Skip to content

vllm.model_executor.layers.quantization.inc.schemes.inc_mxfp4_scheme

Classes:

INCMxfp4Scheme

Bases: INCScheme

MXFP4 (W4A4) scheme for AutoRound checkpoints.

Dispatches to :class:INCMxfp4LinearMethod for linear layers and :class:INCMxfp4MoEMethod for fused MoE layers; see those classes for the per-module weight layout and kernel-selection details.

Source code in vllm/model_executor/layers/quantization/inc/schemes/inc_mxfp4_scheme.py 
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
class INCMxfp4Scheme(INCScheme):
    """MXFP4 (W4A4) scheme for AutoRound checkpoints.

    Dispatches to :class:`INCMxfp4LinearMethod` for linear layers and
    :class:`INCMxfp4MoEMethod` for fused MoE layers; see those classes for the
    per-module weight layout and kernel-selection details.
    """

    @staticmethod
    def can_handle(layer_config: "INCLayerConfig") -> bool:
        return layer_config.is_mxfp4

    def get_linear_method(
        self,
        config: "INCConfig",
        layer: "torch.nn.Module",
        prefix: str,
        layer_config: "INCLayerConfig",
    ):
        del config, layer, prefix
        from .inc_mxfp4_linear import INCMxfp4LinearMethod

        return INCLinearMethod(INCMxfp4LinearMethod(layer_config))

    def get_moe_method(
        self,
        config: "INCConfig",
        layer: "torch.nn.Module",
        prefix: str,
        layer_config: "INCLayerConfig",
    ):
        del config, prefix, layer_config
        from .inc_mxfp4_moe import INCMxfp4MoEMethod

        return INCMxfp4MoEMethod(layer.moe_config)