vllm_omni.diffusion.models.bagel.pipeline_bagel ¶

cfg_img_scale: float = 1.5

cfg_interval: tuple = (0.4, 1.0)

cfg_renorm_min: float = 0.0

cfg_renorm_type: str = 'global'

cfg_text_scale: float = 4.0

num_timesteps: int = 50

timestep_shift: float = 3.0

bagel = Bagel(
    language_model=language_model,
    vit_model=vit_model,
    parallel_config=parallel_config,
    quant_config=quant_config,
    prefix="bagel",
    config=BagelConfig(
        llm_config=llm_config,
        vae_config=vae_cfg,
        vit_config=vit_cfg,
        vit_max_num_patch_per_side=int(
            get("vit_max_num_patch_per_side", 70)
        ),
        connector_act=str(
            get("connector_act", "gelu_pytorch_tanh")
        ),
        interpolate_pos=bool(get("interpolate_pos", False)),
        latent_patch_size=int(get("latent_patch_size", 2)),
        max_latent_size=int(get("max_latent_size", 32)),
        timestep_shift=float(get("timestep_shift", 1.0)),
    ),
)

device = get_local_device()

image_processor = from_pretrained(
    model_path, local_files_only=True
)

language_model = Qwen2MoTForCausalLM(
    llm_config,
    parallel_config=parallel_config,
    quant_config=quant_config,
    prefix="bagel.language_model",
)

od_config = od_config

scheduler: object | None = None

scheduler_kwargs: dict = {}

tokenizer = from_pretrained(
    model_path,
    local_files_only=True,
    trust_remote_code=True,
)

transformer = model

vae = DistributedAutoEncoder(ae_params)

vit_model = SiglipVisionModel(vit_conf)

weights_sources = [
    ComponentSource(
        model_or_path=model,
        subfolder=None,
        revision=revision,
        prefix="",
        fall_back_to_pt=False,
    )
]

vision_model = vision_model