Placeholder linear kernels for online requantization sources.
A checkpoint quantization method used as an online requantization source only
dequantizes its weights, so it needs no kernel. The init_*_linear_kernel
selectors return one of these placeholders instead of dispatching, which keeps
their return types unchanged. Using one for compute raises.
_FakeLinearKernel
Satisfies the abstract interface of every kernel family, never computes.
Source code in vllm/model_executor/kernels/linear/requantization_source.py
| class _FakeLinearKernel:
"""Satisfies the abstract interface of every kernel family, never computes."""
def __init__(self, *args: Any, **kwargs: Any) -> None:
pass
@classmethod
def is_supported(cls, *args: Any, **kwargs: Any) -> tuple[bool, str | None]:
return True, None
@classmethod
def can_implement(cls, *args: Any, **kwargs: Any) -> tuple[bool, str | None]:
return True, None
def input_quant_key(self) -> QuantKey | None:
return None
def process_weights_after_loading(self, *args: Any, **kwargs: Any) -> None:
raise RuntimeError(_NO_KERNEL)
def apply_weights(self, *args: Any, **kwargs: Any) -> torch.Tensor:
raise RuntimeError(_NO_KERNEL)
def apply_scaled_mm(self, *args: Any, **kwargs: Any) -> torch.Tensor:
raise RuntimeError(_NO_KERNEL)
|