(
self, forward_input, curriculum_seqlen=None, labels=None, neox_args=None
)
| 122 | _set_use_cache(self.sequential, False) |
| 123 | |
| 124 | def forward( |
| 125 | self, forward_input, curriculum_seqlen=None, labels=None, neox_args=None |
| 126 | ): |
| 127 | |
| 128 | if ( |
| 129 | curriculum_seqlen is not None |
| 130 | and isinstance(forward_input, tuple) |
| 131 | and len(forward_input) == 3 |
| 132 | ): |
| 133 | neox_args.update_value("curriculum_seqlen", curriculum_seqlen) |
| 134 | tokens = forward_input[0] |
| 135 | input_ids = forward_input[1] |
| 136 | attention_mask = forward_input[2] |
| 137 | if curriculum_seqlen < input_ids.size()[1]: |
| 138 | # seqlen-based curriculum learning |
| 139 | # input_ids, position_ids, labels have size [batch size, seqlen] |
| 140 | input_ids = input_ids[:, :curriculum_seqlen].contiguous() |
| 141 | tokens = tokens[:, :curriculum_seqlen].contiguous() |
| 142 | # position_ids = position_ids[:, :curriculum_seqlen].contiguous() |
| 143 | if labels is not None: |
| 144 | labels = labels[:, :curriculum_seqlen].contiguous() |
| 145 | # attention_mask has size [1, 1, seqlen, seqlen] |
| 146 | attention_mask = attention_mask[ |
| 147 | :, :, :curriculum_seqlen, :curriculum_seqlen |
| 148 | ].contiguous() |
| 149 | forward_input = (tokens, input_ids, attention_mask) |
| 150 | |
| 151 | def exec_range_func(start, end): |
| 152 | """Helper function to be used with checkpoint() |
| 153 | Adapted from torch.utils.checkpoint:checkpoint_sequential() |
| 154 | """ |
| 155 | |
| 156 | def exec_func(*inputs): |
| 157 | # Single tensor inputs need to be unwrapped |
| 158 | if len(inputs) == 1: |
| 159 | inputs = inputs[0] |
| 160 | for idx, layer in enumerate(self.sequential[start:end]): |
| 161 | inputs = layer(inputs) |
| 162 | return inputs |
| 163 | |
| 164 | return exec_func |
| 165 | |
| 166 | if self.activation_checkpoint_interval == 0: |
| 167 | func = exec_range_func(0, len(self.sequential)) |
| 168 | x = func(forward_input) |
| 169 | else: |
| 170 | num_layers = len(self.sequential) |
| 171 | x = forward_input |
| 172 | for start_idx in range(0, num_layers, self.activation_checkpoint_interval): |
| 173 | end_idx = min( |
| 174 | start_idx + self.activation_checkpoint_interval, num_layers |
| 175 | ) |
| 176 | |
| 177 | funcs = self.sequential[start_idx:end_idx] |
| 178 | # Since we either pass tensors or tuples of tensors without unpacking, we |
| 179 | # need to be careful not to double-wrap tensors with tuple. |
| 180 | if not isinstance(x, tuple): |
| 181 | x = (x,) |
nothing calls this directly
no test coverage detected