vllm_omni.diffusion.attention.backends.trtllm_calibration ¶

logger `module-attribute` ¶

logger = logging.getLogger(__name__)

apply_skip_softmax_calibration(
    attention_config: Any, model: Any
) -> None

apply_to_pipeline(pipeline, calibration: dict) -> int

is_ignored(module_name: str, ignore_patterns) -> bool

layer_match_names(module_name: str) -> tuple[str, ...]

parse_secondary_expert_calibration(
    model: str | None,
) -> dict | None

parse_sparse_attention_config(
    sac: dict | None,
) -> dict | None

propagate_skip_softmax_calibration(
    specs: list, model: str | None, tf_config: Any
) -> None

resolve_layer_calibration(
    module_name: str, calibration: dict
) -> dict | None

select_expert(module_name: str, expert_keys) -> str | None