(self, input_tensor: torch.Tensor, temb: torch.Tensor, *args, **kwargs)
| 317 | ) |
| 318 | |
| 319 | def forward(self, input_tensor: torch.Tensor, temb: torch.Tensor, *args, **kwargs) -> torch.Tensor: |
| 320 | if len(args) > 0 or kwargs.get("scale", None) is not None: |
| 321 | deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`." |
| 322 | deprecate("scale", "1.0.0", deprecation_message) |
| 323 | |
| 324 | hidden_states = input_tensor |
| 325 | |
| 326 | hidden_states = self.norm1(hidden_states) |
| 327 | hidden_states = self.nonlinearity(hidden_states) |
| 328 | |
| 329 | if self.upsample is not None: |
| 330 | # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984 |
| 331 | if hidden_states.shape[0] >= 64: |
| 332 | input_tensor = input_tensor.contiguous() |
| 333 | hidden_states = hidden_states.contiguous() |
| 334 | input_tensor = self.upsample(input_tensor) |
| 335 | hidden_states = self.upsample(hidden_states) |
| 336 | elif self.downsample is not None: |
| 337 | input_tensor = self.downsample(input_tensor) |
| 338 | hidden_states = self.downsample(hidden_states) |
| 339 | |
| 340 | hidden_states = self.conv1(hidden_states) |
| 341 | |
| 342 | if self.time_emb_proj is not None: |
| 343 | if not self.skip_time_act: |
| 344 | temb = self.nonlinearity(temb) |
| 345 | temb = self.time_emb_proj(temb)[:, :, None, None] |
| 346 | |
| 347 | if self.time_embedding_norm == "default": |
| 348 | if temb is not None: |
| 349 | hidden_states = hidden_states + temb |
| 350 | hidden_states = self.norm2(hidden_states) |
| 351 | elif self.time_embedding_norm == "scale_shift": |
| 352 | if temb is None: |
| 353 | raise ValueError( |
| 354 | f" `temb` should not be None when `time_embedding_norm` is {self.time_embedding_norm}" |
| 355 | ) |
| 356 | time_scale, time_shift = torch.chunk(temb, 2, dim=1) |
| 357 | hidden_states = self.norm2(hidden_states) |
| 358 | hidden_states = hidden_states * (1 + time_scale) + time_shift |
| 359 | else: |
| 360 | hidden_states = self.norm2(hidden_states) |
| 361 | |
| 362 | hidden_states = self.nonlinearity(hidden_states) |
| 363 | |
| 364 | hidden_states = self.dropout(hidden_states) |
| 365 | hidden_states = self.conv2(hidden_states) |
| 366 | |
| 367 | if self.conv_shortcut is not None: |
| 368 | # Only use contiguous() during training to avoid DDP gradient stride mismatch warning. |
| 369 | # In inference mode (eval or no_grad), skip contiguous() for better performance, especially on CPU. |
| 370 | # Issue: https://github.com/huggingface/diffusers/issues/12975 |
| 371 | if self.training: |
| 372 | input_tensor = input_tensor.contiguous() |
| 373 | input_tensor = self.conv_shortcut(input_tensor) |
| 374 | |
| 375 | output_tensor = (input_tensor + hidden_states) / self.output_scale_factor |
| 376 |
nothing calls this directly
no test coverage detected