 4 is NOT supported! Got {}".format(
@@ -131,7 +131,7 @@ def determinate_rng(
  else:
  assert (
  seed_ not in _rng_name_to_seed.values()
- ), "Seed Confilt! current seed: {}, current sharding expr: {}, generated seed: {}".format(
+ ), "Seed Conflict! current seed: {}, current sharding expr: {}, generated seed: {}".format(
  seed_, sharding_expr, _rng_name_to_seed
  )
  _rng_name_to_seed[sharding_expr] = seed_

python/paddle/distributed/auto_parallel/static/cost/estimate_cost.py

@@ -24,7 +24,7 @@
 
 
 class CostEstimator:
- _sepical_op_type = ["fused_attention", "fused_feedforward"]
+ _special_op_type = ["fused_attention", "fused_feedforward"]
 
  def __init__(
  self, program, cluster, mode="modeling", rank=None, loop_count=10
@@ -207,7 +207,7 @@ def _estimate_core(self, dist_context, resharder, block):
 
  if dist_op_cost is None:
  assert (
- dist_op.serial_op.type in CostEstimator._sepical_op_type
+ dist_op.serial_op.type in CostEstimator._special_op_type
  )
  continue
  for item in dist_op_cost:

python/paddle/distributed/auto_parallel/static/cost/op_runtime_cost.py

@@ -88,7 +88,7 @@ def _filter_vars_with_zero_in_degree_and_ignore_feed_fetch_vars():
  for op in cloned_main_block.ops:
  op: Operator
  if is_comm_op(op):
- # ignore communication op from graph, bacause sometimes we want to profile a sub-graph
+ # ignore communication op from graph, because sometimes we want to profile a sub-graph
  # and these dangling operators will not work (no graph to communicate to/from)
  continue
  input_var_names, output_var_names = _collect_op_input_var_names(
@@ -262,7 +262,7 @@ def measure_program_real_op_cost(
  >>> from paddle.distributed.auto_parallel.static.utils import measure_program_real_op_cost
  >>> place: str = paddle.device.get_device() # here we assume place = "cuda:x"
  >>> place = paddle.CUDAPlace(int(place.split(':')[1]))
- >>> # here "program" is an inner object that has alredy been built before
+ >>> # here "program" is an inner object that has already been built before
  >>> measure_program_real_op_cost(program, verbose_level=1)
  '''
 

python/paddle/distributed/auto_parallel/static/operators/dist_assign.py

@@ -22,7 +22,7 @@ def __init__(self, op_type):
  super().__init__(op_type)
 
 
-# TODO reomve assign dist op
+# TODO remove assign dist op
 # register_distributed_operator_impl_container(DistributedAssign("assign"))
 
 

python/paddle/distributed/auto_parallel/static/operators/dist_concat.py

@@ -43,16 +43,16 @@ def update_dims_mapping(dist_op):
  input_specs = []
  for name in input_arg_names:
  input_specs.append(get_dist_tensor_spec(dist_op, name))
- output_sepc = get_dist_tensor_spec(dist_op, output_arg_names[0], False)
+ output_spec = get_dist_tensor_spec(dist_op, output_arg_names[0], False)
 
  # step2: infer spmd
  rule = get_phi_spmd_rule("concat")
- # tensor order following order in PHI defition
+ # tensor order following order in PHI definition
  fw_results = rule.infer_forward(input_specs, axis)
- bw_results = rule.infer_backward(input_specs, output_sepc, axis)
+ bw_results = rule.infer_backward(input_specs, output_spec, axis)
 
  # step3: update dist_attr
- # tensor order following order in PHI defition
+ # tensor order following order in PHI definition
  changed = update_op_dims_mapping(
  dist_op,
  input_arg_names,

python/paddle/distributed/auto_parallel/static/operators/dist_cross_entropy.py

@@ -64,7 +64,7 @@ def update_dims_mapping(dist_op):
 
  # step2: infer spmd
  rule = get_phi_spmd_rule("softmax_with_cross_entropy")
- # tensor order following order in PHI defition
+ # tensor order following order in PHI definition
  fw_results = rule.infer_forward(
  logits_spec,
  label_spec,
@@ -87,7 +87,7 @@ def update_dims_mapping(dist_op):
  )
 
  # step3: update dist_attr
- # tensor order following order in PHI defition
+ # tensor order following order in PHI definition
  changed = update_op_dims_mapping(
  dist_op,
  [logits_name, label_name],

python/paddle/distributed/auto_parallel/static/operators/dist_default.py

@@ -141,12 +141,12 @@ def update_dims_mapping(dist_op):
 
  # step2: infer spmd
  rule = get_phi_spmd_rule("default_")
- # tensor order following order in PHI defition
+ # tensor order following order in PHI definition
  fw_results = rule.infer_forward(input_specs, output_specs)
  bw_results = rule.infer_backward(input_specs, output_specs)
 
  # step3: update dist_attr
- # tensor order following order in PHI defition
+ # tensor order following order in PHI definition
  changed = update_op_dims_mapping(
  dist_op, input_arg_names, output_arg_names, fw_results, bw_results
  )

python/paddle/distributed/auto_parallel/static/operators/dist_dropout.py

@@ -58,12 +58,12 @@ def update_dims_mapping(dist_op):
 
  # step2: infer spmd
  rule = get_phi_spmd_rule("dropout")
- # tensor order following order in PHI defition
+ # tensor order following order in PHI definition
  fw_results = rule.infer_forward(x_spec)
  bw_results = rule.infer_backward(x_spec, output_spec)
 
  # step3: update dist_attr
- # tensor order following order in PHI defition
+ # tensor order following order in PHI definition
  changed = update_op_dims_mapping(
  dist_op, [x_name], [out_name], fw_results, bw_results
  )