vllm.model_executor.layers.quantization.kernels.scaled_mm.flashinfer ¶

class FlashInferFP8ScaledMMLinearKernel(FP8ScaledMMLinearKernel):
    @classmethod
    def is_supported(
        cls, compute_capability: int | None = None
    ) -> tuple[bool, str | None]:
        if not current_platform.is_cuda():
            return False, "requires CUDA."

        if not has_flashinfer():
            return False, "requires FlashInfer to be installed."

        if compute_capability is not None and compute_capability < 100:
            return False, "requires compute capability 100 and above."

        return True, None

    @classmethod
    def can_implement(cls, c: FP8ScaledMMLinearLayerConfig) -> tuple[bool, str | None]:
        per_tensor_activation_scales = (
            c.activation_quant_key.scale.group_shape.is_per_tensor()
        )
        per_tensor_weight_scales = c.weight_quant_key.scale.group_shape.is_per_tensor()

        if not (per_tensor_activation_scales and per_tensor_weight_scales):
            return False, "requires per tensor activation and weight scales."

        return True, None

    def apply_scaled_mm(
        self,
        *,
        A: torch.Tensor,
        B: torch.Tensor,
        out_dtype: torch.dtype,
        As: torch.Tensor,
        Bs: torch.Tensor,
        bias: torch.Tensor | None,
        output_shape: list,
    ) -> torch.Tensor:
        return flashinfer_scaled_fp8_mm(
            A, B, out_dtype=out_dtype, scale_a=As, scale_b=Bs, bias=bias
        )