Function _resolve_sliding_window

src/whichllm/models/fetcher.py:259–287

Resolve (sliding_window, global_ratio) for honored SWA architectures. Returns (None, None) for every model outside the allowlist so the KV estimate stays at full context (conservative).

(
    config: dict, model_id: str, gguf_arch: str | None = None
)

def _resolve_sliding_window(
    config: dict, model_id: str, gguf_arch: str | None = None
) -> tuple[int | None, float | None]:
    """Resolve (sliding_window, global_ratio) for honored SWA architectures.

    Returns (None, None) for every model outside the allowlist so the KV
    estimate stays at full context (conservative).
    """
    # Respect an explicit opt-out before doing any work.
    if config.get("use_sliding_window") is False:
        return None, None

    key = _swa_arch_key(config, model_id, gguf_arch)
    if key is None:
        return None, None

    default_window, default_ratio = _SWA_ARCH_DEFAULTS[key]

    window = config.get("sliding_window")
    if not isinstance(window, int) or window <= 0:
        window = default_window

    pattern = config.get("sliding_window_pattern")
    if isinstance(pattern, int) and pattern > 0:
        global_ratio = 1.0 / pattern
    else:
        global_ratio = default_ratio

    return window, global_ratio


def _normalize_param_count(