def get_sd3_models_for_export()

in optimum/exporters/openvino/convert.py [0:0]

• 88 lines of code
• 4 McCabe index (conditional complexity)

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,
        )
        export_config.runtime_options = {"ACTIVATIONS_SCALE_FACTOR": "8.0"}
        models_for_export["text_encoder_3"] = (text_encoder_3, export_config)

    return models_for_export