KVQuantSpec
Per-side K/V quantize-on-append scheme: element-coded (dtype + scale), block-quantized (block set), or per-token online scales.
block (property)
(self) -> clika_runtime._core.nn.KVBlockScheme
dtype (property)
(self) -> clika_runtime._core.DataType
per_token (property)
(self) -> bool
scale (property)
(self) -> float
zero_point (property)
(self) -> int | None
__init__
__init____init__(self, *, dtype: clika_runtime._core.DataType = DataType.Int8, scale: float = 1.0, zero_point: int | None = None, block: clika_runtime._core.nn.KVBlockScheme = KVBlockScheme.None_, per_token: bool = False) -> None
init(self, *, dtype: clika_runtime._core.DataType = DataType.Int8, scale: float = 1.0, zero_point: int | None = None, block: clika_runtime.core.nn.KVBlockScheme = KVBlockScheme.None, per_token: bool = False) -> None