Coverage for hyper_parallel / platform / torch / pipeline_parallel / _utils.py: 71%

1# Copyright 2025 Huawei Technologies Co., Ltd 

2# 

3# Licensed under the Apache License, Version 2.0 (the "License"); 

4# you may not use this file except in compliance with the License. 

5# You may obtain a copy of the License at 

6# 

7# http://www.apache.org/licenses/LICENSE-2.0 

8# 

9# Unless required by applicable law or agreed to in writing, software 

10# distributed under the License is distributed on an "AS IS" BASIS, 

11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 

12# See the License for the specific language governing permissions and 

13# limitations under the License. 

14# ============================================================================ 

15"""pipeline parallel utils""" 

16import hyper_parallel 

17from torch import nn 

18 

19 

20class _MicroBatch(nn.Module): 

21 """ 

22 Split inputs into micro_batch in pipeline parallel. 

23 

24 Args: 

25 micro_batch_num (int): The number of micro-batch. 

26 args_batch_dim (list, optional): Specify the batch dim of the args. 

27 Default ``None``. 

28 kwargs_batch_dim(dict, optional): Specify the batch dim of the kwargs. 

29 Default ``None``. 

30 Inputs: 

31 - **args** (list) - Input args. 

32 - **kwargs** (dict) - Input kwargs. 

33 

34 Outputs: 

35 - **args_after_split** (list) - Input args after split into micro_batches. 

36 - **kwargs_after_split** (list) - Input kwargs after split into micro_batches. 

37 """ 

38 

39 def __init__(self, micro_batch_num, args_batch_dim=None, kwargs_batch_dim=None): 

40 super().__init__() 

41 self.micro_batch_num = micro_batch_num 

42 self.args_batch_dim = args_batch_dim 

43 self.kwargs_batch_dim = kwargs_batch_dim 

44 

45 def forward(self, args, kwargs): 

46 """forward of _MicroBatch""" 

47 args_after_split = [] 

48 kwargs_after_split = [] 

49 for micro_idx in range(self.micro_batch_num): 

50 micro_args = [] 

51 micro_kwargs = {} 

52 for arg_idx, cur_arg in enumerate(args): 

53 cur_arg_batch_dim = 0 

54 if self.args_batch_dim and self.args_batch_dim[arg_idx] is not None: 

55 cur_arg_batch_dim = self.args_batch_dim[arg_idx].batch_dim 

56 if isinstance(cur_arg, hyper_parallel.DTensor): 

57 micro_arg = self.split_inputs_with_custom_shard(cur_arg, cur_arg_batch_dim, micro_idx) 

58 else: 

59 micro_arg = self.split_inputs(cur_arg, cur_arg_batch_dim, micro_idx) 

60 micro_args.append(micro_arg) 

61 args_after_split.append(micro_args) 

62 

63 for key, cur_kwarg in kwargs.items(): 

64 cur_kwarg_batch_dim = 0 

65 if self.kwargs_batch_dim is not None: 

66 cur_kwarg_batch_dim = self.kwargs_batch_dim[key].batch_dim 

67 if isinstance(cur_kwarg, hyper_parallel.DTensor): 

68 micro_kwarg = self.split_inputs_with_custom_shard(cur_kwarg, cur_kwarg_batch_dim, micro_idx) 

69 else: 

70 micro_kwarg = self.split_inputs(cur_kwarg, cur_kwarg_batch_dim, micro_idx) 

71 micro_kwargs[key] = micro_kwarg 

72 kwargs_after_split.append(micro_kwargs) 

73 return args_after_split, kwargs_after_split 

74 

75 def split_inputs_with_custom_shard(self, input_tensor, cur_arg_batch_dim, micro_idx): 

76 input_layout = input_tensor.layout 

77 func_wrap = hyper_parallel.custom_shard(self.split_inputs, 

78 device_mesh=input_layout.mesh, 

79 out_placements=(input_layout.placements,), 

80 in_placements=(input_layout.placements, None, None) 

81 ) 

82 return func_wrap(input_tensor, cur_arg_batch_dim, micro_idx) 

83 

84 def split_inputs(self, input_tensor, cur_arg_batch_dim, micro_idx): 

85 """ 

86 Split the input along the specified batch_dim and micro_idx 

87 """ 

88 if cur_arg_batch_dim == -1: 

89 return input_tensor 

90 batch_dim_shape = input_tensor.shape[cur_arg_batch_dim] 

91 if batch_dim_shape % self.micro_batch_num != 0: 

92 raise ValueError(f"Batch dimension size {batch_dim_shape} is not divisible by \ 

93 micro_batch_num {self.micro_batch_num}") 

94 micro_batch_size = batch_dim_shape // self.micro_batch_num 

95 

96 # Calculate start and end idx 

97 start = micro_batch_size * micro_idx 

98 end = micro_batch_size * (micro_idx + 1) 

99 

100 # Create slicing tuple 

101 slices = [slice(None)] * input_tensor.ndim 

102 slices[cur_arg_batch_dim] = slice(start, end) 

103 return input_tensor[slices]