(self, input_)
| 304 | stride=stride, return_master_weight=keep_master_weight_for_test) |
| 305 | |
| 306 | def forward(self, input_): |
| 307 | # Set up backprop all-reduce. |
| 308 | if self.input_is_parallel: |
| 309 | input_parallel = input_ |
| 310 | else: |
| 311 | input_parallel = scatter_to_model_parallel_region(input_) |
| 312 | # Matrix multiply. |
| 313 | output_parallel = F.linear(input_parallel, self.weight) |
| 314 | # All-reduce across all the partitions. |
| 315 | output_ = reduce_from_model_parallel_region(output_parallel) |
| 316 | if self.bias is not None: |
| 317 | output = output_ + self.bias |
| 318 | else: |
| 319 | output = output_ |
| 320 | return output |
| 321 |
nothing calls this directly
no test coverage detected