vllm.model_executor.kernels.linear.scaled_mm ¶

class AiterInt8ScaledMMLinearKernel(CutlassInt8ScaledMMLinearKernel):
    @classmethod
    def is_supported(
        cls, compute_capability: int | None = None
    ) -> tuple[bool, str | None]:
        if not current_platform.is_rocm():
            return False, "Requires ROCm."

        if compute_capability is not None and compute_capability < 90:
            return False, "requires compute capability 90 and above."

        try:
            import aiter  # noqa: F401 # deliberately attempt to import aiter
        except Exception:
            return False, "requires `aiter` to be installed."

        if not rocm_aiter_ops.is_linear_enabled():
            return (
                False,
                "requires setting `VLLM_ROCM_USE_AITER=1` "
                "and `VLLM_ROCM_USE_AITER_LINEAR=1`. "
                "`VLLM_ROCM_USE_AITER_LINEAR` default is True.",
            )
        return True, None

    @classmethod
    def can_implement(cls, c: Int8ScaledMMLinearLayerConfig) -> tuple[bool, str | None]:
        if not c.input_symmetric:
            return False, "supports symmetric quantization only."
        return True, None

    def apply_weights(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        bias: torch.Tensor | None = None,
    ) -> torch.Tensor:
        """
        `AiterInt8ScaledMMLinearKernel` implements a fused version of
            `output = torch.mm((scale_a * a), (scale_b * b)).to(out_dtype)`
        where scale_a * a and scale_b * b are implemented using numpy-style
        broadcasting.
        Currently only support per-tensor-per-tensor GEMM
        and per-token-per-channel GEMM through AITER
        w8a8 scaled gemm. `AiterInt8ScaledMMLinearKernel` also does not support
        ATIER block scaled GEMM and mix-precision GEMM.
        """
        w_q, w_s, i_s, i_zp, azp_adj = self._get_layer_params(layer)

        # ops.scaled_int8_quant supports both dynamic and static quant:
        # * dynamic, i_s is None and x_s computed from x.
        # * static, i_s is scalar and x_s is i_s.
        symmetric = azp_adj is None
        assert symmetric, (
            "AiterInt8ScaledMMLinearKernel only supports symmetric quantization."
        )
        x_q, x_s, x_zp = ops.scaled_int8_quant(x, i_s, i_zp, symmetric=symmetric)

        assert x_zp is None, (
            "AiterInt8ScaledMMLinearKernel only supports symmetric quantization."
        )
        out_dtype = x.dtype

        assert w_q.shape[0] % 16 == 0 and w_q.shape[1] % 16 == 0
        assert out_dtype is torch.bfloat16 or out_dtype is torch.float16
        assert bias is None or bias.shape[0] == w_q.shape[1] and bias.dtype == out_dtype

        m = x_q.shape[0]  # a
        n = w_q.shape[1]  # b

        per_tensor_scale_a = x_s.numel() == 1
        per_tensor_scale_b = w_s.numel() == 1
        per_token_scale_a = x_s.numel() == m
        per_channel_scale_b = w_s.numel() == n

        # @TODO:
        # Maybe broadcast the per-tensor-scale into per-channel-scale
        # if one of the scale is a per-channel-scale.
        # For now, it only supports:
        # - per-tensor-per-tensor a8w8 scaled GEMM, and
        # - per-token-per-channel a8w8 scaled GEMM
        assert (per_tensor_scale_a and per_tensor_scale_b) or (
            per_token_scale_a and per_channel_scale_b
        ), (
            "Currently only support per-tensor-per-tensor GEMM "
            " and per-token-per-channel GEMM through AITER"
            " w8a8 scaled gemm. `AiterInt8ScaledMMLinearKernel` "
            "does not support AITER block scaled GEMM."
        )

        # gemm_a8w8_CK(a, b, scale_a, scale_b, bias) expects
        # a to be [M, K]
        # b to be [N, K]
        # CutlassInt8ScaledMMLinearKernel prepare weight `w_q` in [K, N] format
        return rocm_aiter_ops.gemm_a8w8(x_q, w_q.t(), x_s, w_s, bias, out_dtype)

class MarlinFP8ScaledMMLinearKernel(FP8ScaledMMLinearKernel):
    """
    FP8 Marlin kernel for GPUs that lack FP8 hardware support.
    Leverages the Marlin kernel for fast weight-only FP8 quantization.
    """

    @classmethod
    def is_supported(
        cls, compute_capability: int | None = None
    ) -> tuple[bool, str | None]:
        if not current_platform.is_cuda():
            return False, "requires CUDA."
        # Check if platform supports FP8 Marlin
        if not is_fp8_marlin_supported():
            return False, "FP8 Marlin requires compute capability 7.5 or higher"
        if envs.VLLM_BATCH_INVARIANT:
            return False, "FP8 Marlin not supported for batch invariant execution."
        if (
            compute_capability is not None
            and compute_capability >= 89
            and not envs.VLLM_TEST_FORCE_FP8_MARLIN
        ):
            return (
                False,
                "To apply FP8 Marlin on high-capability GPUs, please set "
                "VLLM_TEST_FORCE_FP8_MARLIN=1",
            )
        return True, None

    @classmethod
    def can_implement(cls, c: FP8ScaledMMLinearLayerConfig) -> tuple[bool, str | None]:
        return True, None

    def __init__(
        self, c: FP8ScaledMMLinearLayerConfig, layer_param_names: Sequence[str]
    ) -> None:
        super().__init__(c, layer_param_names)
        self.marlin_input_dtype = None
        self.block_quant = self.config.weight_quant_key in {kFp8Static128BlockSym}
        self.size_k_first = not self.block_quant

    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        if self.block_quant:
            weight, weight_scale_inv = process_fp8_weight_block_strategy(
                layer.weight, layer.weight_scale_inv
            )
            # Update layer with new values
            replace_parameter(layer, "weight", weight.data)
            replace_parameter(layer, "weight_scale_inv", weight_scale_inv.data)
        else:
            w_q, *_ = self._get_layer_params(layer)
            # Compressed tensors transposes the weight to (K, N)
            # for channel and tensor quant strategies.
            # So we can skip the transpose if the layout is
            # already (K, N).
            # TODO: Remove this check once the layouts have been
            # canonicalized to a standard (N, K) dimension. See issue
            # #33314 for more details.
            if w_q.shape != (
                layer.input_size_per_partition,
                layer.output_size_per_partition,
            ):
                # transpose the weights to (K,N)
                replace_parameter(
                    layer,
                    "weight",
                    w_q.t(),
                )

        layer.input_scale = None
        prepare_fp8_layer_for_marlin(
            layer, self.size_k_first, input_dtype=self.marlin_input_dtype
        )
        del layer.input_scale

    def apply_weights(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        bias: torch.Tensor | None = None,
    ) -> torch.Tensor:
        if self.block_quant:
            weight_scale = layer.weight_scale_inv
        else:
            weight_scale = layer.weight_scale
        return apply_fp8_marlin_linear(
            input=x,
            weight=layer.weight,
            weight_scale=weight_scale,
            workspace=layer.workspace,
            size_n=layer.output_size_per_partition,
            size_k=layer.input_size_per_partition,
            input_dtype=self.marlin_input_dtype,
            bias=bias,
        )

    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:
        pass