sd3 pipeline support

optimum/commands/export/openvino.py

@@ -318,6 +318,10 @@ def run(self):
                 from optimum.intel import OVStableDiffusionPipeline
 
                 model_cls = OVStableDiffusionPipeline
+            elif class_name == "StableDiffusion3Pipeline":
+                from optimum.intel import OVStableDiffusion3Pipeline
+
+                model_cls = OVStableDiffusion3Pipeline
             else:
                 raise NotImplementedError(f"Quantization in hybrid mode isn't supported for class {class_name}.")
 

optimum/exporters/openvino/__main__.py

@@ -491,7 +491,7 @@ def maybe_convert_tokenizers(library_name: str, output: Path, model=None, prepro
                         f"models won't be generated. Exception: {exception}"
                     )
         elif model:
-            for tokenizer_name in ("tokenizer", "tokenizer_2"):
+            for tokenizer_name in ("tokenizer", "tokenizer_2", "tokenizer_3"):
                 tokenizer = getattr(model, tokenizer_name, None)
                 if tokenizer:
                     export_tokenizer(tokenizer, output / tokenizer_name, task=task)

optimum/exporters/openvino/convert.py

@@ -12,6 +12,7 @@
 #  See the License for the specific language governing permissions and
 #  limitations under the License.
 
+import copy
 import functools
 import gc
 import logging
@@ -31,7 +32,12 @@
 from optimum.exporters.onnx.convert import check_dummy_inputs_are_allowed
 from optimum.exporters.onnx.convert import export_pytorch as export_pytorch_to_onnx
 from optimum.exporters.onnx.convert import export_tensorflow as export_tensorflow_onnx
-from optimum.exporters.utils import _get_submodels_and_export_configs as _default_get_submodels_and_export_configs
+from optimum.exporters.utils import (
+    _get_submodels_and_export_configs as _default_get_submodels_and_export_configs,
+)
+from optimum.exporters.utils import (
+    get_diffusion_models_for_export,
+)
 from optimum.intel.utils.import_utils import (
     _nncf_version,
     _open_clip_version,
@@ -618,23 +624,27 @@ def export_from_model(
             model, library_name, task, preprocessors, custom_export_configs, fn_get_submodels
         )
 
-    logging.disable(logging.INFO)
-    export_config, models_and_export_configs, stateful_submodels = _get_submodels_and_export_configs(
-        model=model,
-        task=task,
-        monolith=False,
-        custom_export_configs=custom_export_configs if custom_export_configs is not None else {},
-        custom_architecture=custom_architecture,
-        fn_get_submodels=fn_get_submodels,
-        preprocessors=preprocessors,
-        library_name=library_name,
-        model_kwargs=model_kwargs,
-        _variant="default",
-        legacy=False,
-        exporter="openvino",
-        stateful=stateful,
-    )
-    logging.disable(logging.NOTSET)
+    if library_name == "diffusers":
+        export_config, models_and_export_configs = get_diffusion_models_for_export_ext(model, exporter="openvino")
+        stateful_submodels = False
+    else:
+        logging.disable(logging.INFO)
+        export_config, models_and_export_configs, stateful_submodels = _get_submodels_and_export_configs(
+            model=model,
+            task=task,
+            monolith=False,
+            custom_export_configs=custom_export_configs if custom_export_configs is not None else {},
+            custom_architecture=custom_architecture,
+            fn_get_submodels=fn_get_submodels,
+            preprocessors=preprocessors,
+            library_name=library_name,
+            model_kwargs=model_kwargs,
+            _variant="default",
+            legacy=False,
+            exporter="openvino",
+            stateful=stateful,
+        )
+        logging.disable(logging.NOTSET)
 
     if library_name == "open_clip":
         if hasattr(model.config, "save_pretrained"):
@@ -700,6 +710,10 @@ def 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)
 
     export_models(
@@ -888,3 +902,218 @@ def _get_submodels_and_export_configs(
     )
     stateful_per_model = [stateful] * len(models_for_export)
     return export_config, models_for_export, stateful_per_model
+
+
+def get_diffusion_models_for_export_ext(
+    pipeline: "DiffusionPipeline", int_dtype: str = "int64", float_dtype: str = "fp32", exporter: str = "openvino"
+):
+    try:
+        from diffusers import (
+            StableDiffusion3Img2ImgPipeline,
+            StableDiffusion3InpaintPipeline,
+            StableDiffusion3Pipeline,
+        )
+
+        is_sd3 = isinstance(
+            pipeline, (StableDiffusion3Pipeline, StableDiffusion3InpaintPipeline, StableDiffusion3Img2ImgPipeline)
+        )
+    except ImportError:
+        is_sd3 = False
+
+    try:
+        from diffusers import FluxPipeline
+
+        is_flux = isinstance(pipeline, FluxPipeline)
+    except ImportError:
+        is_flux = False
+
+    if not is_sd3 and not is_flux:
+        return None, get_diffusion_models_for_export(pipeline, int_dtype, float_dtype, exporter)
+    if is_sd3:
+        models_for_export = get_sd3_models_for_export(pipeline, exporter, int_dtype, float_dtype)
+    else:
+        models_for_export = get_flux_models_for_export(pipeline, exporter, int_dtype, float_dtype)
+
+    return None, models_for_export
+
+
+def get_sd3_models_for_export(pipeline, exporter, int_dtype, float_dtype):
+    models_for_export = {}
+
+    # Text encoder
+    text_encoder = getattr(pipeline, "text_encoder", None)
+    if text_encoder is not None:
+        text_encoder.config.output_hidden_states = True
+        text_encoder.text_model.config.output_hidden_states = True
+        text_encoder_config_constructor = TasksManager.get_exporter_config_constructor(
+            model=text_encoder,
+            exporter=exporter,
+            library_name="diffusers",
+            task="feature-extraction",
+            model_type="clip-text-with-projection",
+        )
+        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"] = (text_encoder, text_encoder_export_config)
+
+    transformer = pipeline.transformer
+    transformer.config.text_encoder_projection_dim = transformer.config.joint_attention_dim
+    transformer.config.requires_aesthetics_score = getattr(pipeline.config, "requires_aesthetics_score", False)
+    transformer.config.time_cond_proj_dim = None
+    export_config_constructor = TasksManager.get_exporter_config_constructor(
+        model=transformer,
+        exporter=exporter,
+        library_name="diffusers",
+        task="semantic-segmentation",
+        model_type="sd3-transformer",
+    )
+    transformer_export_config = export_config_constructor(
+        pipeline.transformer.config, int_dtype=int_dtype, float_dtype=float_dtype
+    )
+    models_for_export["transformer"] = (transformer, transformer_export_config)
+
+    # VAE Encoder https://github.com/huggingface/diffusers/blob/v0.11.1/src/diffusers/models/vae.py#L565
+    vae_encoder = copy.deepcopy(pipeline.vae)
+    vae_encoder.forward = lambda sample: {"latent_parameters": vae_encoder.encode(x=sample)["latent_dist"].parameters}
+    vae_config_constructor = TasksManager.get_exporter_config_constructor(
+        model=vae_encoder,
+        exporter=exporter,
+        library_name="diffusers",
+        task="semantic-segmentation",
+        model_type="vae-encoder",
+    )
+    vae_encoder_export_config = vae_config_constructor(
+        vae_encoder.config, int_dtype=int_dtype, float_dtype=float_dtype
+    )
+    models_for_export["vae_encoder"] = (vae_encoder, vae_encoder_export_config)
+
+    # VAE Decoder https://github.com/huggingface/diffusers/blob/v0.11.1/src/diffusers/models/vae.py#L600
+    vae_decoder = copy.deepcopy(pipeline.vae)
+    vae_decoder.forward = lambda latent_sample: vae_decoder.decode(z=latent_sample)
+    vae_config_constructor = TasksManager.get_exporter_config_constructor(
+        model=vae_decoder,
+        exporter=exporter,
+        library_name="diffusers",
+        task="semantic-segmentation",
+        model_type="vae-decoder",
+    )
+    vae_decoder_export_config = vae_config_constructor(
+        vae_decoder.config, int_dtype=int_dtype, float_dtype=float_dtype
+    )
+    models_for_export["vae_decoder"] = (vae_decoder, vae_decoder_export_config)
+
+    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
+        export_config_constructor = TasksManager.get_exporter_config_constructor(
+            model=text_encoder_2,
+            exporter=exporter,
+            library_name="diffusers",
+            task="feature-extraction",
+            model_type="clip-text-with-projection",
+        )
+        export_config = export_config_constructor(text_encoder_2.config, int_dtype=int_dtype, float_dtype=float_dtype)
+        models_for_export["text_encoder_2"] = (text_encoder_2, export_config)
+
+    text_encoder_3 = getattr(pipeline, "text_encoder_3", None)
+    if text_encoder_3 is not None:
+        export_config_constructor = TasksManager.get_exporter_config_constructor(
+            model=text_encoder_3,
+            exporter=exporter,
+            library_name="diffusers",
+            task="feature-extraction",
+            model_type="t5-encoder-model",
+        )
+        export_config = export_config_constructor(
+            text_encoder_3.config,
+            int_dtype=int_dtype,
+            float_dtype=float_dtype,
+        )
+        models_for_export["text_encoder_3"] = (text_encoder_3, export_config)
+
+    return models_for_export
+
+
+def get_flux_models_for_export(pipeline, exporter, int_dtype, float_dtype):
+    models_for_export = {}
+
+    # Text encoder
+    text_encoder = getattr(pipeline, "text_encoder", None)
+    if text_encoder is not None:
+        text_encoder_config_constructor = TasksManager.get_exporter_config_constructor(
+            model=text_encoder,
+            exporter=exporter,
+            library_name="diffusers",
+            task="feature-extraction",
+            model_type="clip-text-model",
+        )
+        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"] = (text_encoder, text_encoder_export_config)
+
+    transformer = pipeline.transformer
+    transformer.config.text_encoder_projection_dim = transformer.config.joint_attention_dim
+    transformer.config.requires_aesthetics_score = getattr(pipeline.config, "requires_aesthetics_score", False)
+    transformer.config.time_cond_proj_dim = None
+    export_config_constructor = TasksManager.get_exporter_config_constructor(
+        model=transformer,
+        exporter=exporter,
+        library_name="diffusers",
+        task="semantic-segmentation",
+        model_type="flux-transformer",
+    )
+    transformer_export_config = export_config_constructor(
+        pipeline.transformer.config, int_dtype=int_dtype, float_dtype=float_dtype
+    )
+    models_for_export["transformer"] = (transformer, transformer_export_config)
+
+    # VAE Encoder https://github.com/huggingface/diffusers/blob/v0.11.1/src/diffusers/models/vae.py#L565
+    vae_encoder = copy.deepcopy(pipeline.vae)
+    vae_encoder.forward = lambda sample: {"latent_parameters": vae_encoder.encode(x=sample)["latent_dist"].parameters}
+    vae_config_constructor = TasksManager.get_exporter_config_constructor(
+        model=vae_encoder,
+        exporter=exporter,
+        library_name="diffusers",
+        task="semantic-segmentation",
+        model_type="vae-encoder",
+    )
+    vae_encoder_export_config = vae_config_constructor(
+        vae_encoder.config, int_dtype=int_dtype, float_dtype=float_dtype
+    )
+    models_for_export["vae_encoder"] = (vae_encoder, vae_encoder_export_config)
+
+    # VAE Decoder https://github.com/huggingface/diffusers/blob/v0.11.1/src/diffusers/models/vae.py#L600
+    vae_decoder = copy.deepcopy(pipeline.vae)
+    vae_decoder.forward = lambda latent_sample: vae_decoder.decode(z=latent_sample)
+    vae_config_constructor = TasksManager.get_exporter_config_constructor(
+        model=vae_decoder,
+        exporter=exporter,
+        library_name="diffusers",
+        task="semantic-segmentation",
+        model_type="vae-decoder",
+    )
+    vae_decoder_export_config = vae_config_constructor(
+        vae_decoder.config, int_dtype=int_dtype, float_dtype=float_dtype
+    )
+    models_for_export["vae_decoder"] = (vae_decoder, vae_decoder_export_config)
+
+    text_encoder_2 = getattr(pipeline, "text_encoder_2", None)
+    if text_encoder_2 is not None:
+        export_config_constructor = TasksManager.get_exporter_config_constructor(
+            model=text_encoder_2,
+            exporter=exporter,
+            library_name="diffusers",
+            task="feature-extraction",
+            model_type="t5-encoder-model",
+        )
+        export_config = export_config_constructor(
+            text_encoder_2.config,
+            int_dtype=int_dtype,
+            float_dtype=float_dtype,
+        )
+        models_for_export["text_encoder_2"] = (text_encoder_2, export_config)
+
+    return models_for_export