SD3 support

optimum/exporters/onnx/base.py

@@ -319,6 +319,7 @@ def fix_dynamic_axes(
  input_shapes = {}
  dummy_inputs = self.generate_dummy_inputs(framework="np", **input_shapes)
  dummy_inputs = self.generate_dummy_inputs_for_validation(dummy_inputs, onnx_input_names=onnx_input_names)
+ dummy_inputs = self.rename_ambiguous_inputs(dummy_inputs)
 
  onnx_inputs = {}
  for name, value in dummy_inputs.items():

optimum/exporters/onnx/convert.py

@@ -1183,6 +1183,10 @@ def onnx_export_from_model(
  if tokenizer_2 is not None:
  tokenizer_2.save_pretrained(output.joinpath("tokenizer_2"))
 
+ tokenizer_3 = getattr(model, "tokenizer_3", None)
+ if tokenizer_3 is not None:
+ tokenizer_3.save_pretrained(output.joinpath("tokenizer_3"))
+
  model.save_config(output)
 
  if float_dtype == "bf16":

optimum/exporters/onnx/model_configs.py

@@ -1015,22 +1015,13 @@ def outputs(self) -> Dict[str, Dict[int, str]]:
  "last_hidden_state": {0: "batch_size", 1: "sequence_length"},
  "pooler_output": {0: "batch_size"},
  }
+
  if self._normalized_config.output_hidden_states:
  for i in range(self._normalized_config.num_layers + 1):
  common_outputs[f"hidden_states.{i}"] = {0: "batch_size", 1: "sequence_length"}
 
  return common_outputs
 
- def generate_dummy_inputs(self, framework: str = "pt", **kwargs):
- dummy_inputs = super().generate_dummy_inputs(framework=framework, **kwargs)
-
- # TODO: fix should be by casting inputs during inference and not export
- if framework == "pt":
- import torch
-
- dummy_inputs["input_ids"] = dummy_inputs["input_ids"].to(dtype=torch.int32)
- return dummy_inputs
-
  def patch_model_for_export(
  self,
  model: Union["PreTrainedModel", "TFPreTrainedModel", "ModelMixin"],
@@ -1160,6 +1151,76 @@ def outputs(self) -> Dict[str, Dict[int, str]]:
  }
 
 
+class PooledProjectionsDummyInputGenerator(DummyInputGenerator):
+ SUPPORTED_INPUT_NAMES = "pooled_projections"
+
+ def __init__(
+ self,
+ task: str,
+ normalized_config: NormalizedConfig,
+ batch_size: int = DEFAULT_DUMMY_SHAPES["batch_size"],
+ **kwargs,
+ ):
+ self.task = task
+ self.batch_size = batch_size
+ self.pooled_projection_dim = normalized_config.config.pooled_projection_dim
+
+ def generate(self, input_name: str, framework: str = "pt", int_dtype: str = "int64", float_dtype: str = "fp32"):
+ return self.random_float_tensor(
+ [self.batch_size, self.pooled_projection_dim], framework=framework, dtype=float_dtype
+ )
+
+
+class DummyTransformerTimestpsInputGenerator(DummyTimestepInputGenerator):
+ def generate(self, input_name: str, framework: str = "pt", int_dtype: str = "int64", float_dtype: str = "fp32"):
+ if input_name == "timestep":
+ shape = [self.batch_size]
+ return self.random_float_tensor(shape, max_value=self.vocab_size, framework=framework, dtype=float_dtype)
+ return super().generate(input_name, framework, int_dtype, float_dtype)
+
+
+class SD3TransformerOnnxConfig(UNetOnnxConfig):
+ DUMMY_INPUT_GENERATOR_CLASSES = (
+ (DummyTransformerTimestpsInputGenerator,)
+ + UNetOnnxConfig.DUMMY_INPUT_GENERATOR_CLASSES
+ + (PooledProjectionsDummyInputGenerator,)
+ )
+ NORMALIZED_CONFIG_CLASS = NormalizedConfig.with_args(
+ image_size="sample_size",
+ num_channels="in_channels",
+ hidden_size="joint_attention_dim",
+ vocab_size="attention_head_dim",
+ allow_new=True,
+ )
+
+ @property
+ def inputs(self):
+ common_inputs = super().inputs
+ common_inputs["pooled_projections"] = {0: "batch_size"}
+ return common_inputs
+
+ def rename_ambiguous_inputs(self, inputs):
+ # The input name in the model signature is `x, hence the export input name is updated.
+ hidden_states = inputs.pop("sample", None)
+ if hidden_states is not None:
+ inputs["hidden_states"] = hidden_states
+ return inputs
+
+
+class T5EncoderOnnxConfig(CLIPTextOnnxConfig):
+ @property
+ def inputs(self):
+ return {
+ "input_ids": {0: "batch_size", 1: "sequence_length"},
+ }
+
+ @property
+ def outputs(self):
+ return {
+ "last_hidden_state": {0: "batch_size", 1: "sequence_length"},
+ }
+
+
 class GroupViTOnnxConfig(CLIPOnnxConfig):
  pass
 

optimum/exporters/tasks.py

@@ -335,6 +335,10 @@ class TasksManager:
  }
 
  _DIFFUSERS_SUPPORTED_MODEL_TYPE = {
+ "t5-encoder": supported_tasks_mapping(
+ "feature-extraction",
+ onnx="T5EncoderOnnxConfig",
+ ),
  "clip-text-model": supported_tasks_mapping(
  "feature-extraction",
  onnx="CLIPTextOnnxConfig",
@@ -347,6 +351,10 @@ class TasksManager:
  "semantic-segmentation",
  onnx="UNetOnnxConfig",
  ),
+ "sd3-transformer": supported_tasks_mapping(
+ "semantic-segmentation",
+ onnx="SD3TransformerOnnxConfig",
+ ),
  "vae-encoder": supported_tasks_mapping(
  "semantic-segmentation",
  onnx="VaeEncoderOnnxConfig",
@@ -1170,12 +1178,14 @@ class TasksManager:
  "transformers": _SUPPORTED_MODEL_TYPE,
  }
  _UNSUPPORTED_CLI_MODEL_TYPE = {
- "unet",
- "vae-encoder",
- "vae-decoder",
  "clip-text-model",
  "clip-text-with-projection",
+ "sd3-transformer",
+ "t5-encoder",
  "trocr", # supported through the vision-encoder-decoder model type
+ "unet",
+ "vae-encoder",
+ "vae-decoder",
  }
  _SUPPORTED_CLI_MODEL_TYPE = (
  set(_SUPPORTED_MODEL_TYPE.keys())

optimum/exporters/utils.py

@@ -50,6 +50,14 @@
  StableDiffusionXLInpaintPipeline,
  StableDiffusionXLPipeline,
  )
+
+ if check_if_diffusers_greater("0.30.0"):
+ from diffusers import (
+ StableDiffusion3Img2ImgPipeline,
+ StableDiffusion3InpaintPipeline,
+ StableDiffusion3Pipeline,
+ )
+
  from diffusers.models.attention_processor import (
  Attention,
  AttnAddedKVProcessor,
@@ -87,56 +95,95 @@ def _get_submodels_for_export_diffusion(
  Returns the components of a Stable Diffusion model.
  """
 
+ models_for_export = {}
+
  is_stable_diffusion_xl = isinstance(
  pipeline, (StableDiffusionXLPipeline, StableDiffusionXLImg2ImgPipeline, StableDiffusionXLInpaintPipeline)
  )
- if is_stable_diffusion_xl:
- projection_dim = pipeline.text_encoder_2.config.projection_dim
- else:
- projection_dim = pipeline.text_encoder.config.projection_dim
+ is_stable_diffusion_3 = isinstance(
+ pipeline, (StableDiffusion3Pipeline, StableDiffusion3Img2ImgPipeline, StableDiffusion3InpaintPipeline)
+ )
 
- models_for_export = {}
+ is_torch_greater_or_equal_than_2_1 = version.parse(torch.__version__) >= version.parse("2.1.0")
 
  # Text encoder
  text_encoder = getattr(pipeline, "text_encoder", None)
  if text_encoder is not None:
- if is_stable_diffusion_xl:
+ if is_stable_diffusion_xl or is_stable_diffusion_3:
  text_encoder.config.output_hidden_states = True
+ text_encoder.text_model.config.output_hidden_states = True
+
+ if is_stable_diffusion_3:
+ text_encoder.config.export_model_type = "clip-text-with-projection"
+ else:
+ text_encoder.config.export_model_type = "clip-text-model"
+
  models_for_export["text_encoder"] = text_encoder
 
- # U-NET
- # ONNX export of torch.nn.functional.scaled_dot_product_attention not supported for < v2.1.0
- is_torch_greater_or_equal_than_2_1 = version.parse(torch.__version__) >= version.parse("2.1.0")
- if not is_torch_greater_or_equal_than_2_1:
- pipeline.unet.set_attn_processor(AttnProcessor())
+ # Text encoder 2
+ text_encoder_2 = getattr(pipeline, "text_encoder_2", None)
+ if text_encoder_2 is not None:
+ text_encoder_2.config.output_hidden_states = True
+ text_encoder_2.text_model.config.output_hidden_states = True
+ text_encoder_2.config.export_model_type = "clip-text-with-projection"
 
- pipeline.unet.config.text_encoder_projection_dim = projection_dim
- # The U-NET time_ids inputs shapes depends on the value of `requires_aesthetics_score`
- # https://github.com/huggingface/diffusers/blob/v0.18.2/src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py#L571
- pipeline.unet.config.requires_aesthetics_score = getattr(pipeline.config, "requires_aesthetics_score", False)
- models_for_export["unet"] = pipeline.unet
+ models_for_export["text_encoder_2"] = text_encoder_2
 
- # VAE Encoder https://github.com/huggingface/diffusers/blob/v0.11.1/src/diffusers/models/vae.py#L565
+ # Text encoder 3
+ text_encoder_3 = getattr(pipeline, "text_encoder_3", None)
+ if text_encoder_3 is not None:
+ text_encoder_3.config.export_model_type = "t5-encoder"
+ models_for_export["text_encoder_3"] = text_encoder_3
+
+ # U-NET
+ unet = getattr(pipeline, "unet", None)
+ if unet is not None:
+ # ONNX export of torch.nn.functional.scaled_dot_product_attention not supported for < v2.1.0
+ if not is_torch_greater_or_equal_than_2_1:
+ unet.set_attn_processor(AttnProcessor())
+
+ # The U-NET time_ids inputs shapes depends on the value of `requires_aesthetics_score`
+ # https://github.com/huggingface/diffusers/blob/v0.18.2/src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py#L571
+ unet.config.requires_aesthetics_score = getattr(pipeline.config, "requires_aesthetics_score", False)
+ unet.config.time_cond_proj_dim = getattr(pipeline.unet.config, "time_cond_proj_dim", None)
+ unet.config.text_encoder_projection_dim = pipeline.text_encoder.config.projection_dim
+ unet.config.export_model_type = "unet"
+ models_for_export["unet"] = unet
+
+ # Transformer
+ transformer = getattr(pipeline, "transformer", None)
+ if transformer is not None:
+ # ONNX export of torch.nn.functional.scaled_dot_product_attention not supported for < v2.1.0
+ if not is_torch_greater_or_equal_than_2_1:
+ transformer.set_attn_processor(AttnProcessor())
+
+ transformer.config.requires_aesthetics_score = getattr(pipeline.config, "requires_aesthetics_score", False)
+ transformer.config.time_cond_proj_dim = getattr(pipeline.transformer.config, "time_cond_proj_dim", None)
+ transformer.config.text_encoder_projection_dim = pipeline.text_encoder.config.projection_dim
+ transformer.config.export_model_type = "sd3-transformer"
+ models_for_export["transformer"] = transformer
+
+ # VAE Encoder
  vae_encoder = copy.deepcopy(pipeline.vae)
+
+ # ONNX export of torch.nn.functional.scaled_dot_product_attention not supported for < v2.1.0
  if not is_torch_greater_or_equal_than_2_1:
  vae_encoder = override_diffusers_2_0_attn_processors(vae_encoder)
+
  # we return the distribution parameters to be able to recreate it in the decoder
  vae_encoder.forward = lambda sample: {"latent_parameters": vae_encoder.encode(x=sample)["latent_dist"].parameters}
  models_for_export["vae_encoder"] = vae_encoder
 
- # VAE Decoder https://github.com/huggingface/diffusers/blob/v0.11.1/src/diffusers/models/vae.py#L600
+ # VAE Decoder
  vae_decoder = copy.deepcopy(pipeline.vae)
+
+ # ONNX export of torch.nn.functional.scaled_dot_product_attention not supported for < v2.1.0
  if not is_torch_greater_or_equal_than_2_1:
  vae_decoder = override_diffusers_2_0_attn_processors(vae_decoder)
+
  vae_decoder.forward = lambda latent_sample: vae_decoder.decode(z=latent_sample)
  models_for_export["vae_decoder"] = vae_decoder
 
- text_encoder_2 = getattr(pipeline, "text_encoder_2", None)
- if text_encoder_2 is not None:
- text_encoder_2.config.output_hidden_states = True
- text_encoder_2.text_model.config.output_hidden_states = True
- models_for_export["text_encoder_2"] = text_encoder_2
-
  return models_for_export
 
 
@@ -294,31 +341,58 @@ def get_diffusion_models_for_export(
  `Dict[str, Tuple[Union[`PreTrainedModel`, `TFPreTrainedModel`], `ExportConfig`]: A Dict containing the model and
  export configs for the different components of the model.
  """
+
  models_for_export = _get_submodels_for_export_diffusion(pipeline)
 
  # Text encoder
  if "text_encoder" in models_for_export:
  text_encoder_config_constructor = TasksManager.get_exporter_config_constructor(
- model=pipeline.text_encoder,
- exporter=exporter,
- library_name="diffusers",
- task="feature-extraction",
+ model=pipeline.text_encoder, exporter=exporter, library_name="diffusers", task="feature-extraction"
  )
  text_encoder_export_config = text_encoder_config_constructor(
  pipeline.text_encoder.config, int_dtype=int_dtype, float_dtype=float_dtype
  )
  models_for_export["text_encoder"] = (models_for_export["text_encoder"], text_encoder_export_config)
 
+ # Text encoder 2
+ if "text_encoder_2" in models_for_export:
+ export_config_constructor = TasksManager.get_exporter_config_constructor(
+ model=pipeline.text_encoder_2, exporter=exporter, library_name="diffusers", task="feature-extraction"
+ )
+ export_config = export_config_constructor(
+ pipeline.text_encoder_2.config, int_dtype=int_dtype, float_dtype=float_dtype
+ )
+ models_for_export["text_encoder_2"] = (models_for_export["text_encoder_2"], export_config)
+
+ # Text encoder 3
+ if "text_encoder_3" in models_for_export:
+ export_config_constructor = TasksManager.get_exporter_config_constructor(
+ model=pipeline.text_encoder_3, exporter=exporter, library_name="diffusers", task="feature-extraction"
+ )
+ export_config = export_config_constructor(
+ pipeline.text_encoder_3.config, int_dtype=int_dtype, float_dtype=float_dtype
+ )
+ models_for_export["text_encoder_3"] = (models_for_export["text_encoder_3"], export_config)
+
  # U-NET
- export_config_constructor = TasksManager.get_exporter_config_constructor(
- model=pipeline.unet,
- exporter=exporter,
- library_name="diffusers",
- task="semantic-segmentation",
- model_type="unet",
- )
- unet_export_config = export_config_constructor(pipeline.unet.config, int_dtype=int_dtype, float_dtype=float_dtype)
- models_for_export["unet"] = (models_for_export["unet"], unet_export_config)
+ if "unet" in models_for_export:
+ export_config_constructor = TasksManager.get_exporter_config_constructor(
+ model=pipeline.unet, exporter=exporter, library_name="diffusers", task="semantic-segmentation"
+ )
+ unet_export_config = export_config_constructor(
+ pipeline.unet.config, int_dtype=int_dtype, float_dtype=float_dtype
+ )
+ models_for_export["unet"] = (models_for_export["unet"], unet_export_config)
+
+ # Transformer
+ if "transformer" in models_for_export:
+ export_config_constructor = TasksManager.get_exporter_config_constructor(
+ model=pipeline.transformer, exporter=exporter, library_name="diffusers", task="semantic-segmentation"
+ )
+ transformer_export_config = export_config_constructor(
+ pipeline.transformer.config, int_dtype=int_dtype, float_dtype=float_dtype
+ )
+ models_for_export["transformer"] = (models_for_export["transformer"], transformer_export_config)
 
  # VAE Encoder https://github.com/huggingface/diffusers/blob/v0.11.1/src/diffusers/models/vae.py#L565
  vae_encoder = models_for_export["vae_encoder"]
@@ -344,19 +418,6 @@ def get_diffusion_models_for_export(
  vae_export_config = vae_config_constructor(vae_decoder.config, int_dtype=int_dtype, float_dtype=float_dtype)
  models_for_export["vae_decoder"] = (vae_decoder, vae_export_config)
 
- if "text_encoder_2" in models_for_export:
- export_config_constructor = TasksManager.get_exporter_config_constructor(
- model=pipeline.text_encoder_2,
- exporter=exporter,
- library_name="diffusers",
- task="feature-extraction",
- model_type="clip-text-with-projection",
- )
- export_config = export_config_constructor(
- pipeline.text_encoder_2.config, int_dtype=int_dtype, float_dtype=float_dtype
- )
- models_for_export["text_encoder_2"] = (models_for_export["text_encoder_2"], export_config)
-
  return models_for_export