Skip to content

vllm.model_executor.kernels.linear.mxfp4

Modules:

Classes:

MxFp4LinearKernel

Bases: ABC

Base class for MXFP4 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, _supports_quant_scheme and can_implement to find the best match for the current hardware.

Methods:

Source code in vllm/model_executor/kernels/linear/mxfp4/base.py 
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
class MxFp4LinearKernel(ABC):
    """Base class for MXFP4 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``, ``_supports_quant_scheme`` and
    ``can_implement`` to find the best match for the current hardware.
    """

    def __init__(self, config: MxFp4LinearLayerConfig) -> 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
    def _supports_quant_scheme(
        cls,
        weight_key: QuantKey | None,
        activation_key: QuantKey | None,
    ) -> bool:
        """Return whether this kernel implements *weight_key* x *activation_key*.

        Selection uses this to dispatch on the requested activation format
        (from the checkpoint or ``--quantization-config.linear.activation``).
        The default covers MXFP4 weights with MXFP4 or unquantized activations;
        kernels for other activation formats override it.
        """
        return weight_key == kMxfp4Static and activation_key in (None, kMxfp4Dynamic)

    @classmethod
    @abstractmethod
    def can_implement(cls, config: MxFp4LinearLayerConfig) -> 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

_supports_quant_scheme(weight_key, activation_key) classmethod

Return whether this kernel implements weight_key x activation_key.

Selection uses this to dispatch on the requested activation format (from the checkpoint or --quantization-config.linear.activation). The default covers MXFP4 weights with MXFP4 or unquantized activations; kernels for other activation formats override it.

Source code in vllm/model_executor/kernels/linear/mxfp4/base.py 
54
55
56
57
58
59
60
61
62
63
64
65
66
67
@classmethod
def _supports_quant_scheme(
    cls,
    weight_key: QuantKey | None,
    activation_key: QuantKey | None,
) -> bool:
    """Return whether this kernel implements *weight_key* x *activation_key*.

    Selection uses this to dispatch on the requested activation format
    (from the checkpoint or ``--quantization-config.linear.activation``).
    The default covers MXFP4 weights with MXFP4 or unquantized activations;
    kernels for other activation formats override it.
    """
    return weight_key == kMxfp4Static and activation_key in (None, kMxfp4Dynamic)

apply_weights(layer, x, bias=None) abstractmethod

Run the quantized GEMM.

Source code in vllm/model_executor/kernels/linear/mxfp4/base.py 
85
86
87
88
89
90
91
92
93
@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

can_implement(config) abstractmethod classmethod

Return whether this kernel can handle config.

Source code in vllm/model_executor/kernels/linear/mxfp4/base.py 
69
70
71
72
73
@classmethod
@abstractmethod
def can_implement(cls, config: MxFp4LinearLayerConfig) -> tuple[bool, str | None]:
    """Return whether this kernel can handle *config*."""
    raise NotImplementedError

is_supported(compute_capability=None) abstractmethod classmethod

Return whether this kernel can run on the current platform.

Source code in vllm/model_executor/kernels/linear/mxfp4/base.py 
46
47
48
49
50
51
52
@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

process_weights_after_loading(layer) abstractmethod

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.

Source code in vllm/model_executor/kernels/linear/mxfp4/base.py 
75
76
77
78
79
80
81
82
83
@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

MxFp4LinearLayerConfig dataclass

Configuration for an MXFP4 linear layer.

All MXFP4 layers share the same structure: packed uint8 weights (2 FP4 values per byte) and per-block weight scales (group size 32).

Attributes:

  • activation_quant_key (QuantKey | None) –

    Identifies the activation quantization format, or None when activations must not be quantized.

Source code in vllm/model_executor/kernels/linear/mxfp4/base.py 
16
17
18
19
20
21
22
23
24
25
26
27
28
29
@dataclass
class MxFp4LinearLayerConfig:
    """Configuration for an MXFP4 linear layer.

    All MXFP4 layers share the same structure: packed uint8 weights (2 FP4 values per
    byte) and per-block weight scales (group size 32).

    Attributes:
        activation_quant_key: Identifies the activation quantization format,
            or `None` when activations must not be quantized.

    """

    activation_quant_key: QuantKey | None = None