Skip to content

vllm_omni.model_extras.mammothmodal2_preview ¶

MAMMOTHMODA2_PREVIEW_EXTRA_BODY_PARAMS `module-attribute` ¶

MAMMOTHMODA2_PREVIEW_EXTRA_BODY_PARAMS = frozenset(
    {
        "text_guidance_scale",
        "cfg_range",
        "num_inference_steps",
    }
)

MAMMOTHMODA2_PREVIEW_EXTRA_OUTPUT_PARAMS `module-attribute` ¶

MAMMOTHMODA2_PREVIEW_EXTRA_OUTPUT_PARAMS = frozenset()

MAMMOTHMODA2_PREVIEW_INIT_EXTRA_ARGS_FOR_NON_DIFFUSION_STAGES `module-attribute` ¶

MAMMOTHMODA2_PREVIEW_INIT_EXTRA_ARGS_FOR_NON_DIFFUSION_STAGES = True

build_text_to_image_prompt ¶

build_text_to_image_prompt(
    prompt: str,
    negative_prompt: str | None,
    height: int | None = None,
    width: int | None = None,
) -> dict[str, Any]

Build the MammothModa2 AR-stage prompt for text-to-image generation.

Reproduces the prompt string and the structural additional_information that the former bespoke example (run_mammothmoda2_t2i.py) constructed, deriving the AR image grid from height / width.

Model-specific sampling knobs (text_guidance_scale, cfg_range, num_inference_steps) flow separately via extra_body -> extra_args. Config-derived token ids (eol_token_id, visual_token_start_id, visual_token_end_id, visual_ids) are sourced inside the AR stage rather than passed by the caller -- see the ar2dit stage input processor.

MammothModa2 t2i uses classifier-free guidance via text_guidance_scale and has no explicit negative-prompt path, so negative_prompt is accepted for signature compatibility but not injected.