vllm.model_executor.kernels.linear ¶

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 CutlassNvFp4LinearKernel(NvFp4LinearKernel):
    """NVFP4 GEMM via the vLLM CUTLASS kernel."""

    @classmethod
    def is_supported(
        cls, compute_capability: int | None = None
    ) -> tuple[bool, str | None]:
        if cutlass_fp4_supported():
            return True, None
        return False, "CUTLASS FP4 kernels not available"

    @classmethod
    def can_implement(cls, config: NvFp4LinearLayerConfig) -> tuple[bool, str | None]:
        return True, None

    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        layer.weight_scale = torch.nn.Parameter(
            swizzle_blockscale(layer.weight_scale.data), requires_grad=False
        )
        padded_weight, weights_padding_cols = pad_nvfp4_weight_for_cutlass(
            layer.weight.data
        )
        layer.weight = torch.nn.Parameter(padded_weight, requires_grad=False)
        layer.weights_padding_cols = weights_padding_cols

    def apply_weights(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        bias: torch.Tensor | None = None,
    ) -> torch.Tensor:
        output_size = layer.output_size_per_partition
        output_dtype = x.dtype
        output_shape = [*x.shape[:-1], output_size]

        x_fp4, x_blockscale = scaled_fp4_quant(
            x,
            layer.input_global_scale_inv,
            is_sf_swizzled_layout=True,
            backend="cutlass",
        )

        x_fp4 = pad_nvfp4_activation_for_cutlass(
            x_fp4, getattr(layer, "weights_padding_cols", 0)
        )

        out = cutlass_scaled_fp4_mm(
            x_fp4,
            layer.weight,
            x_blockscale,
            layer.weight_scale,
            layer.alpha,
            output_dtype,
        )

        out = slice_nvfp4_output(out, output_size)

        if bias is not None:
            out = out + bias
        return out.view(*output_shape)

class EmulationMxfp8LinearKernel(Mxfp8LinearKernel):
    """Software emulation fallback for MXFP8 (dequant to BF16)."""

    @classmethod
    def is_supported(
        cls, compute_capability: int | None = None
    ) -> tuple[bool, str | None]:
        return True, None

    @classmethod
    def can_implement(cls, c: Mxfp8LinearLayerConfig) -> tuple[bool, str | None]:
        return True, None

    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        weight = layer.weight.data  # [N, K]
        N, K = weight.shape
        scale_k = K // MXFP8_BLOCK_SIZE

        weight_scale = layer.weight_scale.data[:N, :scale_k].contiguous()

        layer.weight = Parameter(weight.contiguous(), requires_grad=False)
        layer.weight_scale = Parameter(weight_scale, requires_grad=False)

    def apply_weights(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        bias: torch.Tensor | None = None,
    ) -> torch.Tensor:
        weight_scale = layer.weight_scale
        if weight_scale.dtype != MXFP8_SCALE_DTYPE:
            raise ValueError(
                f"Emulation backend requires {MXFP8_SCALE_DTYPE} "
                f"weight_scale dtype, got {weight_scale.dtype}."
            )
        if weight_scale.ndim != 2:
            raise ValueError(
                f"Emulation backend requires 2D weight_scale, "
                f"got {weight_scale.ndim}D. "
                f"Ensure process_weights_after_loading was called."
            )

        weight_bf16 = dequant_mxfp8_to_bf16(layer.weight, weight_scale)
        output = torch.nn.functional.linear(x, weight_bf16, bias)
        return output.to(x.dtype)

class EmulationNvFp4LinearKernel(NvFp4LinearKernel):
    """Software emulation fallback for NVFP4 (dequant → BF16 matmul)."""

    @classmethod
    def is_supported(
        cls, compute_capability: int | None = None
    ) -> tuple[bool, str | None]:
        # Always available as a last-resort fallback.
        return True, None

    @classmethod
    def can_implement(cls, config: NvFp4LinearLayerConfig) -> tuple[bool, str | None]:
        return True, None

    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        # Move the E2M1 lookup table to the device now, because
        # `.to(device)` is not allowed during CUDA graph capture.
        kE2M1ToFloat_handle.val = kE2M1ToFloat_handle.val.to(layer.weight.device)

    def apply_weights(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        bias: torch.Tensor | None = None,
    ) -> torch.Tensor:
        out = run_nvfp4_emulations(
            x=x,
            input_global_scale=layer.input_global_scale_inv,
            weight=layer.weight,
            weight_scale_swizzled=layer.weight_scale,
            weight_global_scale=layer.weight_global_scale,
            swizzle=False,
        )
        if bias is not None:
            out = out + bias
        return out

class FbgemmNvFp4LinearKernel(NvFp4LinearKernel):
    """NVFP4 GEMM via FBGEMM."""

    @classmethod
    def is_supported(
        cls, compute_capability: int | None = None
    ) -> tuple[bool, str | None]:
        if has_fbgemm_gpu():
            return True, None
        return False, "fbgemm_gpu required"

    @classmethod
    def can_implement(cls, config: NvFp4LinearLayerConfig) -> tuple[bool, str | None]:
        return True, None

    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        swizzled = swizzle_blockscale(layer.weight_scale.data)
        layer.weight_scale = torch.nn.Parameter(
            swizzled.view(-1).view(torch.uint8), requires_grad=False
        )

    def apply_weights(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        bias: torch.Tensor | None = None,
    ) -> torch.Tensor:
        import fbgemm_gpu  # noqa: F401 - registers torch.ops.fbgemm.*

        output_size = layer.output_size_per_partition
        output_dtype = x.dtype
        output_shape = [*x.shape[:-1], output_size]

        x_fp4, x_blockscale = scaled_fp4_quant(
            x,
            layer.input_global_scale_inv,
            is_sf_swizzled_layout=True,
            backend="fbgemm",
        )

        out = torch.ops.fbgemm.f4f4bf16(
            x_fp4,
            layer.weight,
            x_blockscale.view(-1).view(torch.uint8),
            layer.weight_scale,
            layer.alpha,
            use_mx=False,
        ).to(output_dtype)

        out = slice_nvfp4_output(out, output_size)

        if bias is not None:
            out = out + bias
        return out.view(*output_shape)

class FlashInferCudnnNvFp4LinearKernel(NvFp4LinearKernel):
    """NVFP4 GEMM via FlashInfer's cuDNN wrapper."""

    @classmethod
    def is_supported(
        cls, compute_capability: int | None = None
    ) -> tuple[bool, str | None]:
        if has_flashinfer():
            return True, None
        return False, "FlashInfer required"

    @classmethod
    def can_implement(cls, config: NvFp4LinearLayerConfig) -> tuple[bool, str | None]:
        return True, None

    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        # cuDNN uses the same swizzled + padded layout as CUTLASS
        layer.weight_scale = torch.nn.Parameter(
            swizzle_blockscale(layer.weight_scale.data), requires_grad=False
        )
        padded_weight, weights_padding_cols = pad_nvfp4_weight_for_cutlass(
            layer.weight.data
        )
        layer.weight = torch.nn.Parameter(padded_weight, requires_grad=False)
        layer.weights_padding_cols = weights_padding_cols

    def apply_weights(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        bias: torch.Tensor | None = None,
    ) -> torch.Tensor:
        output_size = layer.output_size_per_partition
        output_dtype = x.dtype
        output_shape = [*x.shape[:-1], output_size]

        x_fp4, x_blockscale = scaled_fp4_quant(
            x,
            layer.input_global_scale_inv,
            is_sf_swizzled_layout=True,
            backend="flashinfer-cudnn",
        )

        x_fp4 = pad_nvfp4_activation_for_cutlass(
            x_fp4, getattr(layer, "weights_padding_cols", 0)
        )

        out = flashinfer_scaled_fp4_mm(
            x_fp4,
            layer.weight,
            x_blockscale,
            layer.weight_scale,
            layer.alpha,
            output_dtype,
            backend="cudnn",
        )

        out = slice_nvfp4_output(out, output_size)

        if bias is not None:
            out = out + bias
        return out.view(*output_shape)

class FlashInferCutlassMxfp8LinearKernel(Mxfp8LinearKernel):
    """MXFP8 W8A8 GEMM via FlashInfer CUTLASS (SM100+)."""

    @classmethod
    def is_supported(
        cls, compute_capability: int | None = None
    ) -> tuple[bool, str | None]:
        if current_platform.has_device_capability(100):
            return True, None
        return False, "requires >=sm_100 (Blackwell)"

    @classmethod
    def can_implement(cls, c: Mxfp8LinearLayerConfig) -> tuple[bool, str | None]:
        return True, None

    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        weight = layer.weight.data  # [N, K]
        N, K = weight.shape

        scale_k = K // MXFP8_BLOCK_SIZE
        weight_scale_2d = layer.weight_scale.data[:N, :scale_k].contiguous()
        weight_scale_swizzled = swizzle_mxfp8_scale(weight_scale_2d, M=N, K=K)

        layer.weight = Parameter(weight.contiguous(), requires_grad=False)
        layer.weight_scale = Parameter(
            weight_scale_swizzled.contiguous(), requires_grad=False
        )

    def apply_weights(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        bias: torch.Tensor | None = None,
    ) -> torch.Tensor:
        weight = layer.weight
        weight_scale = layer.weight_scale
        out_dtype = x.dtype
        N, K = weight.shape

        input_shape = x.shape
        input_2d = x.view(-1, K)
        M_orig = input_2d.shape[0]

        min_dim = 128

        assert min_dim <= K, (
            f"mm_mxfp8 requires K >= {min_dim}, got K={K}. "
            f"in_features is too small for mm_mxfp8."
        )
        assert K % MXFP8_BLOCK_SIZE == 0, (
            f"mm_mxfp8 requires K to be divisible by {MXFP8_BLOCK_SIZE}, got K={K}."
        )
        assert min_dim <= N, (
            f"mm_mxfp8 requires N >= {min_dim}, got N={N}. "
            f"out_features is too small for mm_mxfp8."
        )

        M_padded = ((M_orig + min_dim - 1) // min_dim) * min_dim
        if M_padded != M_orig:
            pad_rows = M_padded - M_orig
            input_2d = torch.nn.functional.pad(input_2d, (0, 0, 0, pad_rows))

        input_mxfp8, input_scale = mxfp8_e4m3_quantize(
            input_2d, is_sf_swizzled_layout=True
        )

        if not weight.is_contiguous():
            weight = weight.contiguous()

        output = vllm_flashinfer.mm_mxfp8(
            input_mxfp8,
            weight.t(),
            input_scale,
            weight_scale,
            out_dtype=out_dtype,
            backend="cutlass",
        )

        if M_padded != M_orig:
            output = output[:M_orig, :]

        if bias is not None:
            output = output + bias

        output_shape = (*input_shape[:-1], N)
        return output.view(output_shape)

class FlashInferCutlassNvFp4LinearKernel(NvFp4LinearKernel):
    """NVFP4 GEMM via FlashInfer's CUTLASS wrapper."""

    @classmethod
    def is_supported(
        cls, compute_capability: int | None = None
    ) -> tuple[bool, str | None]:
        from vllm.model_executor.layers.quantization.utils.nvfp4_utils import (
            cutlass_fp4_supported,
        )

        if (
            cutlass_fp4_supported()
            and current_platform.has_device_capability(100)
            and has_flashinfer()
        ):
            return True, None
        return False, "FlashInfer + >=sm_100 required"

    @classmethod
    def can_implement(cls, config: NvFp4LinearLayerConfig) -> tuple[bool, str | None]:
        return True, None

    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        layer.weight_scale = torch.nn.Parameter(
            swizzle_blockscale(layer.weight_scale.data), requires_grad=False
        )
        padded_weight, weights_padding_cols = pad_nvfp4_weight_for_cutlass(
            layer.weight.data
        )
        layer.weight = torch.nn.Parameter(padded_weight, requires_grad=False)
        layer.weights_padding_cols = weights_padding_cols

    def apply_weights(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        bias: torch.Tensor | None = None,
    ) -> torch.Tensor:
        output_size = layer.output_size_per_partition
        output_dtype = x.dtype
        output_shape = [*x.shape[:-1], output_size]

        x_fp4, x_blockscale = scaled_fp4_quant(
            x,
            layer.input_global_scale_inv,
            is_sf_swizzled_layout=True,
            backend="flashinfer-cutlass",
        )

        x_fp4 = pad_nvfp4_activation_for_cutlass(
            x_fp4, getattr(layer, "weights_padding_cols", 0)
        )

        out = flashinfer_scaled_fp4_mm(
            x_fp4,
            layer.weight,
            x_blockscale,
            layer.weight_scale,
            layer.alpha,
            output_dtype,
            backend="cutlass",
        )

        out = slice_nvfp4_output(out, output_size)

        if bias is not None:
            out = out + bias
        return out.view(*output_shape)

class FlashInferFp8DeepGEMMDynamicBlockScaledKernel(
    Fp8BlockScaledDynamicMMLinearKernel
):
    """
    Conditional FlashInfer / DeepGEMM FP8 block-scaled GEMM.

    Dispatches between two kernels based on input batch size:
    - Small batches (M < 32): FlashInfer's swapAB trick for better utilisation.
    - Large batches (M >= 32): DeepGEMM for peak throughput.

    apply_input_quant is False because FlashInfer accepts BF16 input and
    handles FP8 conversion internally.  The DeepGEMM branch therefore
    quantises BF16→FP8 inside apply_mm via a closure before dispatching to
    the DeepGEMM kernel — keeping both branches compatible with the single
    BF16 tensor operand list passed by torch.cond.
    """

    base_type: ClassVar[type[FlashInferFp8BlockScaledMMKernel]] = (
        FlashInferFp8BlockScaledMMKernel
    )
    fallback_type: ClassVar[type[DeepGemmFp8BlockScaledMMKernel]] = (
        DeepGemmFp8BlockScaledMMKernel
    )
    apply_input_quant: ClassVar[bool] = False

    def __init__(self, config: FP8ScaledMMLinearLayerConfig):
        super().__init__(config)
        self.base: FlashInferFp8BlockScaledMMKernel
        self.fallback: DeepGemmFp8BlockScaledMMKernel

    def process_weights_after_loading(self, layer: torch.nn.Module):
        # DeepGEMM need post-processing; both kernels share the same
        # parameter tensor layout so processing once is sufficient.
        self.fallback.process_weights_after_loading(layer)

    def apply_block_scaled_mm(
        self,
        A: torch.Tensor,
        B: torch.Tensor,
        As: torch.Tensor,
        Bs: torch.Tensor,
    ) -> torch.Tensor:
        group_size = self.weight_group_shape.col
        use_deep_gemm_e8m0 = self.fallback.use_deep_gemm_e8m0

        return torch.ops.vllm.dynamic_flashinfer_deepgemm_blockscale_gemm(
            A, B, Bs, group_size, use_deep_gemm_e8m0
        )

class FlashInferTrtllmNvFp4LinearKernel(NvFp4LinearKernel):
    """NVFP4 GEMM via FlashInfer's TensorRT-LLM wrapper."""

    @classmethod
    def is_supported(
        cls, compute_capability: int | None = None
    ) -> tuple[bool, str | None]:
        if has_flashinfer():
            return True, None
        return False, "FlashInfer required"

    @classmethod
    def can_implement(cls, config: NvFp4LinearLayerConfig) -> tuple[bool, str | None]:
        return True, None

    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        from flashinfer import shuffle_matrix_a, shuffle_matrix_sf_a

        weight = layer.weight.data
        weight_scale = layer.weight_scale.data
        epilogue_tile_m = 128

        layer.weight = torch.nn.Parameter(
            shuffle_matrix_a(weight.view(torch.uint8), epilogue_tile_m),
            requires_grad=False,
        )
        layer.weight_scale = torch.nn.Parameter(
            shuffle_matrix_sf_a(weight_scale.view(torch.uint8), epilogue_tile_m)
            .reshape(weight_scale.shape)
            .view(torch.float8_e4m3fn),
            requires_grad=False,
        )

    def apply_weights(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        bias: torch.Tensor | None = None,
    ) -> torch.Tensor:
        output_size = layer.output_size_per_partition
        output_dtype = x.dtype
        output_shape = [*x.shape[:-1], output_size]

        x_fp4, x_blockscale = scaled_fp4_quant(
            x,
            layer.input_global_scale_inv,
            is_sf_swizzled_layout=True,
            backend="flashinfer-trtllm",
        )

        out = flashinfer_scaled_fp4_mm(
            x_fp4,
            layer.weight,
            x_blockscale,
            layer.weight_scale,
            layer.alpha,
            output_dtype,
            backend="trtllm",
        )

        out = slice_nvfp4_output(out, output_size)

        if bias is not None:
            out = out + bias
        return out.view(*output_shape)

class MarlinMxfp8LinearKernel(Mxfp8LinearKernel):
    """MXFP8 W8A16 GEMM via Marlin (SM80+)."""

    @classmethod
    def is_supported(
        cls, compute_capability: int | None = None
    ) -> tuple[bool, str | None]:
        from vllm.model_executor.layers.quantization.utils.marlin_utils_fp8 import (
            is_fp8_marlin_supported,
        )

        if is_fp8_marlin_supported():
            return True, None
        return False, "Marlin FP8 not available"

    @classmethod
    def can_implement(cls, c: Mxfp8LinearLayerConfig) -> tuple[bool, str | None]:
        return True, None

    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        from vllm.model_executor.layers.quantization.utils.marlin_utils_fp8 import (
            prepare_mxfp8_layer_for_marlin,
        )

        prepare_mxfp8_layer_for_marlin(layer)

    def apply_weights(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        bias: torch.Tensor | None = None,
    ) -> torch.Tensor:
        from vllm.model_executor.layers.quantization.utils.marlin_utils_fp8 import (
            apply_mxfp8_marlin_linear,
        )

        return apply_mxfp8_marlin_linear(
            input=x,
            weight=layer.weight,
            weight_scale=layer.weight_scale,
            workspace=layer.workspace,
            size_n=layer.output_size_per_partition,
            size_k=layer.input_size_per_partition,
            bias=bias,
        )

class MarlinNvFp4LinearKernel(NvFp4LinearKernel):
    """NVFP4 weight-only GEMM via Marlin (W4A16)."""

    @classmethod
    def is_supported(
        cls, compute_capability: int | None = None
    ) -> tuple[bool, str | None]:
        if is_fp4_marlin_supported():
            return True, None
        return False, "Marlin FP4 not available"

    @classmethod
    def can_implement(cls, config: NvFp4LinearLayerConfig) -> tuple[bool, str | None]:
        return True, None

    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        logger.warning_once(
            "Your GPU does not have native support for FP4 computation but "
            "FP4 quantization is being used. Weight-only FP4 compression "
            "will be used leveraging the Marlin kernel. This may degrade "
            "performance for compute-heavy workloads."
        )
        prepare_fp4_layer_for_marlin(layer)

    def apply_weights(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        bias: torch.Tensor | None = None,
    ) -> torch.Tensor:
        return apply_fp4_marlin_linear(
            input=x,
            weight=layer.weight,
            weight_scale=layer.weight_scale,
            weight_global_scale=layer.weight_global_scale,
            workspace=layer.workspace,
            size_n=layer.output_size_per_partition,
            size_k=layer.input_size_per_partition,
            bias=bias,
        )

@dataclass
class Mxfp8LinearLayerConfig:
    """Configuration for an MXFP8 linear layer.

    All MXFP8 layers share the same structure: FP8-E4M3 weights with
    uint8 (E8M0) per-block scales at block size 32.
    """

    pass

class NvFp4LinearKernel(ABC):
    """Base class for NVFP4 quantized linear kernels.

    Each subclass implements a specific GEMM backend (CUTLASS, Marlin, etc).
    The kernel selection mechanism iterates over registered subclasses in
    priority order,calling ``is_supported`` and ``can_implement`` to find the best
    match for the current hardware.
    """

    def __init__(self, config: NvFp4LinearLayerConfig) -> None:
        assert self.can_implement(config)[0]
        assert self.is_supported()[0]
        self.config = config

    @classmethod
    @abstractmethod
    def is_supported(
        cls, compute_capability: int | None = None
    ) -> tuple[bool, str | None]:
        """Return whether this kernel can run on the current platform."""
        raise NotImplementedError

    @classmethod
    @abstractmethod
    def can_implement(cls, config: NvFp4LinearLayerConfig) -> tuple[bool, str | None]:
        """Return whether this kernel can handle *config*."""
        raise NotImplementedError

    @abstractmethod
    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        """Transform weights into the format required by this kernel.

        Called once after checkpoint weights have been loaded onto the
        device.  Implementations should repack / swizzle / pad weights
        and scales in-place on *layer*.
        """
        raise NotImplementedError

    @abstractmethod
    def apply_weights(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        bias: torch.Tensor | None = None,
    ) -> torch.Tensor:
        """Run the quantized GEMM."""
        raise NotImplementedError